There is an easy version of the PocketOS story.

An AI coding agent went rogue. It ignored instructions. It deleted production data. Then, in a moment almost too perfect for the internet, it “confessed” in writing.

That is the viral version.

It is not the version CEOs and Chief Risk Officers should care about.

The harder truth is this: a business allowed a non-human worker to access authority it did not properly understand, limit, supervise, or recover from.

That is not an AI story first.

It is a governance story.

Or, more plainly: the intern had the keys.

According to Jer Crane’s original post and subsequent reporting by The Register, a Cursor AI agent working on what was supposed to be a staging issue found a Railway credential and used it to delete PocketOS’s production data environment. PocketOS serves rental businesses that rely on its software for reservations, payments, customer records, and day-to-day operations.

For those customers, this was not an abstract AI failure.

It was Saturday morning chaos.

Bookings disappeared. Customer records were missing. Operators had to reconstruct reality from payment histories, emails, calendars, and whatever else they could find.

Railway later told The Register that the data had been restored and that additional safeguards had been added. That is important. It changes the ending of the incident.

It does not change the lesson.

Read more

The Prompt

Act as a Military Strategist (John Boyd OODA Loop).

We are facing a market change/competitor move: [Insert Situation].
We are currently stuck.

Diagnose our OODA Loop:
1. Observe: Do we have real-time data, or are we looking at last month's report?
2. Orient: Are we paralysed by "Analysis Paralysis"? What cultural bias is stopping us from seeing the truth?
3. Decide: Who is the bottlenecks? (Is it a person or a process?).
4. Act: Script the "Minimum Viable Move" we can make *today* to disrupt the competitor's rhythm.

Share

The Executive Upgrade: Subscribers Only

The 100% Certainty Myth
General Colin Powell had a rule: The 40-70 Rule.
If you have less than 40% of the information, don’t decide.
If you wait for more than 70% of the information, you are too late.
You must train your team to pull the trigger at 70%. The cost of being wrong is usually less than the cost of being slow.

Here are the two Governance Protocols to force speed.

Read more

Canva’s newest AI launch is not really a design story.

It is a power grab for the marketing workflow.

At Canva Create 2026, Canva launched Canva AI 2.0 and framed it as its biggest product shift since the company started. The company is explicitly trying to move from “a design platform with AI tools” to “an AI platform with design tools.” The release bundles conversational design, connectors into systems like Slack, Gmail, Google Drive, Google Calendar, Notion, Zoom, HubSpot, Microsoft, Atlassian and Linear, plus scheduling, web research, brand intelligence, Sheets AI, and Canva Code 2.0. Canva says the release is available as a research preview, with broader rollout coming over the following weeks.

That framing matters.

Because once you stop reading this as “more AI for design,” the strategic move becomes much clearer.

Canva is trying to become the layer that sits between company context and outward-facing content.

Not just the place where someone makes a deck.

The place where a brief, a transcript, a campaign plan, a sales email, a product update, and a brand system get turned into publishable output.

That is why this launch is worth paying attention to.

And it is also why it deserves more scepticism than the average product keynote gets.

Subscribe now

What is actually valuable here

The strongest part of Canva’s pitch is not the generative layer.

It is the operating layer around it.

Canva’s official announcement says outputs remain fully layered and editable, and that Brand Intelligence applies fonts, colours, and style from the first draft rather than after the fact. That is a much more serious enterprise proposition than “type a prompt and get a pretty picture.” It is a pitch about reducing rework, not just increasing novelty.

That maps directly to a real marketing problem.

In Salesforce’s latest ANZ marketing research, 83% of marketers say customers now expect two-way conversations, 72% of Australian marketers say they still struggle to respond promptly, and up to 80% say they need more personalised content than they can currently produce. That is the gap Canva is targeting: not a shortage of ideas, but a shortage of operational capacity.

This is the line that should make every CMO stop and think:

That is exactly the right lens.

Because the real value in Canva’s release is not that it makes more content possible.

It is that it tries to compress the path from context to usable, on-brand content.

If that works, Canva becomes more than design software.

It becomes marketing operations software wearing a design interface.

Share this with your CMO if your content problem is no longer creativity, but throughput, control, and brand consistency at scale.

Share

Where the marketing faff starts

This is also where the launch starts to drift into theatre.

The announcement is full of the usual language of this cycle:

agentic
memory
orchestration
foundation model
start and end your day in Canva

A lot of that will sound familiar because it is now the standard way SaaS companies describe “we want to be a bigger platform.”

And some of it may prove real.

But the gap between launch language and operating reality is still large.

Digiday’s latest research says trust and complexity remain major blockers, and 54% of respondents say their companies do not use agentic AI in workflows today. That does not mean agentic systems will fail. It does mean vendors are talking as though the operating model has already arrived, while most teams are still nowhere near ready to run it.

There is a second problem.

The more visible the AI becomes, the more brand trust can go backwards.

According to a recent eMarketer summary, when consumers notice AI-generated content in brand marketing, they are roughly four times more likely to trust the brand less than more: 31% versus 7%. That is not a small warning label. It means speed alone is not a win. Scale alone is not a win. More output is often just more branded noise.

So my current split on this announcement is fairly simple:

The value is in control, editability, and workflow compression.

The faff is in the suggestion that every marketing team is now ready to hand over the creative operating loop to agents.

They are not.

Leave a comment if you think most marketing teams are still further from agentic workflows than vendors want boards to believe.

Leave a comment

The procurement reality check

Put a procurement hat on for a moment and the category looks much less crowded than the AI discourse suggests.

For a chief procurement officer and a CMO, the most relevant comparison is still Adobe.

Adobe’s current enterprise pitch is explicitly about a content supply chain: planning, creation, activation, delivery, reporting, insights, brand intelligence, and governance across enterprise workflows. That is the heavier, more mature answer for organisations already standardised on Adobe systems and formal content operations.

Figma is closer than it used to be, but still only partly.

Figma Buzz is clearly aimed at brand and marketing teams producing on-brand assets at scale, with editable restrictions, approvals, template locking, and bulk asset creation. But Figma’s centre of gravity still sits closer to product and interface work than Canva’s broader business-content footprint.

The coding agents and model vendors matter, but mostly as complements, not core substitutes.

That is one of the more revealing parts of Canva’s recent strategy. Canva is not trying to eliminate ChatGPT or Claude. It is integrating with them. Canva now supports design creation inside ChatGPT and Claude Design, which suggests Canva sees the general assistant as an upstream interface and Canva as the structured output layer underneath.

That is a much more realistic position than pretending one tool will absorb the entire stack.

So the procurement answer is:

Yes, Canva has moved up a tier.
No, it is not suddenly the uncontested enterprise standard.
And yes, it brings a genuinely useful new toolset to the table.

But the value case is strongest where marketing teams want more throughput without buying a full Adobe-style content supply chain.

That is meaningful.

It is just not the same thing as category dominance.

Why the timing matters

The market context is not incidental here.

In August 2025, Reuters reported that Canva’s employee share sale valued the company at $42 billion, ahead of a reported IPO push. Since then, software markets have become much less forgiving. Reuters reported that software and services stocks shed $830 billion in market value over six trading days in February, and this week reported that the software ETF is down about 16% in 2026 while the semiconductor ETF is up more than 43%. That is not just volatility. It is a live repricing of who investors think benefits from AI and who gets squeezed by it.

Reuters has also reported that the best defence for software companies may be proprietary data and deep workflow embedding. That is the subtext sitting underneath Canva’s product story. The company is not just trying to prove that it has AI features. It is trying to prove that it owns a defensible position inside the flow of work.

That is why this announcement reads as both product strategy and capital-markets strategy.

The product strategy says:

we are more than a design tool.

The market strategy says:

we deserve to be valued as an AI workflow platform, not as a vulnerable SaaS application waiting to be abstracted away.

That does not make the launch fake.

It makes it legible.

The CEO and CMO read

The right way to read Canva’s launch is not to ask whether the demos looked impressive.

The right question is whether Canva is building a stronger position at the point where company context becomes market-facing communication.

That is the prize.

If Canva can pull inputs from the systems where work already happens, keep the outputs editable, apply brand rules from the start, and make content production materially faster without making it visibly worse, then this launch will prove genuinely important.

If it cannot, then this will age like a very polished IPO deck.

My view is that Canva has identified the right strategic position.

But it has not yet proved that it owns it.

That is the distinction CEOs and CMOs should care about.

Because in this market, AI features are cheap.

Workflow position is not.

Share this with a peer who still sees generative AI as a software feature, not a structural fight over leverage, lock-in, and operating control.

Share

Claude Mythos Preview is Anthropic’s unreleased general-purpose generative AI frontier model, notable because its cyber capabilities appear materially stronger than previous systems. Anthropic has restricted access through Project Glasswing, where selected organisations use it for defensive security work, and the U.K. AI Security Institute says the model is a genuine step up on cyber tasks, though the published results are still based on controlled evaluations rather than open real-world deployment.

For CEOs and CISOs, the relevant question is not whether Mythos is dramatic enough to justify the headlines.

It is whether exploit discovery is starting to move faster than leadership decisions and defensive response.

That is the real issue here. Anthropic says it does not plan to make Claude Mythos Preview generally available for now, but is already giving Project Glasswing partners access to find and fix vulnerabilities in critical software. AISI says the model is the first to complete its 32-step simulated attack range end to end, and that it succeeds on expert-level cyber tasks 73% of the time. If those signals are even directionally right, the pressure point for companies is obvious: the time gap between discovering weaknesses and fixing them may be shrinking.

That is why this is not mainly an “AI product” story.

It is a decision-speed story.

A governance story.

A resilience story.

The White House has discussed Mythos with Anthropic, that federal agencies are preparing for controlled access, and that U.S. officials have briefed major bank CEOs on the potential implications. That is not the pattern of a product quietly sitting in a lab. It is the pattern of a strategically restricted capability that institutions already think matters.

If you are the CEO, the question is whether your company can still make and execute risk decisions fast enough.

If you are the CISO, the question is whether your controls still assume attackers are slower, less automated, and more resource-constrained than they may soon be.

That is the frame.

What Mythos changes for CEOs and CISOs

Read more

Most executives are not leading.

They are routing.

Information comes in from one person, gets passed to another, and somehow this gets mistaken for executive work.

A customer issue needs escalation.
A team needs alignment.
A project is “at risk.”
A decision is waiting on three departments.

So the executive steps in.

They forward.
They summarise.
They chase.
They clarify.
They sit in the meeting.
They send the follow-up.

They become the bridge between functions that should already know how to operate.

It feels important because it is visible.

It feels valuable because it is close to the action.

It feels like leadership because everyone seems to need you.

Usually, it is none of those things.

It is coordination drag with a senior title.

And many well-paid executives spend most of their week doing it.

That is not scale.

That is human middleware.

If this feels uncomfortably familiar, share this with a peer or leadership team member who may be stuck in the same trap.

Share

The Problem Is Not Laziness. It Is Misclassification.

Most executives are not working too little.

They are working in the wrong category of work.

That is why Eisenhower’s framework still matters.

It cuts through the theatre of busyness.

Not every task that demands attention deserves executive ownership.

The four categories are simple:

1. Urgent and important
   Crises. Critical decisions. Live failures.
   Do these.

2. Urgent and not important
   Email churn. Status meetings. Approval theatre. Coordination pings.
   Delegate these.

3. Not urgent and important
   Strategy. Hiring. Organisational design. Capital allocation. Thinking.
   This is the real job.

4. Not urgent and not important
   Noise. Vanity tasks. Habitual checking. Legacy meetings.
   Delete these.

The framework is old.

The mistake is current.

Most executives still spend too much of their week in the urgent categories because urgency is socially rewarded.

You get to look responsive.
You get to look involved.
You get to look indispensable.

But looking indispensable is often just another way of saying the system is weak.

Your Real Job Is Usually Hiding in the Neglected Quadrant

The highest-value executive work rarely feels dramatic.

It usually does not arrive as a fire drill.

It sits in the neglected quadrant:

This is where executive leverage actually lives.

Making the hard hiring call before the wrong person creates drag.

Fixing a broken reporting line before it turns into politics.

Clarifying strategic priorities before teams scatter effort.

Rethinking pricing before the market forces the issue.

Building an operating cadence that prevents chaos instead of reacting to it.

This work is easy to postpone because nothing explodes today if you ignore it.

That is exactly why weak leaders neglect it.

Reactive leaders manage consequences.

Serious leaders shape conditions.

If too little of your week is spent on work that compounds, you are not really steering the company.

You are absorbing workflow on behalf of it.

Subscribe now

Why Smart Executives Become Expensive Project Managers

This is where the trap gets subtle.

Most executives do not choose low-leverage work.

They inherit it.

At first, stepping in feels rational.

You know the business better than anyone else.
You can unblock things quickly.
You do not trust the handoff quality.
The team is not mature enough yet.
It is faster if you just do it.

That logic works in short bursts.

Then it becomes the operating model.

And once that happens, the organisation learns the wrong lesson:

important things move through you
unclear things belong to you
cross-functional tension gets escalated to you
decisions wait for you
ownership climbs upward instead of pushing downward

Now you are not leading a system.

You are compensating for the absence of one.

That is why so many executives become glorified routers.

Not because they are weak.

Because the organisation has quietly trained itself to use executive attention as a substitute for structure.

The Cost Is Bigger Than Lost Time

This is not just a productivity problem.

It is a scale problem.

When an executive spends too much time routing, four things happen at once.

First, strategic work gets starved.
The highest-leverage work gets displaced by coordination noise.

Second, the organisation gets weaker.
Teams stop solving problems where they should be solved, because escalation is always available.

Third, decision quality gets worse.
The executive is flooded with low-context, high-volume input and ends up making too many shallow calls.

Fourth, dependency hardens into culture.
Everyone says they want empowered teams, but the system rewards bringing half-formed problems upward.

This is how companies get slower as they grow.

Not because scale automatically creates bureaucracy.

Because leaders keep centralising the wrong work.

“I’m Busy” Is Not a Serious Standard

Most executives overrate the value of their own involvement.

That is understandable.

Your calendar is full.
Your inbox is on fire.
Your Slack never stops.
Everyone wants your input.

But demand is not proof of leverage.

A CEO should not be judged by how many loose ends they personally touch.

They should be judged by whether the company can make smart progress without their constant intervention.

So the standard has to get harder.

Not: did I work hard?
Not: was I responsive?
Not: did I clear a lot of tasks?

The real question is:

That is the audit.

And most executives do not like the answer when they run it honestly.

If you have seen this dynamic inside your own company, I’d be interested in your view in the comments.

Leave a comment

The Productivity Audit

If you want a real diagnosis, stop relying on feeling.

Run an audit.

Take your weekly to-do list and force every item into one of four buckets:

Do - urgent and important
Decide - not urgent and important
Delegate - urgent and not important
Delete - not urgent and not important

The second category matters most.

That is where executive work lives.

That is where you find strategy, judgment, hiring, organisational design, and the thinking work that compounds.

Then do the uncomfortable part:

assign an hourly value to each task.

Not emotionally.
Economically.

Is this $50 per hour work?
$500 per hour work?
$5,000 per hour work?

Most executives discover that too much of their workload sits in the low-value range.

Not because the work does not matter.

Because it should not be done by someone at that level.

That is the difference.

A task can matter and still be the wrong use of executive time.

Use This Prompt

What follows is the practical part: how to stop taking monkeys off your team, how to force completed staff work, and how to delegate with enough precision that your calendar stops filling with work that should never have reached you.

Read more

The Federal Court of Australia’s new practice note on generative AI is nominally about litigation. In practice, it reads like one of the clearest executive briefs yet on how institutions should use GenAI responsibly (Federal Court of Australia, 2026).

That is why CEOs should pay attention.

Not because you are about to draft affidavits.

Because you are already running an organisation where people are using AI to summarise, draft, analyse, recommend, and persuade.

And if this framing is useful, forward it to the person on your team who owns AI rollout, risk, legal, or operations. They will likely recognise the pattern immediately.

I am not a lawyer, and this is not legal advice. But this note is practical, readable, and more useful to executives than many enterprise AI policies I have seen.

Subscribe now

What the Court is actually saying

The Court is not anti-AI. It explicitly says generative AI can increase efficiency, reduce costs, improve access to justice, and enhance the administration of justice. That opening matters. This is not a prohibition document. It is a governance document.

Then it draws a hard line.

Anyone using GenAI in connection with proceedings is expected to understand its capabilities, limitations, and risks. AI use must not adversely affect the administration of justice. And if the Court requires it, a person must disclose if, and how, generative AI was used in a proceeding.

That is a very clean management model:

Use the tool.
Understand the tool.
Do not outsource accountability to the tool.
Disclose material use when it matters.

That, in one line, is where a lot of enterprise AI policy is heading.

The practical parts CEOs should steal

The most useful part of the note is not the principle. It is the specificity.

The Court warns that GenAI can generate fictitious cases, false citations, misleading legal analysis, factual errors, and confident but wrong verification. It then says the person responsible for the document must confirm the underlying facts, authorities, evidence, chronology, and document lists are actually right.

That logic applies far beyond court filings.

Swap legal authorities for market data.
Swap submissions for board papers.
Swap chronologies for incident reports.
Swap affidavits for customer communications.
Swap expert reports for strategy memos.

The operating principle still holds: if AI helped create a document that others will rely on, the named human still owns the truth claim.

That is the sentence I would want every leadership team discussing this quarter. If you agree, share this with the executive who signs off on high-stakes documents in your business.

Share

Where the note is especially strong

The note is sharpest in three places.

1. Verification is not optional

The Court does not treat hallucination as a novelty. It treats it as an operational risk. False material presented confidently is still false material. And responsibility remains with the person who filed it.

For CEOs, this matters because the biggest near-term GenAI failures in companies are rarely spectacular. They are mundane.

A number in a board deck.
A misstated source in a strategy paper.
A fabricated reference in a policy draft.
A persuasive summary that quietly dropped the most important caveat.

This is exactly why AI governance should not be framed only as “tool access” or “approved vendors.” It should also be framed as: what classes of output require human validation before circulation?

2. Disclosure matters when AI shaped the substance

The practice note requires disclosure where GenAI was used to summarise or analyse information relied on by a witness, or to create synthetic media presented to the Court, or in any other way that might affect admissibility or how the Court uses the evidence. It also says that disclosure should appear at the start of the document and explain, concisely, where and how AI was used.

That is more broadly useful than it first appears.

Most companies do not need to disclose every use of AI. They do need a rule for when AI use becomes material.

My view: if AI materially shaped analysis, recommendations, evidence, or representations that others are expected to trust, you need a disclosure standard. Maybe internal. Maybe external. But definitely deliberate.

That is not bureaucracy. It is institutional honesty.

3. Confidentiality is not a footnote

The Court also goes straight at the data issue. It warns that information entered into generally accessible GenAI tools may become available to others, and users may not know where that information is stored, how it is used, or who can access it. It specifically flags confidential, privileged, suppressed, and private information as high risk.

This is where a lot of executive teams are still underweight.

Public GenAI is not just a productivity layer. It is a data-boundary decision.

That means your AI policy should not merely ask, “Is this useful?” It should ask, “Is this information appropriate to enter into this environment at all?”

Those are different questions, and the second one is often the more important one.

This is bigger than Australia

The Federal Court’s note is part of a broader institutional pattern.

In England and Wales, the judiciary’s updated AI guidance says judicial office holders remain personally responsible for everything produced in their name. It also stresses that AI should be used consistently with maintaining the integrity of the administration of justice and the rule of law (Courts and Tribunals Judiciary, 2025).

At the European level, the Council of Europe’s CEPEJ published guidelines on the use of generative AI in the administration of justice, emphasising practical safeguards including human oversight, legal certainty, transparency, data protection, and training (European Commission for the Efficiency of Justice [CEPEJ], 2025).

Also in Europe, the Council of Bars and Law Societies of Europe issued a guide for lawyers on generative AI, focused on opportunity, risk, confidentiality, and professional responsibility (Council of Bars and Law Societies of Europe, 2025).

In the United States, the American Bar Association’s Formal Opinion 512 makes the same core move: generative AI does not replace duties of competence, confidentiality, client communication, and reasonableness in fees (American Bar Association, 2024).

And in Australia itself, the Victorian Law Reform Commission’s 2025 report on AI in courts and tribunals recommended principles, governance settings, guidance, and training, while rejecting AI as a substitute for judicial decision-making (Victorian Law Reform Commission, 2025).

Different jurisdictions. Same direction of travel.

Institutions are not saying, “Do not use AI.”

They are saying, “Use it, but keep responsibility attached.”

That is the signal.

The CEO takeaway

The best line I can give you from all of this is this:

Do not govern AI as software. Govern it as decision infrastructure.

Once you see it that way, the right questions become clearer.

Where can AI help draft?
Where can it summarise?
Where can it analyse?
Where must humans verify?
Where must use be disclosed?
Where is data too sensitive to enter?
Who is the named owner of the output?

That is the real work.

Not whether your teams have access to the latest model.
Whether your institution knows what it is willing to trust, and under what conditions.

If that framing feels useful, send this to a peer CEO or your GC, CIO, or COO. These conversations are becoming board-level questions faster than many teams realise.

Share

Three actions to take this week

First, define which AI-assisted outputs require human verification before they are circulated upward or outward. This should include anything going to the board, regulators, customers, investors, or courts.

Second, create a material-use disclosure rule. Not for every prompt. For outputs where AI materially shaped analysis, representations, or evidence-like content.

Third, redraw your data boundaries. Treat open GenAI tools as environments with explicit information-handling rules, not as harmless drafting assistants.

None of that is anti-innovation.

It is what serious adoption looks like.

Final thought

The most interesting thing about the Federal Court’s new note is not that a court has commented on AI.

It is that one of Australia’s most important institutions has now written down a practical model for using it.

Helpful, but not authoritative.
Fast, but not self-verifying.
Useful, but not consequence-bearing.
That part remains human.

That is true in litigation.

It is also true in management.

And the CEOs who internalise that early will build better AI capability than the ones still treating this as a tooling decision.

References

American Bar Association. (2024, July 29). ABA issues first ethics guidance on a lawyer’s use of AI tools. https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/

Council of Bars and Law Societies of Europe. (2025, October 2). CCBE guide on the use of generative AI by lawyers. https://www.ccbe.eu/fileadmin/speciality_distribution/public/documents/IT_LAW/ITL_Guides_recommendations/EN_ITL_20251002_CCBE-guide-on-the-use-of-the-use-of-generative-AI-for-lawyers.pdf

Courts and Tribunals Judiciary. (2025, October 31). Artificial intelligence (AI) – judicial guidance (October 2025). https://www.judiciary.uk/guidance-and-resources/artificial-intelligence-ai-judicial-guidance-october-2025/

European Commission for the Efficiency of Justice. (2025, December 19). Guidelines on the use of generative artificial intelligence in the administration of justice and by judicial professionals. Council of Europe. https://rm.coe.int/cepej-2025-18final-en-draft-guidelines-on-the-use-of-generative-ai-for/48802a4ad1

Federal Court of Australia. (2026, April 16). Use of generative artificial intelligence practice note (GPN-AI). https://www.fedcourt.gov.au/law-and-practice/practice-documents/practice-notes/gpn-ai

Victorian Law Reform Commission. (2025). Artificial intelligence in Victoria’s courts and tribunals: Report. https://www.lawreform.vic.gov.au/publication/artificial-intelligence-in-victorias-courts-and-tribunals-report/

There is a lazy story in the market right now:

AI writes code.
So SaaS is in trouble.
So enterprise software is about to get flattened.

That story is too crude.

But the comforting response has been too crude as well.

The comforting response says serious enterprise software is safe because it owns “coordination,” “governance,” and “operational depth.” That sounds reassuring. It is also how incumbents always describe the thing that is about to be abstracted.

The real issue is not whether software matters.

It is whether vendor-owned logic should still sit at the centre of the enterprise.

That is the strategic shift.

Not “no software.”

Not “every SaaS company dies.”

But something more destabilising than either camp wants to admit:

the era of the vendor-owned enterprise is ending.

The market has felt that before it has fully explained it. Reuters reported on April 9 that the S&P 500 Software & Services index was down 25.5% year to date. At the same time, the BVP Nasdaq Emerging Cloud Index showed roughly $1.5 trillion in aggregate market cap, 18.6% average revenue growth, and a much lower 5.1x average revenue multiple. That is not software demand disappearing. It is the market charging much less for software whose right to own the workflow is suddenly in doubt.

The spending backdrop says the same thing in a different language. Gartner expects worldwide software spending to reach about $1.43 trillion in 2026, up 14.7%, while data-centre systems spending is expected to grow 31.7% to about $653.4 billion. The point is not that software vanishes. The point is that value is shifting toward the layers that make intelligence and data liquid, while application logic gets repriced.

Every cycle turns a moat into a migration

Every major technology cycle does the same thing.

It takes a former strategic asset and turns it into a rigid liability.

Server rooms were once a moat. Then cloud turned them into stranded complexity.

That is why the “SaaS as the new server room” analogy lands. It forces the right question:

What if fixed workflows, fixed schemas, vendor-controlled roadmaps, trapped data, and painful exits are not signs of software strength, but signs of a delivery model nearing abstraction?

That is not a fringe question. It is the central question.

Because some of what enterprise software has long called “depth” is just dependency with a better account team.

That is also why “revenue is still up” is a weak defense.

Salesforce reported fiscal 2026 revenue of $41.5 billion, up 10%, with current RPO up 16%. ServiceNow reported Q4 2025 subscription revenue up 21%, with cRPO up 25%. Atlassian reported Q2 FY26 revenue up 23%, with RPO up 44%. Those are real numbers. They are also the wrong comfort. They tell you these platforms still have inertia. They do not tell you the old bargain remains strategically attractive. Long contracts, renewal friction, and embedded process debt can keep revenue healthy long after the logic of the model starts decaying.

If you are a CEO, vendor revenue is not the question.

The question is whether you are paying for value creation or financing your own captivity.

The big mistake: treating coordination as mystical

This is where a lot of pro-SaaS analysis still falls apart.

It treats “coordination” as if it is some sacred enterprise art form that AI cannot touch.

That is too romantic.

A large share of coordination inside enterprises is just logic applied to state:

If the threshold is crossed, escalate.
If legal approves, route to finance.
If the contract is redlined, trigger review.
If the ticket has these attributes, hand it to this queue.
If this vendor fails this check, block the workflow.
If this deal exceeds the limit, require a second sign-off.

That is not mystical.

That is logic.

And once intelligence gets cheap enough to generate, adapt, monitor, and rewrite logic quickly, the strategic question becomes brutal:

Why should that logic belong to the vendor?

The traditional model said: let the vendor own the workflow, the data model, the policy surface, and the release cadence. In return, you get convenience, uptime, and someone else carrying the maintenance burden.

That was a reasonable trade when internal software creation was slow and expensive.

It is a much weaker trade when AI compresses the cost of drafting, extending, and reconfiguring business logic.

The evidence points in that direction. In the UK government’s AI coding assistant trial, developers reported average time savings of 56 minutes per working day. In the 2025 Stack Overflow Developer Survey, 84% of respondents were using or planning to use AI tools in development, and 51% of professional developers used them daily. Brynjolfsson, Li, and Raymond found that generative AI increased customer-support productivity by about 14% on average, with larger gains for less-experienced workers. Dell’Acqua, and colleagues found meaningful gains on realistic knowledge-work tasks, while also describing a “jagged technological frontier” where capabilities remain uneven. The point is not that every company has already standardised on a fully autonomous agentic fleet. The point is that the cost curve on logic creation and maintenance is moving sharply against vendor-owned workflow.

Data liquidity is now strategic

This is the point the old SaaS playbook handles worst.

In an AI-heavy enterprise, data becomes more valuable when it is portable, model-readable, recomposable, and easy to route across systems.

But classic SaaS economics were built on the opposite assumption.

Put your data in my system.
Adopt my schema.
Use my workflow.
Access your own operating history through my interface and my API limits.
Then pay me more every year because leaving is painful.

That model worked when software convenience outweighed software captivity.

Now the captivity is more visible.

If your most important operational data lives inside a vendor-defined logic box, then the vendor is no longer just selling software. They are taxing your future optionality.

This is why “ease of exit” matters so much.

Not because easy-exit vendors are abundant in public markets. They are not.

But because CEOs should stop rewarding the opposite.

A vendor whose economics depend on your inability to leave is not aligned with your long-term strategic position. That used to be tolerated. It should not be now.

And no, the answer is not to go on a fantasy hunt for perfect portability. The answer is to build leverage yourself: own the data layer where you can, decouple logic from interface where you can, and stop buying software bundles that assume the vendor should permanently own your operating state.

The public markets are repricing entitlement

The software selloff is not a clean referendum on product quality. It is a messy repricing of several things at once.

Part of it is rates. The 10-year U.S. Treasury yield was 4.29% on April 9, and the Fed’s target range remained 3.5% to 3.75% after its March meeting. Long-duration software equities are structurally vulnerable in that environment.

Part of it is capital rotation. The AI stack is pulling spend toward data centres, compute, model access, and data infrastructure faster than toward classic application layers. Gartner’s category split captures that clearly.

But part of it is a much deeper reassessment:

If AI can increasingly generate and adapt business logic, should customers keep paying premium multiples for software that owns the logic on their behalf?

That is the question markets are starting, clumsily, to ask. And if that is the question, then “revenue is still growing” is not enough of an answer. What is being repriced is not just software. It is vendor entitlement: the entitlement to own the workflow, define the schema, pace the roadmap, and charge more because exit is painful.

The new strategic unit: state, logic, view

This is the architectural pivot that matters.

CEOs should stop thinking of the software stack as a collection of applications.

They should think of it as three layers:

state - where the business truth lives
logic - how decisions, routing, permissions, and workflows operate
view - the interface through which people inspect or act on the system

The old SaaS model bundled all three under vendor control.

That is what is breaking.

The AI-era opportunity is to separate them.

Own the state.

Own or at least govern the logic.

Treat views as replaceable.

This does not mean every company should rebuild ERP from scratch. That would be cargo-cult modernity.

It means the architecture of power is changing.

The more your vendor owns your state and logic, the more they own your tempo.

The more you can decouple state and logic from the UI layer, the more strategic room you create for yourself.

That is the real reason “AI-native” is such a weak buying criterion. A vendor can be AI-native and still be building a better cage.

The board question is not “buy or build”

It is tempting to turn this into a classic buy-versus-build fight.

That is too simple.

The real board-level question is:

What must the enterprise own to preserve strategic freedom as intelligence gets cheaper?

That usually points to a different answer than either SaaS marketing or engineering bravado wants.

You probably do not need to own every interface.

But you do need far more control over:

your core data exhaust
the logic that routes work and approvals
your ability to swap interfaces and model providers
your ability to extract, inspect, and recompose operational state
the policy layer that governs automation at scale

That is what logic sovereignty looks like in practice.

Not a macho internal-tools fantasy.

A deliberate reduction in vendor control over your operating model.

And that matters because one of SaaS’s oldest arguments is getting weaker: the ownership scare tactic.

For twenty years, SaaS vendors sold “maintenance insurance.” They charged venture-grade and then public-market-grade multiples because hiring dozens of engineers to maintain internal business software was genuinely painful.

In 2026, that insurance premium looks much less defensible.

The same AI wave that pressures SaaS pricing also reduces the cost of a meaningful share of internal software upkeep. The UK trial, the Stack Overflow survey, the NBER field evidence, and the Dell’Acqua et al. study all point in the same direction: AI improves productivity on real software and knowledge tasks, often most for the less-experienced workers who used to make internal capability hard to scale. That does not erase maintenance cost. But it does erode one of SaaS’s oldest structural advantages. If 50 engineers increasingly look like 5 engineers with agentic IDEs and strong internal tooling, the old maintenance-insurance premium starts to look less like prudence and more like an overpriced relic.

What’s coming up next

Below, I get even more operational:

• which software categories CEOs should target first for budget cannibalisation

• how to spot fake portability in a sales process

• what to ask vendors about state ownership, API access, logic escrow, and exit rights

• where AI infrastructure should be funded from inside the existing software budget

• the simple scorecard I’d use before every major renewal

Read more

Your highest performer is not automatically your next leader.

That assumption destroys value in more companies than most org charts reveal.

A top salesperson beats quota for three years, so they become Head of Sales.

A brilliant engineer keeps rescuing critical projects, so they become Engineering Manager.

A reliable operator becomes the team lead because everyone trusts them.

The logic feels obvious.

The outcome usually is not.

Because the skills that create elite individual performance are often different from the skills required to lead a team.

Sometimes radically different.

And when companies confuse the two, they make the same mistake twice:

they remove excellence from the role where it created the most value
then they install someone unprepared into a role with broader consequences

That is not promotion.

It is organisational self-harm.

The Peter Principle Was Never Really About Incompetence

The Peter Principle is usually quoted as a joke:

But the mechanism is serious.

People are promoted based on performance in Role A into Role B, even when Role B rewards a completely different set of capabilities.

So the company is not actually promoting proven aptitude.

It is promoting hopeful inference.

That is how you get:

great closers who cannot coach
great engineers who cannot delegate
great executors who become bottlenecks
great individual contributors who make fragile managers

The issue is not intelligence.

The issue is mismatch.

The person is still talented.

They are just now sitting in a role designed for a different kind of leverage.

Why Companies Keep Making This Mistake

Because most companies still use promotion as reward.

That is the root error.

Promotion should not answer the question: Who deserves recognition?

It should answer the question: Where will this person create the most leverage next?

Those are not the same thing.

Reward is backward-looking.

Role design is forward-looking.

One recognises past performance.

The other allocates future responsibility.

When leaders blur the two, they turn promotions into symbolic compensation.

That feels generous in the moment.

It becomes expensive later.

Because management is not a gold star for top performers.

It is a different job.

And different jobs should be filled based on demonstrated fit, not sentiment.

The Skills Barely Overlap

This is where most leadership teams get careless.

They assume “senior” naturally evolves into “manager.”

But many of the traits that make someone exceptional as an individual contributor can become liabilities in management.

A high-performing individual contributor often wins through:

speed
autonomy
craft
control
personal standards
depth of execution

A strong manager wins through:

delegation
judgment
coaching
conflict navigation
team clarity
system design

One role creates value through direct output.

The other creates value through multiplied output.

Those are different engines.

Sometimes opposing ones.

The perfectionist who insists on high standards may become a chronic micromanager.

The hero operator who always jumps in to fix things may prevent the team from growing.

The introverted specialist who does brilliant work independently may avoid the difficult conversations management requires.

The top performer is not failing because they became less capable.

They are failing because the company changed the game without checking whether their strengths match the new rules.

The Real Cost of the Wrong Promotion

Most organisations understate this cost because they only look at the promoted person.

That is too narrow.

A wrong promotion creates damage in at least four places, you:

1. lose output. The person stops doing the high-value work they were uniquely strong at.

2. weaken management quality. The team now reports to someone who may not know how to coach, prioritise, or lead through others.

3. create organisational drag. Decisions slow down. Delegation becomes inconsistent. Standards become personal instead of systemic.

4. damage morale. The promoted person feels overwhelmed. The team feels unsupported. Everyone becomes less candid because the situation is awkward to name.

This is why a bad promotion is rarely a contained mistake.

It spreads.

And because leadership teams often wait too long to correct it, the cost compounds quietly before it becomes visible.

Stop Promoting for Reward. Promote for Capability.

This is the central shift.

Do not ask whether someone has earned the promotion.

Ask whether they have shown evidence of succeeding in the role.

That means looking beyond output.

You need to assess how they handle:

delegation
ambiguity
peer coordination
feedback conversations
conflict
decision-making through others

Most companies never do this rigorously.

They promote the standout performer, then hope the rest can be learned on the job.

Sometimes that works.

Often it does not.

And hope is a weak operating model.

A better approach is to test management aptitude before assigning management authority.

Not through personality theory.

Through work sample.

Simulation.

Observed behaviour under managerial conditions.

That is where AI can actually be useful.

Not as decision-maker.

As diagnostic partner.

The Promotion Audit

Before promoting someone from high-performing IC to manager, pressure-test the move.

Use this prompt:

This matters because management failure is usually visible in behaviour long before it is visible in performance metrics.

The signals show up early:

they keep taking work back
they avoid hard conversations
they solve instead of coaching
they become the approval bottleneck
they confuse personal excellence with team leadership

A pre-promotion audit makes those risks discussable before the title makes them political.

The Deeper Structural Problem

Most promotion failures are not really individual mistakes.

They are system design failures.

Because in many companies, the only path up is management.

More money means direct reports.

More prestige means people leadership.

More career progression means moving away from the work you are best at.

That system forces the wrong trade-off.

Brilliant builders become reluctant managers because management is the only recognised path to status, compensation, and influence.

Then companies act surprised when both performance and energy decline.

The issue is not ambition.

It is architecture.

If you make management the only path to progression, you will overproduce managers and underprotect expertise.

That is not a talent problem.

It is an incentive problem.

Most companies can see the symptom.

Far fewer redesign the system that keeps producing it.

What follows is the practical part:

how to build a dual-track career ladder that allows top performers to advance without managing people, and how to unwind the wrong promotion gracefully when the mistake has already been made.

The Only Way Up Cannot Be Management

Read more

A one-person AI newsroom launched, published at industrial scale, attracted tens of thousands of visitors, and shut down less than a month later.

That arc matters more than the novelty.

Because it shows both halves of the generative AI story at once. A single operator can now simulate the output layer of an institution. But technical scale arrives much faster than trust, governance, and business viability.

That is why The Daily Perspective matters.

For 29 days, it ran as an AI-powered Australian news site publishing around the clock across politics, business, world, sport, technology, climate, crime, culture and more. According to its own retrospective, it used 33 AI editorial personas, ran on an automated Cloudflare-based pipeline, and published at a volume that would once have implied a real newsroom behind it.

Then it stopped.

For CEOs, that is the signal.

The important question is no longer whether AI can write an article. It can. The more important question is what happens when one builder can create something that looks, feels, and operates like a publication, a research desk, or an analysis team.

Not a demo.
Not a toy workflow.

A synthetic institution.

This is not really a media story

Most AI coverage still treats the technology as a tool question.

Which model.
Which assistant.
Which use case.
Which productivity gain.

That framing is already too narrow.

What matters now is more structural. Generative AI is starting to let a single operator simulate the outward form of an institution. Not just a task. Not just a workflow. An institution-shaped output layer.

That is what this case demonstrated.

Once you strip away the publishing wrapper, the pattern is clear: retrieval, synthesis, routing, policy constraints, automated checks, and low-cost distribution at scale.

That pattern does not stop at media.

It applies anywhere work can be broken into inputs, rules, outputs, and delivery.

That is why this matters beyond journalism.

Share

The real shift is task-stack redesign

For years, organisational breadth required organisational weight.

If you wanted specialist coverage, segmented output, continuity of voice, and high-frequency delivery across multiple categories, you hired teams. That was the cost of institutional surface area.

Increasingly, in some functions, you can now approximate parts of that through system design.

Not perfectly.
Not safely by default.
Not without risk.

But much more cheaply than before.

That shifts the leadership problem.

The question is no longer simply whether AI can help a function move faster. The question is whether whole bundles of work can now be decomposed, rebuilt, and recombined into smaller human-plus-machine operating models.

That is a different level of change.

A newsroom bundles monitoring, sourcing, verification, drafting, editing, packaging, distribution, corrections, and voice consistency. This project appears to have rebuilt enough of that stack in software to create a credible publication surface.

That does not mean it recreated a newsroom in the fullest sense.

It means it recreated enough of the output layer to force a more serious executive question:

Which parts of your own organisation now look more like software design problems than fixed team structures?

Output is not the same as institution

This is where the shutdown matters more than the build.

The project proved that a single operator can create a surprising amount of institutional surface area:

breadth
continuity
specialist bylines
publication cadence
editorial segmentation
multi-channel distribution

What it did not prove is that those things are enough.

Because institutions are not just output machines. They also contain judgment, accountability, legitimacy, trust, and economic structure.

That is the distinction many leadership teams are still underestimating.

AI can help create the appearance, and sometimes the operational reality, of greater capacity than headcount once allowed. But institutional substance does not come bundled with output.

You still have to build the rest.

Governance.
Review architecture.
Legal defensibility.
Correction mechanisms.
Trust signals.
Commercial viability.

The easier it becomes to generate surface area, the more important it becomes to ask what is underneath it.

Subscribe now

The economics are part of the lesson

One of the most useful details in the site’s own postmortem was not technical.

It was economic.

Over five days of display advertising, it reported earning a total of $0.59.

That number matters because it captures a pattern that is about to show up across many categories of AI deployment: the cost of production can collapse long before the economics of the business do.

This is one of the easiest mistakes in AI strategy.

Teams see that content can be produced faster. Reports can be generated more cheaply. Coverage can be broadened. Analysis can be scaled.

All of that may be true.

But lower production cost does not automatically create distribution, demand, trust, margin, or defensibility.

In many cases, it simply creates more output chasing the same scarce inputs: attention, credibility, customer willingness to pay, and platform tolerance.

AI can improve the production function without solving the business model.

Sometimes it exposes the weakness of the business model faster.

Accountability does not disappear. It concentrates

This is the governance lesson many organisations are still underestimating.

If an AI-generated article is wrong, misleading, defamatory, or recklessly framed, accountability does not sit with a fictional byline or a model endpoint. It sits with the humans and organisations behind the system.

The person who designed the logic.
The person who chose the sources.
The person who set the thresholds.
The person who decided human review was unnecessary.
The organisation that published the output.

That principle travels well beyond media.

As AI moves deeper into customer-facing, regulator-facing, investor-facing, and public-facing work, responsibility does not vanish. It hardens.

The model can generate the recommendation.
The human signs the decision.
The organisation carries the liability.
That is not a technical detail.

It is an operating-model issue.

Synthetic scale is becoming a strategic variable

AI does not just change cost structures.

It changes market optics.

A system like this can create the impression of breadth, specialist capability, continuity, and institutional depth. Some of that may be genuine leverage. Some of it may be synthetic theatre. The point is that the surface area can now be created much more cheaply than before.

That has obvious implications outside publishing.

Small companies can look larger.
Niche organisations can look broader.
A single operator can create the outward form of a specialised function.
An AI vendor can appear more mature than its underlying operating model really is.

This is where the CEO lens needs to sharpen.

The question is no longer only what the model can produce.

The more important question is what has actually been built around it.

Because durable advantage rarely comes from the model alone. It comes from the surrounding system: workflow design, review architecture, escalation rules, trust signals, and economic discipline.

That is what determines whether AI creates real leverage or merely manufactures synthetic scale.

The strategic takeaway

The headline is not that AI can write the news.

The headline is that AI can now help a single operator manufacture enough coherence, continuity, and specialisation to look institutional, and that many companies will mistake that surface area for durable capability.

That is the real lesson.

The competitive divide will not simply be between companies that adopt AI and companies that do not.

It will be between companies that use AI to create real operational leverage, and companies that use it to create synthetic scale without the controls, trust, and economics to support it.

The companies that get this right will not just produce more with fewer people.

They will redesign work deliberately.
They will govern AI where it matters.
They will know where accountability sits.
They will build systems whose control layer scales with their output layer.

The companies that get it wrong will look efficient right up until something breaks.

And when it does, the question will not be whether the model made the mistake.

It will be who decided to trust it.

Leave a comment

The harder leadership skill now is not approving more AI experiments. It is knowing what to stop, before weak economics, vague ownership, or synthetic scale turn into expensive drag.

Which AI projects or pilots have you stopped, and why?

Generative AI is already changing workforce decisions.

But not in the way most headlines suggest.

The public debate is still framed too bluntly. On one side sits the claim that AI is already replacing large numbers of workers. On the other sits the false comfort that, because broad labour-market data remain relatively calm, little of consequence has changed.

Neither view is much use to a CEO.

The stronger reading is narrower, and more practical.

The strongest public labour-market data still do not show broad AI-driven disruption. The Budget Lab at Yale says the January and February 2026 CPS releases do not indicate a meaningful shift in most categories, and its broader tracker still points to no meaningful broad impact of AI on the labour market in those public data. At the same time, OECD data show enterprise AI adoption continuing to rise, with 20.2% of organisations in OECD countries with available data reporting AI use in 2025, up from 14.2% in 2024 and 8.7% in 2023.

Those two facts are not contradictory.

They suggest that organisational change is moving ahead of broad labour-market confirmation.

That matters because the first real effects of AI are more likely to show up inside workflows, in management structure, in hiring mix, in capital allocation, and in the design of junior pathways before they show up clearly in macro labour data.

The most useful way to understand that shift is through three channels of workforce change.

1. Task redesign

The first channel is task redesign.

AI does not need to replace a role in full to change workforce decisions. It only needs to absorb or accelerate enough repeatable tasks inside a workflow to change how that workflow is staffed, reviewed, and handed off.

That is what the task-based research points to. The ILO’s refined exposure analysis treats occupations as bundles of tasks rather than indivisible jobs. Microsoft Research’s work on real-world AI use found that common work activities include gathering information and writing, while common AI-performed activities include providing information and assistance, writing, teaching and advising.

That is the first correction most executive teams should make.

The meaningful unit is not the job title.
It is the task bundle inside the workflow.

2. Management-layer and skill-mix change

The second channel is management-layer and skill-mix change.

Once enough tasks inside a workflow are accelerated or absorbed, organisations can start changing the composition of roles around that workflow. That can mean fewer coordinators, flatter structures, altered management spans, and different expectations of junior and mid-level staff.

This is one reason the labour story still looks confusing in public. Broad unemployment can remain relatively stable while internal role design changes materially.

Brookings’ March 2026 review points to younger workers, transitions and task change as more plausible early channels than broad unemployment alone. The World Bank’s labour-demand paper points in a similar direction, finding that by mid-2025 job postings in occupations with above-median AI substitution scores had declined by about 12% relative to those below median. That does not prove broad displacement. It does support the view that hiring pressure and transition effects may appear before aggregate job counts move clearly.

3. Capital reallocation linked to AI

The third channel is capital reallocation linked to AI.

Some workforce decisions are not best understood as direct automation at all. They are better understood as labour-cost reduction and reallocation to fund AI products, infrastructure, cloud capacity, data programmes, or related strategic priorities.

This matters because many “AI layoffs” are actually mixed events.

They may reflect some task redesign, some organisational restructuring, some pressure to fund AI investment, and some pressure to demonstrate future AI leverage. OECD’s enterprise adoption work supports this framing because it treats AI adoption as dependent on complementary investment, organisational capability, data readiness and diffusion. In practice, AI can affect workforce decisions through funding logic as much as through direct substitution.

The company cases are real, but they are not the same story

This is where the commentary often becomes sloppy.

Block, Atlassian and Oracle should not be thrown into one bucket called “AI sackings”.

They do not show the same thing.

Block is the clearest case of leadership explicitly linking a smaller workforce and flatter structure to AI-enabled productivity and organisational redesign. Reuters reported that Block announced more than 4,000 job cuts while Jack Dorsey said a significantly smaller team using the tools the company is building could do more and do it better. Subsequent reporting described a broader AI-led redesign with compressed management layers and “player-coaches”.

Atlassian is more mixed, and more useful for that reason. Atlassian explicitly said its approach is not “AI replaces people”, but also said it would be disingenuous to pretend AI does not change the mix of skills needed or the number of roles required in some areas. That makes it a case of AI-linked adaptation and skill-mix change, not a clean substitution story.

Oracle is different again. Reuters reported that Oracle had begun laying off thousands while increasing investment in AI infrastructure and estimating up to $2.1 billion in restructuring costs for fiscal 2026, mostly from severance. Oracle did not publicly claim that those cuts were caused by realised AI productivity. This looks more like capital reallocation and cost restructuring linked to AI priorities than proof that AI has already made those roles redundant.

The lesson is not that one of these companies is right and the others are wrong.

The lesson is that AI is already influencing workforce decisions through different mechanisms, and executive teams need a framework that can distinguish between them.

What matters now

The current picture is narrower, and more useful, than most public commentary suggests.

Broad labour-market data still do not support a mass AI displacement story. But that does not mean organisations are standing still. What matters now is that workforce decisions are already shifting through task redesign, management-layer and skill-mix change, and capital reallocation linked to AI. Public labour-market data are still too blunt and too delayed to guide leadership decisions on their own. The more useful question is where organisational change is already visible through workflows, role design, management structure, hiring mix, and capital allocation. That is the signal leadership teams need to read now.

The junior pipeline is still the most plausible early pressure point

If there is one place leadership teams should look harder, it is the junior pipeline.

Entry-level and lower-mid-level roles often contain the highest concentration of repeatable information-gathering, drafting, formatting, reconciliation, coordination and support tasks that AI can first absorb or compress.

If enough of those tasks move, organisations can slow hiring, raise the threshold for what counts as entry-level work, redesign junior roles around validation, exception handling and stakeholder support, and weaken apprenticeship pathways without making large incumbent cuts.

That does not require broad immediate displacement.

It only requires enough workflow change to alter the economics of the pathway.

This is one reason the most important AI workforce risk is not simply job loss. It is the possibility that organisations improve short-term efficiency while quietly weakening their future capability base. Brookings’ review and the World Bank’s labour-demand evidence both support the view that hiring and transitions deserve close attention here.

Paid subscribers can read the full executive analysis below

In the rest of this briefing, I cover:

• why work design changes before broad employment data

• why task decomposition matters more than job labels

• how to classify Block, Atlassian and Oracle without collapsing them into the same story

• why AI-cited cuts are relevant, but should remain supporting evidence

• the executive dashboard I use to track AI-linked workforce change inside organisations

Read more

Your spreadsheet is a comfort blanket

Revenue up 15%.

Margin holds.
Hiring stays controlled.
AI investment starts paying back.
Execution stays on plan.

It looks sensible.

That is what makes it dangerous.

Because it assumes the world stays broadly cooperative.

It won’t.

There is a war on.
Energy is unstable.
Shipping is exposed.
Customers are slower.
Capital is tighter.
AI is moving faster than most operating models can absorb.

And yet many leadership teams are still planning as if next quarter is just a cleaner version of last quarter.

That is not strategy.

That is extrapolation.

Linear growth is a myth

Forecasting works when the range of outcomes is narrow.

This is not that environment.

Right now, one geopolitical event can hit costs, demand, freight, pricing, and confidence at the same time.

So the problem is not that your forecast will be slightly wrong.

The problem is that it is built on the wrong premise.

That the system will keep behaving.

The question most plans avoid

Most strategy decks ask:

What happens if things go right?

Wrong question.

The real one is:

What happens if the world turns against our plan?

Not in theory.

In reality.

War expands.
Energy spikes.
Buying slows.
AI adoption fragments.
Costs stay high.
Execution gets harder.

Does your current strategy still work?

If not, it is not a strategy.

It is a fair-weather budget.

The tool

Use a 2x2 Scenario Matrix.

Pick the two uncertainties that matter most.

For most businesses right now, they are some version of these:

1. AI adoption accelerates vs stalls
2. Operating conditions stabilise vs deteriorate

That gives you four worlds.

Acceleration
AI moves fast. Conditions hold. Speed wins.

Extraction
AI moves fast. Conditions deteriorate. Productivity matters. Waste gets punished.

Delay
AI stalls. Conditions hold. Pilots continue. Payoff drifts.

Crunch
AI stalls. Conditions deteriorate. Budgets tighten. Fragility gets exposed.

Now do the only part that matters.

Take your current strategy and drop it into the worst quadrant.

Then ask:

Does it still protect cash?
Does it still reduce dependency?
Does it still defend customers?
Does it still improve our position?

If not, you have found the weakness.

Before the market does.

What strong leadership does differently

Amateurs bet on the most likely scenario.

Professionals make no-regret moves.

Decisions that work in all four worlds.

Higher talent density.
Lower vendor dependency.
Better cash discipline.
Shorter execution cycles.
Fewer fragile initiatives.
More room to pivot.

That is what strategy looks like in uncertain times.

Not prediction.

Positioning.

Prompt: the core scenario stress test

Read more

The uncomfortable starting point

Most organisations have more layers than they need.

They just haven’t had to confront it.

Until now.

Earlier this year, a layer disappeared

Earlier this year, I wrote that AI is collapsing the layer between data and executive decision-making.

Executives are getting closer to raw signal.

Fewer dashboards.
Fewer summaries.
Fewer intermediaries.

That shift is already underway.

This is the next layer

Now the same thing is happening lower down.

Not at the top of the organisation.

In the middle.

What the data actually shows

The most important AI insight from the past 12 months is this:

Value is not coming from better models.

It’s coming from redesigning work.

A 2025 McKinsey study found that organisations seeing real impact from AI are reimagining end-to-end workflows, not just automating tasks.

More specifically:

Companies that redesign workflows are significantly more likely to see financial impact.
High-performing organisations are changing how work flows, not just adding tools.

This is the shift most teams are underestimating.

What’s already changing inside work

AI is already reducing time spent on coordination-heavy tasks.

A large-scale field experiment (6,000 workers) found:

Around 25% less time spent on email.
Faster document production.
Limited change to meetings, for now.

That matters.

Because email, reporting, and document prep are not edge work.

They are coordination work.

This is already visible inside companies

In consulting firms like McKinsey, BCG, Deloitte, and PwC:

Internal AI tools are being used across large portions of the workforce.
Tasks like research, summarisation, and document prep are being automated.
Teams are saving meaningful time on routine work.

The work hasn’t disappeared.

But the effort required to coordinate it has.

What the middle layer actually does

Most middle layers exist to:

Move information
Translate work
Aggregate updates
Coordinate teams

That work was necessary.

Because coordination was expensive.

That assumption just broke

AI reduces the cost of coordination.

Not to zero.

But enough to matter.

Summaries are generated instantly.
Reports are drafted automatically.
Context is surfaced without manual effort.

The work still exists.

But less human effort is required to move it.

What’s happening now

This is not delayering.

Not yet.

It’s hollowing.

The structure remains.

But less work is required inside it.

Most organisations won’t act on this yet.

They will:

Add AI tools
Increase output
Improve reporting

But keep the same structure.

Until the inefficiency becomes visible.

Too many layers.
Too much coordination.
Too little value added per role.

Below is the diagnostic to identify which parts of your organisation are now under pressure, and what to do about it before it becomes a cost problem.

Read more

1. Must-Haves: If missing, they hate you. If present, they feel nothing (e.g., a hotel having hot water).

2. Performance: More is better (e.g., battery life).

3. Delighters: Unexpected magic.

You are likely spending 80% of your budget polishing “Must-Haves” that are already good enough. Stop it. You get zero credit for making hot water hotter.

The Prompt

Act as a Product Strategy Consultant (Kano Model framework).

Here are the top 5 initiatives on our roadmap: [Paste List].

1. The Classification: Categorise each into:
- Must-Have (Threshold).
- Performance (Linear).
- Delighter (Exciter).
2. The Audit: Are we over-investing in a "Must-Have" that is already at market parity? (If yes, mark it for a budget cut).
3. The Gap: Do we have at least one "Delighter" that differentiates us from [Competitor], or are we just playing catch-up?

Share

The Executive Upgrade: Subscribers Only

Feature Creep” & “The Bloat.
Engineers and Designers love to perfect things. They will spend 3 months optimising a “Settings” page that 1% of users visit.
As an executive, your job is to be the Editor.
You must ruthlessly cut “Gold Plating” (perfecting the irrelevant) to fund the “Magic” (the features that sell the product).

Here are the two Governance Protocols to clean up the product.

Read more

The narrative is wrong

The dominant story is simple.

AI replaces jobs.

That makes for good headlines.

It’s not what’s actually happening.

Not yet.

What we said last year

Last year, I wrote that AI would change how we work before it changes who works.

That’s now playing out.

Roles are still there.

But the expectations inside those roles are shifting much faster than most teams realise.

If you haven’t read it, this was the argument in full:

The signal most people missed

When Commonwealth Bank announced its AI workforce strategy, the headlines focused on job cuts.

~300 roles removed.

That’s the visible part.

The more important move was the $90 million investment behind it.

A program to redesign how skills are built, how roles evolve, and how people move inside the organisation:

That is not a redundancy program.

It is a redesign of work.

I unpacked that shift in more detail here:

This is what AI adoption actually looks like

CBA is not alone.

Across industries, the pattern is consistent.

AI is not being deployed to remove entire roles.

It is being deployed to change how those roles operate.

At CBA:

Skills are being mapped dynamically

Careers are being redesigned around capabilities, not job titles

Work is being restructured to fit AI, not the other way around

This is the shift most companies are underestimating.

The bar moves

When AI increases output, expectations follow.

Quietly.

A role doesn’t disappear.

But what “good” looks like inside it changes.

Faster turnaround
Higher output
Better decisions

This is already visible inside teams.

And it compounds.

The tension leaders are starting to feel

This is not just theoretical.

You can hear it directly from leadership.

Commonwealth Bank CEO Matt Comyn put it clearly last week:

“This technology needs to lead to productivity and improvements in living standards, not valuations.”

That distinction matters.

Because right now, the market is rewarding the expectation of AI.

Not the realised output.

Most organisations are still being measured on the first.

While trying to build the second.

And inside organisations, that creates tension.

Performance expectations are rising.

But the systems, skills, and structures to support them are still catching up.

That gap is where most teams are operating today.

This is where the real impact sits

AI is not compressing headcount evenly.

It is widening performance.

The gap between:

People who can work with AI
People who cannot

Is increasing.

And it is becoming visible.

Most leadership teams are still asking the wrong question.

“How many jobs will AI remove?”

The better question is:

“What level of performance will we now expect from every role?”

Because this is where the real organisational change happens.

Below are the structural shifts already playing out inside teams, and the decisions leaders need to make now to stay ahead of them.

Read more

The shift already happened

Most leaders still talk about AI like it’s new.

It isn’t.

It’s already embedded in products, workflows, and customer experiences.

Which means the conversation has changed.

This is no longer about what AI can do.

It’s about what it costs.

The signal most people missed

The most important AI move this year wasn’t a model release.

It was Microsoft making it clear it doesn’t want to depend on a single model provider.

Multiple models inside Copilot.

Internal models moving into production.

Azure positioned as a marketplace, not a pipeline.

That’s not innovation.

That’s cost control.

AI is now a line item

This is already showing up in real companies.

Klarna didn’t just use AI to increase productivity.

It used it to remove cost.

$10 million saved in marketing.

Thousands of assets generated internally.

That’s not a feature.

That’s margin.

At the same time, companies like Atlassian and Block are restructuring teams and cost bases around AI.

Not because the technology is perfect.

Because the market is already pricing in productivity gains.

The expectation is set.

Now the cost structure has to follow.

The mistake most companies are making

Most AI strategies still sound like this:

“Which model is best?”
“How fast can we deploy?”

That’s the wrong question now.

Because once AI is embedded in your product:

Every prompt has a cost.
Every workflow has a cost.
Every customer interaction has a cost.

And those costs scale with usage.

Where this hits first

You won’t see this in demos.

You’ll see it in:

Gross margin
Customer acquisition cost
Cost to serve

Because once AI is part of your product, it stops being optional.

It becomes structural.

Most teams are still optimising for capability.

They are not optimising for cost.

And that creates a predictable problem.

The companies moving fastest today may be locking in the worst economics tomorrow.

Below are the three cost traps already forming inside enterprise AI strategies, and how leading teams are starting to avoid them.

Read more

The number everyone trusts

Return on Equity.

It shows up in every board pack.

It trends up, and the conversation gets easier.

Performance looks strong.
Capital looks efficient.
Strategy feels validated.

But ROE doesn’t tell you what’s actually happening.

It compresses three very different things into one number.

How much you make
How fast you move
How much you borrow

And those are not the same.

The problem with ROE

A company can improve ROE in two fundamentally different ways.

The first is real.

Higher margins
Better pricing
Faster asset utilisation

The second is cosmetic.

More leverage
More debt
More financial engineering

Both increase ROE.

Only one improves the business.

The other increases risk.

The model that exposes it

The DuPont framework breaks ROE into three components.

Net Margin
Your pricing power

Asset Turnover
Your speed

Financial Leverage
Your dependence on debt

This is not academic.

It is diagnostic.

Because once you separate these drivers, the story becomes clear.

Where things quietly go wrong

Most leadership teams don’t improve all three.

They optimise one.

And compensate with another.

The pattern is predictable.

Margins flatten or decline
Asset turnover slows
Leverage increases

ROE holds steady.

Or improves.

On paper, nothing looks wrong.

In reality, the business is becoming more fragile.

The early warning signal

If your ROE is rising while margins are falling, you are not growing.

You are borrowing time.

The business is working harder for less return.

And using leverage to hide it.

That works.

Until it doesn’t.

Most teams stop at the number.

They don’t break it down.

And that’s where the mistake is made.

Because ROE doesn’t just tell you how you’re performing.

It tells you how you’re exposed.

The next step is not analysis.

It is a forced choice.

Below are the structural traps that destroy return quality, and the decision frameworks leadership teams use to fix them.

Read more

Content is no longer the constraint

AI has quietly removed one of the biggest constraints in marketing.

Content.

What used to take weeks now takes days.

What used to take teams now takes individuals.

Zalando has reduced campaign production time from 6–8 weeks to 3–4 days, cutting costs by around 90%, with AI generating the majority of campaign imagery.

Unilever has reduced product content production costs by more than 50% using AI.

Klarna has generated thousands of assets while cutting marketing costs by $10 million annually.

Production is no longer scarce.

Retail is already operating at AI speed

In retail, this shift is no longer theoretical.

H&M is using AI-generated “digital twins” of models to create campaign assets without traditional shoots.

At the same time, large-scale retail experiments show: → 0% to 16.3% increases in sales, depending on execution quality

And not all outcomes are positive.

AI-generated campaigns from major brands have faced backlash for being “soulless” or inauthentic.

The pattern is consistent:

AI is making content faster and cheaper.

But not consistently better.

But performance isn’t scaling with volume

Here’s the part most leadership teams are missing.

More content does not automatically mean better performance.

Across organisations:

productivity is increasing
output is scaling
costs are falling

But outcomes are uneven.

Deloitte’s 2025 analysis is clear:

→ AI returns remain difficult to measure and inconsistent across organisations

Even in advertising:

some campaigns improve significantly
others show only marginal gains

AI increases supply.

It does not guarantee quality.

The new bottleneck is taste

When supply becomes infinite, selection becomes scarce.

This is the shift.

Brands can now generate:

hundreds of ad variations
thousands of product descriptions
personalised campaigns at scale

But only a fraction of that output actually:

converts
resonates
strengthens the brand

The advantage moves from creation … to selection.

The risk isn’t that your team produces bad content.

It’s that they produce too much average content and mistake activity for progress.

That’s where most AI strategies are already drifting.

Most AI strategies look right on slides.

They fail in execution.

Because the hard part isn’t the model.

It’s knowing:

where it actually changes your economics
where it introduces risk
where it creates advantage

If you’re using AI to make real decisions, not just follow the conversation, the rest of this briefing goes deeper on this.

Read more

Blaming AI Is Like Blaming Excel

“AI is great, except when it gets everything wrong all the time”

That is the headline of a recent column.

The argument rests on a simple experience.

A user asked AI to identify a band from a partial description.

The system failed.

From that anecdote comes a broader conclusion:

“AI is great, except when it gets everything wrong all the time.”

This is an understandable reaction.

It is also a misunderstanding.

Not because AI never gets things wrong.

It clearly does.

But because the example reveals something more important than the technology’s limits.

It reveals how people misunderstand the tool.

The Excel Problem

Imagine opening Excel for the first time.

You type a formula incorrectly.

The spreadsheet produces the wrong answer.

You then conclude spreadsheets are unreliable.

That sounds absurd.

But it is essentially the same reasoning that often appears in commentary about generative AI.

These systems are not encyclopaedias.

They are probabilistic reasoning engines.

And like any tool, the quality of the output depends heavily on the skill of the person using it.

AI Is Not Google

The column describes using AI the way many casual users do:

as a replacement for search.

That works sometimes.

But it is not where the technology creates its greatest value.

Generative systems are far more powerful when used to:

structure analysis

draft documents

generate code

summarise complex material

iterate ideas rapidly

Used this way, they compress hours of work into minutes.

That is why people who use AI every day often report dramatically different experiences from those who try it occasionally.

The tool has not changed.

The workflow has.

The Real Constraint

In the column, the writer describes asking the system why it failed to identify the band.

The AI suggests the problem might be the description.

The writer responds that a human would have asked that question earlier.

That observation is actually the key insight.

Human expertise matters.

Because generative AI does not know when it is wrong.

It produces outputs that are statistically plausible.

Which means something important follows.

If the human operator cannot evaluate the answer, the system becomes unreliable.

Not because the model failed.

Because the supervision failed.

The Leadership Blind Spot

This is where the issue becomes strategic.

Many organisations talk about “human-in-the-loop” AI systems.

But the phrase is often misunderstood.

The human role is not symbolic oversight.

It is expert judgment.

If the person reviewing the output lacks domain knowledge, the loop is effectively broken.

The system is operating without meaningful supervision.

That is the real operational risk.

AI Amplifies Judgment

There is a historical precedent.

When calculators appeared in classrooms, critics argued they would destroy mathematical understanding.

The opposite happened.

Students who understood the concepts moved faster.

Students who did not still struggled.

Calculators did not eliminate expertise.

They exposed it.

Generative AI works the same way.

It amplifies judgment.

And it amplifies mistakes.

At exactly the same speed.

Subscribe now

The Question Leaders Should Ask

So when someone concludes that AI “gets everything wrong all the time,” the response should not be defensive.

It should be analytical.

What exactly failed?

The model?

Or the workflow surrounding it?

Because generative AI does not democratise expertise.

It rewards it.

For CEOs deploying these systems across their organisations, that distinction matters.

The most important question is not whether the technology works.

It is whether the people supervising it can recognise when it doesn’t.

If they cannot, the problem is not the AI system.

It is the operating model.

And AI will scale that weakness faster than any previous technology.

The Prompt

Act as a Project Management Auditor (Apple DRI style).

Here is a list of our current strategic initiatives and the "Teams" assigned to them:
[Paste List, e.g., "Website Relaunch - Marketing Team"].

1. The Audit: Flag every initiative that lacks a single human name attached to it.
2. The Assignment: Define the specific scope of the DRI for each project. (They don't do all the work, but they make all the decisions).
3. The Conflict: If two people are listed as "Co-Leads," explain why this will fail and force me to pick one.

Share

The Executive Upgrade: Subscribers Only

Consensus vs. Consultation.
Assigning a DRI scares people. They think it means they can’t ask for help.
You need to clarify decision rights.
The biggest bottleneck in your company is people waiting for you to decide, because they don’t know they are allowed to.
We use Bain’s RAPID framework (Recommend, Agree, Perform, Input, Decide) to fix this.

Here are the two Governance Protocols to clear the bottleneck.

Read more

The Pressure Is Showing Up in Headcount

Two recent announcements hint at a shift happening inside the technology sector.

Atlassian announced plans to cut roughly 10% of its workforce as it pivots toward AI and enterprise sales.

Block has also moved to reduce staff while emphasising the growing role of AI across its operations.

The pattern is familiar:

cost reductions
AI investment
productivity narratives

But the interesting question is not why companies are investing in AI.

It is why they are moving so quickly to show productivity gains.

Because the evidence suggests something unusual may be happening.

Markets appear to be pricing the gains before they fully exist.

The AI Productivity Evidence So Far

There is credible research showing generative AI can increase productivity at the task level.

A large field experiment involving more than 5,000 customer-support agents found generative AI increased issues resolved per hour by about 14%, with the biggest improvements among less experienced workers.

Developers are seeing similar signals.

Controlled studies from GitHub report that programmers using AI coding assistants can complete certain tasks up to 55% faster.

These results are meaningful.

But they share one important feature.

They measure individual task productivity, not enterprise-wide outcomes.

Turning micro-level efficiency into measurable company-level gains is harder.

It requires workflow redesign, organisational change, and the redeployment of labour saved through automation.

Those transitions take time.

The Market Is Moving Faster Than the Evidence

Despite that lag, financial markets appear to be moving ahead.

Technology investors increasingly divide companies into two groups:

those that will benefit from AI productivity
those whose economics may be disrupted by it

That distinction is already influencing valuations.

Some companies are rewarded for demonstrating AI-led efficiency.

Others are punished if investors believe AI will weaken their competitive position.

The result is subtle but powerful.

Management teams now operate under pressure to demonstrate AI leverage.

Even if the long-term productivity gains have not fully materialised yet.

This Pattern Has Happened Before

Economists have seen this dynamic with other general-purpose technologies.

The electrification of factories in the early 20th century required decades of organisational redesign before productivity improvements appeared in national statistics.

Early factories simply replaced steam engines with electric motors.

The real gains only emerged when companies redesigned production lines around distributed power.

Computing followed a similar path.

For years economists talked about the productivity paradox, observing that computers were everywhere except in productivity statistics.

Only later, once organisations reorganised around digital workflows, did the gains become measurable.

The same pattern may be repeating with AI.

Initial adoption improves individual tasks.

But broader economic gains appear only after organisations redesign how work is structured.

The Market Is Already Acting

Investors may already be assuming that AI will deliver productivity gains across the technology sector.

But most companies are still experimenting with how to operationalise it.

That gap between expectation and execution is now starting to influence real decisions inside organisations.

Below I unpack:

• why markets appear to be pricing AI productivity ahead of reality

• the three different layers of the AI productivity story

• the strategic dilemma CEOs now face as expectations rise faster than results

If you lead a company deploying AI, understanding that gap matters more than the technology itself.

The Three Layers of the AI Productivity Story

Right now, three different timelines are colliding.

Read more