The NHS Just Handed Every Transformation Leader a Very Expensive Lesson

Posted on by

 

Four NHS trusts have now admitted their discharge delay figures were wrong. Not slightly wrong. The kind of wrong where numbers fall from the thousands to zero overnight, then climb straight back up within weeks, a pattern no functioning hospital produces naturally. Those figures sat underneath NHS England’s proudest claim about its £330 million Palantir contract: a 15 per cent fall in delayed discharges, held up as proof the Federated Data Platform was working.

A Financial Times investigation has now found irregularities in the discharge data of 42 per cent of all NHS trusts, across four years of records. The UK’s statistics regulator, the Office for Statistics Regulation, is investigating how the figures were used to justify the technology. A cross-party group of MPs has written to ministers urging the government to use the contract’s break clause, which the government can exercise from February 2027. And NHS England’s own chief executive, Sir Jim Mackey, told a select committee this week that he had personally challenged whether the benefits claims “have been objective and can be fully stood up if challenged.”

The person running the organisation just told Parliament, on the record, that he isn’t confident the headline number survives scrutiny. That’s a long way past a minor caveat.

It doesn’t stop at discharge delays either. A separate Freedom of Information request from the campaign group Foxglove found that close to a third of trusts using the platform’s scheduling tool carried out fewer procedures after adopting it than before, and that a single trust, Chelsea and Westminster, accounted for 84 per cent of the reported fall in outpatient waiting lists across the entire programme. One hospital’s good year, dressed up as a national result.

I’ve sat in enough steering committees to know exactly how this happens. And it isn’t really a story about Palantir.

 

The Data Was Never Built to Do This Job

Charles Tallack, formerly head of operational research and evaluation at NHS England, put it plainly: the evidence for the platform’s impact “looked increasingly flimsy.” His reasoning matters more than the headline. “The delayed discharge dataset may be suitable for day-to-day management purposes, but not for evaluation,” he said.

That’s the whole story in one sentence. NHS England’s own website admits the data undergoes only “minimal validation”, because the speed of collection doesn’t allow for more; it’s explicitly badged as “fit-for-purpose” for NHS management information.

It was built so ward managers could see who needs discharging today, not so a select committee could weigh whether a £330 million technology contract earned its keep. Those are two different jobs, needing two different levels of rigour. Somewhere along the way, one got quietly substituted for the other.

Every transformation leader has watched this substitution happen. Operational dashboards get repurposed as benefit trackers because they’re already there, already live, already familiar to the room. Nobody sits down and consciously decides to treat management data as evaluation-grade evidence. It just drifts that way, one board pack at a time, until a number designed to flag today’s bottleneck is being quoted as proof a nine-figure programme delivered its business case.

 

When the Numbers Look Too Clean, Get Suspicious

A drop from thousands of delayed discharges to zero, then straight back up, should never have made it into a report unchallenged. That’s not an improvement curve. That’s a data pipeline breaking.

I’d go further: any benefit metric that moves in a straight line, with no noise, no seasonality, no awkward months, should raise your suspicion before it raises your confidence. Real operational change is messy. It has plateaus, regressions, a bad winter, a strike, a system outage. A number that behaves too perfectly is usually telling you something broke upstream, not that something improved downstream. And a national result that traces back to one outperforming site, as the Foxglove data suggests happened here, is a local win being marketed as a systemic one.

 

Whoever Owns the Contract Shouldn’t Own the Evidence

The underlying dataset sits with NHS England, not Palantir. But that’s precisely the point worth stressing. The organisation whose reputation, and whose vendor relationship, depended on this figure looking good was also the organisation compiling it, with minimal quality checks, and no independent evaluation running alongside it until the regulator forced the question.

One NHS official told the FT that trusts “are being asked to put their name to statements about improvements before the tools are fully embedded and before the evaluations are done.” Read that twice. Governance failed here. Data quality is just where it happened to show up first, and I’ve seen it inside plenty of transformation programmes that had nothing to do with the NHS or with Palantir.

If the same team that needs the benefit case to land is also the team producing the evidence for it, the incentive to tell a good story will always beat the incentive to tell the true one.

 

Three Questions Worth Asking Before You Quote a Benefit Number Externally

Before any number from your programme reaches a board pack, a press release, or a select committee, it’s worth asking:

Was this dataset designed to answer the question I’m now asking of it, or was it designed for something else entirely and repurposed under pressure?

Who compiled this figure, and do they have a stake in it looking good?

Would this number survive an independent audit conducted by someone with no relationship to the programme?

If you can’t answer all three with confidence, what you’ve got is a hypothesis pretending to be a benefits case.

 

The Real Cost Isn’t the Contract

NHS England will likely survive this, whatever happens to the Palantir contract when the break clause opens in 2027. What’s harder to repair is trust in the next number this organisation, or any organisation, puts in front of Parliament, staff, or the public. Sir Jim Mackey said an objective review “would be helpful and necessary” but would take months. That’s months of every subsequent claim being read with one eyebrow raised.

Build your evaluation evidence with the same rigour you’d want turned on you, before someone else turns it on for you.

Handling Stakeholder Expectations in Digital Transformation: The Honesty Problem

Posted on by

 

Most failed digital transformations were not derailed by technology.

They were derailed by a gap between what was promised at the start and what was achievable in reality. That gap existed from day one, embedded in the business case, and neither the sponsors nor the delivery team chose to address it directly until the programme was already in trouble.

The stakeholder expectation problem in digital transformation is routinely framed as a communication challenge. Better updates. More frequent steering committee engagement. Clearer reporting. These are sensible practices. They are also, in most cases, insufficient, because the problem is not that stakeholders were not kept informed. It is that they were kept informed using numbers and timelines that were not honest about the uncertainty behind them.

That is an honesty problem, not a communication problem. And the fix for it happens at the beginning, not during delivery.

 

The Business Case That Everyone Signed Off On

Digital transformation business cases are almost universally optimistic. Not because the people who write them are dishonest, but because the incentive structure in most organisations rewards ambition and penalises conservatism. A realistic business case, one that acknowledges uncertainty ranges, models downside scenarios, and commits to fewer benefits with higher confidence, is harder to get approved than an ambitious one. So the ambitious one gets written.

The consequence is that the business case becomes a set of commitments rather than a set of hypotheses. By the time the programme moves into delivery, the numbers in the original document are treated as targets rather than as estimates, even when the assumptions underlying them have already been revised. The expectation gap was always there. It was just papered over.

BCG’s analysis of more than 850 companies found that only 35% reach their stated digital transformation goals. Gartner’s October 2024 survey of more than 3,100 CIOs found only 48% of digital initiatives meet or exceed their business outcome targets. The gap is not primarily a delivery capability problem. It is a framing problem.

 

Scope at Altitude

The second structural cause of expectation gaps is scope defined at too high a level of abstraction. Transformation programmes are typically scoped during a phase when the delivery architecture is not yet understood, which means scope boundaries are drawn based on intent rather than on a detailed model of what delivery will actually require.

Both sides, sponsor and delivery team, leave the scoping phase with genuine but different understandings of what is included. Neither party is misrepresenting anything. They simply have not gone deep enough to discover the ambiguity. That ambiguity is then carried into the contract, into the programme plan, and eventually into the steering committee deck, where it will surface as a scope dispute at the moment least convenient for everyone involved.

The fix is not to define scope more tightly in the abstract. It is to define scope at a level of specificity that forces the ambiguity into the open before commitments are made. That is harder and slower than moving quickly to contract. It is also significantly less expensive than managing the dispute six months into delivery.

 

What the Programme Board Actually Hears

Stakeholder management, in practice, often means giving senior sponsors the confidence to remain supportive rather than giving them the information they need to make good decisions. Status reporting in large programmes tends to converge toward reassurance. RAG ratings stay amber longer than conditions warrant, because the consequences of going red feel disproportionate in the moment. Risks that have materialised are carried as risks rather than re-classified as issues. Forecasts are revised gradually rather than reset to reflect the actual picture.

This is not cynical. It is human. Nobody wants to be the person who delivers bad news. The programme team has worked hard. The delays feel temporary. There is always a reasonable argument for holding the line a little longer.

The problem is that by the time the gap between expectation and reality is reported honestly, it is too large to close without a significant reset. The reset conversation is much harder than it needed to be, because the sponsor was not kept informed of how the gap was developing.

 

The Expectation Reset Conversation

When the gap surfaces, and it always surfaces, the response that preserves the programme is not to defend the original business case. It is to reframe the conversation around what is still achievable, with what degree of confidence, on what timeline. That requires the delivery team to be willing to put a revised view in front of the sponsor, acknowledge that the original framing was overoptimistic, and propose a credible path forward.

That conversation is significantly easier when the relationship between sponsor and delivery has been built on honest reporting from the start. It is significantly harder when the sponsor has been receiving optimistic status updates and now feels misled, not because anyone intended to mislead them, but because the communication was shaped by the desire to maintain confidence rather than the obligation to maintain accuracy.

 

Less Ambiguity, Earlier

The conditions that produce the stakeholder expectation problem are well understood: optimistic business cases, scope defined at altitude, and status reporting shaped by the incentive to maintain confidence. None of these are inevitable.

The organisations that manage stakeholder expectations well are not the ones with the best communication strategies. They are the ones with the discipline to be specific about uncertainty before programmes begin, to define scope at a level of detail that surfaces ambiguity early, and to build reporting practices that give senior sponsors the information they need to make real decisions rather than the reassurance they want to maintain support.

Less ambiguity earlier. Fewer numbers presented as certain when they are estimated. Fewer commitments made before the delivery architecture is understood.

That is the fix. It is harder to sell at the outset and significantly easier to live with throughout delivery.

The AI Infrastructure Race Has Already Been Decided, Just Not Where You’re Looking

Posted on by

 

Four companies have committed more capital to a single region’s AI infrastructure than most countries spend on national defence in a year.

AWS, Google, Microsoft, and Oracle have collectively committed more than 160 billion US dollars to building AI infrastructure across Asia-Pacific between January 2024 and May 2026, according to McKinsey’s analysis of the region’s data centre demand. That is not a forecast or an aspiration. It is capital already committed, over a 28-month window, by the four organisations best positioned in the world to judge where AI compute demand is actually heading.

Most enterprise conversations about AI strategy still treat the geography of AI capability as fixed, anchored in North America and Europe. That assumption stopped being accurate somewhere in the last two years, and the redrawing is happening in Asia-Pacific, largely unnoticed by the boardrooms it will eventually affect.

 

The “Follower” Narrative Was Already Wrong

The conventional framing of AI outside North America and Europe has always been one of catching up, adopting capability built elsewhere, closing a gap set by others. That framing was already inaccurate before this capital started moving, and the UAE is the sharpest evidence of it. Microsoft’s AI Economy Institute put the UAE’s working-age AI adoption at 70.1 per cent in its Q1 2026 diffusion report, the highest of any economy measured, against a global average of 17.8 per cent. Abu Dhabi is also home to Stargate UAE, a 5-gigawatt AI campus built with OpenAI, Oracle, and Nvidia that is now the largest AI infrastructure deployment outside the United States. Neither of those is a country catching up. That is a country leading.

The hyperscaler capital now moving into Asia-Pacific specifically, the $160 billion figure McKinsey tracks across AWS, Google, Microsoft, and Oracle, sits in a different regional bucket to the UAE in most analysts’ own classifications, McKinsey included, which places the Gulf within EMEA rather than APAC. But read together, the pattern is bigger than either region’s infrastructure story on its own. AI leadership has already decentralised away from North America and Europe in adoption terms. The capital now following it into Asia-Pacific is the same shift playing out in infrastructure terms, just in a different part of the map. The four largest cloud infrastructure providers on earth are not building in Asia-Pacific because the region is catching up. They are building there because the assumption that AI capability originates in the West and diffuses outward has already been disproven elsewhere, and they are positioning for where demand actually sits next.

That distinction matters for anyone making a ten-year technology strategy decision today. Compute infrastructure built now does not simply serve current workloads. It becomes the physical foundation that shapes what is commercially viable to build on top of it for the following decade. The parallel worth drawing is North American cloud infrastructure investment around 2015, which quietly determined which companies had a structural cost and latency advantage for the cloud-native decade that followed. Most of those advantages were locked in years before most executives recognised the pattern.

 

Building Faster Than Anyone Can Govern

What makes this moment genuinely worth attention is not just the scale of the capital commitment. It is the gap between that commitment and what has followed it.

The physical infrastructure, the data centres, the compute capacity, the power agreements, is being built at a pace the market has not seen before. The governance, integration, and organisational capability needed to actually use that infrastructure well has not kept pace at anything like the same speed. This is the same structural gap showing up across every AI signal this year: deployment and physical capacity moving faster than the organisational readiness required to extract value from either.

For enterprises operating in or adjacent to Asia-Pacific markets, this creates a specific and immediate strategic question, not a hypothetical one for next year’s planning cycle. The infrastructure being built now will define whose AI workloads run cheaply, quickly, and reliably in the region from 2027 onwards. Enterprises without a considered position on that infrastructure are not neutral bystanders. They are watching the operating environment for their future competitors being constructed, largely without their input.

 

What This Actually Requires From Leadership

Chasing every regional infrastructure headline is not the answer. Two things are.

The first is a straightforward board-level question that most technology strategy committees have never actually asked: does our AI roadmap account for where the compute capacity underneath it is being built, and by whom? Most enterprises with APAC exposure have not asked this, because compute geography has always been treated as an IT procurement detail rather than a strategic input.

The second is timing discipline. The infrastructure decisions with the longest shelf life, cloud provider selection, data residency architecture, regional partnership structures, are being made now, this year, by enterprises that recognise the window. Wait for the 2027 competitive gap to become visible and the decision will already have been made, by whichever provider has the compute capacity and the customer relationship in the region first.

The infrastructure race rarely announces itself as urgent while it is still open to influence. It only looks urgent in hindsight, once the capital is spent and the advantage is locked in. Right now, for Asia-Pacific, it is still open.

Your Enterprise Architecture Is Built on a Map That No Longer Exists

Posted on by

Most enterprise technology decisions are made on a map that no longer reflects the territory. Vendors occupy defined positions. ERP here, CRM there, ITSM in its lane, workflow tooling beneath. Procurement, architecture, and integration planning all proceed on the assumption that those positions are reasonably stable.

Two announcements made within days of each other in May 2026 did not just shift those positions. They rendered the map itself unreliable.

At Knowledge 2026, ServiceNow unveiled Autonomous CRM, covering sales, service, quoting, order fulfilment, invoice disputes, renewals, and the full customer lifecycle. A direct entry into Salesforce’s core market from a vendor whose identity has been workflow and ITSM. At Sapphire 2026, SAP declared itself a business AI company, launched its Autonomous Enterprise vision, in which AI agents execute end-to-end business processes, with humans directing strategy rather than managing individual steps, and acquired Reltio to make enterprise data AI-ready. An ERP vendor positioning as an AI orchestration layer across the entire enterprise.

Most commentary has treated these as two separate vendor stories. They are not.

 

Two Announcements. One Signal.

What ServiceNow and SAP announced is not primarily about features. It is about boundaries, and the dissolution of them.

For the past decade, enterprise technology portfolios have been built on a category model. You choose an ERP, a CRM, an ITSM platform, a workflow layer, and you integrate them. Vendors in each category compete within it. The architecture question is mostly about how the categories connect, not whether the categories themselves hold.

That model has broken. ServiceNow is not extending into adjacent territory around the edges. It is standing directly in Salesforce’s most defensible ground, covering sales pipeline, quoting, order management, and customer lifecycle, with AI as the differentiator. SAP is not adding AI features to its ERP. It is repositioning as the orchestration intelligence for the entire autonomous enterprise, with a master data acquisition to back it.

The category model that procurement teams, architecture boards, and technology roadmaps are built on is now operating on assumptions that neither vendor supports.

 

The Roadmap Problem

This matters less as a vendor story and more as a planning problem.

Enterprise technology decisions have long lag times. A CRM strategy agreed in 2023 reflects assumptions about what Salesforce competes with and how. An ERP consolidation approved in 2024 was scoped against SAP doing one thing and other vendors doing adjacent things. Integration architectures designed eighteen months ago were designed for a world where the platforms stayed in their lanes.

None of those assumptions survived May 2026. And the organisations currently finalising multi-year enterprise software contracts, negotiating renewal terms, or approving architecture blueprints need to know that before they sign.

The problem is not that ServiceNow and SAP have made bold moves. Vendors always make bold moves. The problem is that decisions downstream of those moves, about what to buy, what to build, what to integrate, and which vendor relationships to deepen, are still being made against the old map.

Two specific conversations are worth having before any major enterprise software decision closes in the next twelve months.

The first is the vendor dependency audit. When your workflow vendor is also your CRM, and your ERP vendor is also your AI orchestration layer, the concentration risk in your technology portfolio changes. So does the negotiating leverage. So does the cost and complexity of exit if you need it later. These are not hypothetical concerns. They are the direct consequence of vendors collapsing categories.

The second is the integration investment review. Integration work designed to connect cleanly bounded platforms does not simply carry across when those platforms expand into each other’s territory. Some of it becomes redundant. Some of it creates conflict. Some of it was justified by a separation of function that no longer exists. If your architecture team has not reviewed integration design against what ServiceNow and SAP announced in May 2026, that is a gap worth closing quickly.

 

What the Briefing Should Have Said

Technology updates for executive teams and programme boards tend to cover vendor announcements as market news. ServiceNow does this, SAP does that, here is what it means for the industry. That framing is too distant to be useful.

The briefing transformation leaders needed, and in most cases did not receive, is this: the vendor categories on which your enterprise architecture is built are collapsing. Two of the largest platform vendors are now competing directly across the boundaries that your current roadmap treats as fixed. The decisions you need to revisit are specific, and the window before contracts close is finite.

That is a different conversation from a product announcement. It requires someone in the organisation to have read the news, synthesised its implications, and brought it to the table as a strategy question rather than a technology update.

Most organisations are not structured for that kind of synthesis. Technology teams report on what vendors do. Strategy functions rarely track vendor positioning in this level of detail. The CIO’s office is often the only place where the two sets of knowledge meet, and it is frequently operating at capacity on current programmes rather than monitoring future architecture exposure.

The gap this creates is real, and it has a cost. Not immediately, but at contract renewal, at architecture review, at the moment an integration investment turns out to have been built against a boundary that no longer exists.

The ServiceNow and SAP announcements are not the story. The story is that the category model your technology decisions are built on changed, and most of the people who need to know have not been briefed. That is an organisational problem, not a technology one, and the organisations that address it before the next contract closes will be in a materially different position from the ones that address it after.

Pre-Mortem: Eight Companies, No Published Accountability Standard

Posted on by

The Pre-Mortem is a weekly series on this blog. Each piece applies five questions to a major technology commitment before the outcome is known.

In February 2026, the United States Department of War signed agreements with eight of the world’s leading artificial intelligence companies, OpenAI, Google, Microsoft, SpaceX, Oracle, Amazon Web Services, NVIDIA, and Reflection, to deploy their advanced AI models inside its classified networks. Impact Level 6 (IL6) covers data classified at the Secret level. Impact Level 7 (IL7) covers compartmented intelligence and the most sensitive operational systems, where the United States military runs its actual warfighting decision support. This is the first time that large language models have operated within IL7 environments. What has not been published is who carries accountability when one of them gets something wrong.

 

The Bet

The Department of War’s stated aim is to establish the United States military as an AI-first fighting force, achieving what its AI Acceleration Strategy calls decision superiority across all domains of warfare. The eight agreements are the mechanism. The AI systems will summarise surveillance feeds, synthesise intelligence data, and suggest tactical options to human operators. The Department of War’s five AI ethics principles, responsible, equitable, traceable, reliable, and governable, are on the record. The bet is that those principles are sufficient architecture for what happens inside a classified environment.

 

The Assumption

The whole bet turns on this: that “humans remain accountable for AI outcomes” as a stated principle is equivalent to a published accountability framework.

That distinction is where there is a gap. The Department of War’s Responsible AI Strategy and Implementation Pathway establishes process. It does not name the specific individual, command role, or governance layer accountable when an AI-assisted intelligence summary inside an IL7 environment shapes a decision that turns out to be wrong. Principle and framework are not the same thing, and in a classified environment that distinction cannot be tested publicly.

 

The Sequence

In July 2025, Anthropic’s Claude became the first frontier AI model approved for use on classified networks. The Pentagon subsequently sought to renegotiate those terms, demanding Anthropic permit its models to be used for all lawful purposes without limitation. Anthropic declined, citing concerns about mass domestic surveillance and autonomous weapons. On 27 February 2026, President Trump ordered all federal agencies to stop using Anthropic. The following day, OpenAI signed its classified deal with commitments that included prohibitions on domestic mass surveillance and human responsibility for the use of force, positions that aligned with the guardrails Anthropic had sought to retain. By May 2026, the remaining seven of the eight, Google, Microsoft, SpaceX, Oracle, Amazon Web Services, NVIDIA, and Reflection, had signed equivalent agreements.

The sequence reveals something structural. The accountability architecture for classified military AI was settled by commercial negotiation and political designation, not by a published governance framework.

 

The Pager

Legal scholars on autonomous weapons identify the same accountability fracture that applies in the decision-support context here. When an AI-assisted output causes harm in a classified environment, accountability distributes: software developers could not have anticipated all operational contexts, commanding officers disclaim responsibility for machine-generated outputs, vendors invoke contractual limitation of liability. The human-in-the-loop design means a person reviews AI suggestions before acting. It does not mean accountability for acting on a wrong AI output has been named anywhere in the command chain.

No published document names the specific individual role, command layer, or governance body accountable for a wrong AI-assisted output inside an IL7 environment. No congressional oversight mechanism covers classified operational AI use. No published error reporting standard exists. By the nature of classified operations, none can.

 

The Proof

Eight companies, the highest classification levels, large language models operating on top-secret data for the first time: the scale of the commitment is confirmed. The outcome data will not follow. Classified operational AI performance is not publicly reviewed, by design. This is the only deployment in this series where the proof question cannot be answered from the outside, not because the data is not collected, but because it cannot be published.

The accountability question is not whether humans are in the loop. They are, by stated commitment. The question is whether the framework for who carries it specifically, when they get something wrong, inside a system that cannot publish what it got wrong, exists in any enforceable form.

 

The Verdict

If the Department of War’s five principles are operationalised into a named, enforceable command accountability chain for AI-assisted decisions at every classification level, if the commercial guardrails in all eight agreements are independently verifiable by a body with appropriate clearance, and if a congressional oversight mechanism specific to classified AI operational failure is established, then this is what responsible military AI deployment at scale should look like.

Without all three, eight of the most powerful AI systems on earth are running inside the most classified networks in the world. The decisions they shape will not be publicly reviewed. The wrong ones will not be counted.

The accountability is a principle. The framework has not been built yet.

AI Gets the Blame. Governance Built the Problem

Posted on by

When an AI system causes serious harm, the story is easy to write. Biased algorithm. Dangerous model. Technology that cannot be trusted. The headline names the AI and moves on.

What the headline does not name is the governance meeting that never happened. The audit trail that was never built. The accountability structure that should have existed before the first automated decision was made, and did not.

That absence is the story.

Across healthcare, hiring, public services, and criminal justice, the pattern repeats with such consistency that it has stopped looking like bad luck and started looking like a structural failure. AI is deployed to solve a genuine operational problem. The deployment decision is made without the governance architecture that would constrain it, audit it, or catch its errors before they compound. Consequential harm compounds. And then AI gets the blame, while the process failure is buried somewhere in paragraph twelve.

These are not AI stories. They are governance stories. AI is the mechanism.

 

The Cases That Make It Visible

In January 2026, a class action was brought against Eightfold AI, a hiring platform used by major employers globally. The case was filed by Jenny Yang, former chair of the Equal Employment Opportunity Commission. It does not argue that the algorithm was biased, though that question remains open. It argues that the system operated in secret. Eightfold had scored over one billion workers on a scale of zero to five, and candidates ranked at the bottom were discarded before a human being ever saw their application.

Seventy per cent of companies using AI in hiring allow AI to reject candidates at the initial screening stage, with no human review at that point. One in five goes further, allowing AI to reject candidates at every stage of the process with zero human involvement at any point. That is not a technology decision. It is a governance decision. Someone, somewhere, made the deliberate choice to remove the human from the loop. Nobody built an accountability structure around what happened next.

The pattern is not new. In 2021, the Dutch government resigned after an AI system falsely accused twenty thousand families of child welfare fraud. Courts ordered repayments of tens of thousands of euros per family. In Australia, the Robodebt programme issued four hundred thousand wrongful fraud accusations before it was ruled unlawful and the government repaid over one billion dollars. In Michigan, a 2024 settlement reimbursed three thousand plaintiffs for what a benefits fraud algorithm had wrongly taken from them.

In each case, the AI system did what it was designed to do. What nobody designed was the mechanism that would question whether it was doing the right thing.

 

The Structural Argument

The research makes the pattern numerical.

An analysis of a hundred and forty enterprise AI implementations found that only twenty-three per cent of failures were caused by model performance, data quality, or technical integration. The remaining seventy-seven per cent came down to strategy, governance, and change management.

Three-quarters of AI failures have nothing to do with the technology.

Only one in five organisations has a mature governance model for autonomous AI agents. This is the population deploying AI at scale, across consequential decisions in hiring, healthcare, benefits, credit, and criminal justice, mostly without the mechanisms needed to know whether the AI is producing correct outcomes, or what to do when it does not.

This is not a portrait of reckless technology. It is a portrait of reckless deployment. The AI worked. The organisation around it did not.

 

Governance as Delivery Discipline

Most organisations treating AI governance as a compliance function are building the next wave of failures right now. Compliance asks whether the system meets a threshold at the point of deployment. Delivery discipline asks whether the system is behaving as intended across every subsequent decision it makes, and whether there is anyone accountable when it does not.

These are not the same question. That gap is where accountability ends and harm begins.

Effective AI governance is not about slowing deployment. It is about building the accountability architecture alongside the deployment. An agreed definition of what the system is supposed to achieve. A method for measuring whether it is achieving it. A human being, with authority and accountability, responsible for reviewing outcomes at meaningful intervals. An appeals mechanism when the system gets it wrong, because it will. Documentation that allows an audit when something goes wrong, rather than after the damage is done.

 

The Question Worth Asking Before the Next Deployment

The Eightfold case will not be the last of its kind. The healthcare billing figures will grow before they shrink. More governments will face the political and financial cost of systems that automated consequential decisions without the mechanisms to catch errors before they multiply.

The organisations that avoid this are not the ones that move slower on AI. They are the ones that treat governance as part of what delivery means, not as a separate conversation to have later, when the headlines arrive.

By then, the structure has already failed. The question worth asking now is whether yours is being built.

Your Enterprise AI Programme Is Structured Backwards.

Posted on by

There is a research paper that has been making rounds in enterprise AI circles and deserves more attention than the single line most people have taken from it.

The Stanford Digital Economy Lab published its Enterprise AI Playbook in April 2026. It is drawn from 51 live, production-grade deployments across 41 organisations, seven countries, and more than one million employees. And what it found cuts against almost every assumption the technology industry has built its messaging around.

In every deployment that succeeded, and in every deployment that failed, the determining factor was not the model. It was not the vendor, the feature set, the benchmark performance, or the integration timeline. In every case, the variable that decided outcome was the organisation: executive sponsorship, governance architecture, and the quality of workforce change management. Seventy-seven per cent of the implementation challenges in the study traced to non-technical factors: change management, data quality, and process redesign.

 

One Finding in 51 Deployments

The research does not say technology does not matter. It says it matters significantly less than most programmes treat it as mattering, and that the organisations which led with governance and change management consistently outperformed those that led with model selection.

One adjacent data point makes this more concrete. Sixty-one per cent of the successful deployments in the study followed at least one prior failed attempt. Those organisations had not found a better model on the second attempt. They had changed what they were actually doing. And in almost every case, that meant addressing the organisational variables they had underestimated the first time: governance structure, change management approach, and clarity of executive ownership.

The technology was not the lesson. The organisation was.

 

How Most Programmes Are Actually Structured

I have sat in enough enterprise AI programme kick-offs to recognise the pattern before the second slide.

The programme begins with a vendor selection process. A proof of concept is scoped, model performance is evaluated, pricing tiers are compared, latency is benchmarked. The technology conversation consumes the first two to six months of the programme. By the time it concludes, the organisation has committed significant capital and credibility to a specific platform before the questions the Stanford data confirms are the real determinants of success have been seriously engaged.

Those questions are not complicated. Who is the executive sponsor, and what does their sponsorship mean in terms of decision-making authority and resource commitment, not just endorsement? What is the governance architecture for the AI programme, not for the AI system, but for the programme? How does the organisation plan to manage the workforce transition that a serious deployment requires, and what does it know about the change-readiness of the teams it is deploying into?

These are not questions that get answered in a vendor evaluation. They are not questions that appear on most programme charters. They are the questions that decide whether the programme succeeds.

 

What Leading With Governance Actually Means

Leading with governance does not mean delaying deployment while a committee produces documentation nobody will read. It means defining, before the technology is in the ground, who owns the programme’s outcomes, how the workforce transition will be handled, and what the decision-making structure looks like when the deployment hits the friction that every serious AI implementation hits at scale.

Because the friction will come. It always does. And how an organisation responds to it reveals which frame it used at the start.

Programmes built on a technology frame diagnose the friction as a technology problem. The model is adjusted. The integration is patched. The interface is redesigned. The organisational dynamics actually driving the resistance go unexamined, because the programme was never looking at them. Programmes built on a change management frame diagnose the same friction differently. The conversation shifts to whether the right people were involved in design, whether transition support was adequate, whether the governance gave teams the clarity they needed to work confidently with the new system. Those questions lead somewhere. The technology-first version usually leads to another vendor call.

 

The Argument That Is Now Evidence

This is not a new insight for experienced transformation leaders. I have been making a version of this argument for years, and so has almost everyone else who has led a serious enterprise change programme. The frustration, and the genuine value of Stanford’s work, is that it can now be asserted with data.

For transformation leaders making the case for governance investment in leadership conversations where the pressure is almost always to accelerate on technology, the Stanford Playbook is the data point that turns an argument from opinion into evidence. It does not require arguing against the technology. It requires arguing about sequence and proportion, and it gives you the empirical foundation to do it.

The organisations still leading with model selection are systematically delaying the decisions that actually determine whether a deployment succeeds.

Fifty-one real deployments confirm it. That should be enough.

Workplaces Don’t Just Pay You. They Shape You

Posted on by

There is a version of career advice that treats every job as essentially equivalent, a set of roles, responsibilities, and remuneration packages to be evaluated primarily on their individual merits. You assess what the role requires, what it pays, and what it advances you toward. This is a reasonable framework. It is also incomplete.

The workplace you spend three or five years in does not just pay you. It shapes you. It forms the decision-making patterns, the risk tolerance, the communication instincts, and the professional identity that you carry into every subsequent role. Often in ways you cannot fully see until you have left.

 

How Environments Form Capability

The formation happens gradually and mostly below the surface. In an organisation with strong analytical discipline, you learn, often without realising it, to interrogate assumptions before committing to a course of action. In an organisation where pace is valued above rigour, you learn to move quickly and make decisions with incomplete information. In an organisation where challenge is welcomed, you develop the confidence to push back. In one where it is not, you develop the habit of working around problems rather than confronting them.

None of these formations is necessarily good or bad in isolation. All of them travel with you.

The professional who has spent five years in a high-accountability, high-clarity environment arrives in their next role with a set of instincts (about what questions to ask, what risks to flag, what evidence to require before deciding) that their counterpart from a lower-accountability environment simply does not have. The gap between them is not qualifications or experience in any formal sense. It is formed professional judgement.

That formation is not something a training course produces. It is the cumulative output of the environment itself: the decisions you are asked to make, the standard your work is held to, the conversations you are included in or excluded from, the quality of the thinking you are surrounded by.

 

What the Research Confirms

LinkedIn’s 2025 Workplace Learning Report found that 88 per cent of organisations say employee retention is a concern, and that providing learning opportunities is their top retention strategy. That finding is usually read as an HR insight. It is also a signal about what people themselves understand: the environment you work in is the primary determinant of your professional growth, and its absence is what drives capable people out of the door.

The same research found that only 36 per cent of organisations qualify as genuine career development champions, with robust and actively used development programmes in place. Thirty-three per cent have no meaningful initiatives at all, or are just beginning to build them.

The gap between those two groups is not a gap in intention. Most organisations understand that development matters. It is a gap in practice. Capable people read the environment accurately. When it stops investing in them, they leave for ones that do. The organisations that cannot close that gap are not simply struggling with retention. They are handing their best people, already formed by the investment made in them, to the competitors who will benefit from what that formation produced.

 

The Compound Effect

Development is not linear. The professional who is given increasing challenge, exposed to better quality thinking, and held to a rising standard does not improve at the rate of their individual training investments. They compound. Each year of a strong environment builds on the previous one in ways that create something qualitatively different from the sum of the parts.

The professional who spends the same years in a low-challenge environment is not standing still. They are becoming expert at a narrower set of conditions. They are developing the instincts that environment rewards and allowing the ones it does not reward to atrophy. They will be competent within those conditions for as long as they apply, and brittle when they change.

This is not a moral observation about which organisations are virtuous. It is a structural one about how professional capability is formed. The environment is the curriculum.

 

The Decision You Are Actually Making

When you accept a role, you are making two decisions. The explicit decision is about the role: the responsibilities, the scope, the compensation. The implicit decision is about the environment: who you will learn from, what standard you will be held to, what kinds of problems you will be asked to think about, and who will be in the room when consequential decisions are made.

That first decision shapes the next two or three years. The second shapes the decade that follows.

Most career conversations focus almost entirely on the explicit decision. Salary, title, scope, promotion trajectory. These are real considerations. They are also the shorter-term ones.

The question worth asking, and the one asked too rarely, is what this environment will make of you. Not what you will do in it. What it will do in you.

An organisation that invests seriously in the development of its people, one that exposes them to hard problems, holds them to high standards, and creates the conditions for genuine challenge, is offering something that does not appear in the compensation package. It is offering the formation of a professional who will be more capable, more resilient, and more effective in every role that follows.

 

What This Means for Leaders

For leaders, the obligation this creates is clear. You are not simply extracting the capability your people currently have. You are shaping the capability they will have. The quality of that formation is your responsibility.

The organisations that take this seriously produce people others want to hire. That is a leading indicator, not a lagging one. If the market consistently values what comes out of your environment, you are building something. If it does not, the environment is telling you something.

Choose the environments you build as carefully as you choose the people you put in them.

Your AI Risk Register Does Not Reflect Your Actual Risk

Posted on by

 

On 22 June 2026, the intelligence agencies of the United States, United Kingdom, Australia, Canada, and New Zealand spoke in a single voice about enterprise AI risk, and what they said demands attention.

The Five Eyes cybersecurity agencies issued a joint statement warning that frontier AI models are improving at a pace that will allow them to bypass prevailing enterprise cybersecurity defences within months. Not within years. Not in the next planning cycle. Within months. The statement’s own language: “The timeline is not years, it is months.”

 

This Is Not an Abstract Warning

Joint statements from the Five Eyes agencies carry a different category of authority than vendor advisories or consultancy threat reports. These are national intelligence services with access to classified threat intelligence, speaking to government and enterprise leaders simultaneously. When they frame a risk as both imminent and enterprise-specific, take it at face value.

What sets this advisory apart from every AI security conversation most enterprises have been having is one thing: specificity. The Five Eyes statement does not describe abstract AI risks. It specifically names the enterprise AI tools deployed at scale in the last 18 months: copilots, AI assistants, browser-connected agents, and systems with access to operational and customer data. The primary attack mechanism, developed across Five Eyes guidance published earlier this year, is prompt injection: an adversary embeds hidden instructions in content the AI system processes, causing it to act outside its intended scope.

That specificity matters. It means the tools that most large enterprises have already deployed are the attack surface being described.

 

The Threat Moved Faster Than Your Review

Most organisations that have rolled out AI copilots, enterprise agents, or browser-integrated assistants have conducted security reviews of those deployments. The Five Eyes advisory is not questioning whether those reviews happened. It is saying that the threat has moved faster than the defences, and that a review conducted six months ago may no longer accurately reflect the risk profile today. The gap is not in intent. It is in elapsed time against a threat that has not stood still.

The advisory is explicit that this is not solely a security-team problem. The statement directs its recommendations at leadership, framing AI-driven cyber risk as a governance and board-level accountability question. The statement’s own title: “The AI shift in cyber risk: why leaders must act now.” That framing has direct implications for how risk registers are built and how AI deployment decisions are reported to boards.

 

Three Things Worth Doing Before Your Next Board Meeting

The advisory points to three things transformation leaders should act on before their next board meeting.

The first is a current security review. Every AI deployment connected to operational data, whether customer records, financial systems, or internal communications, needs a review that specifically addresses prompt injection risk. Not the review conducted at go-live. A current one, calibrated to the threat capability the Five Eyes describe as arriving within months.

The second is an updated risk register. Most enterprise risk frameworks assessed AI security risk at the point of initial deployment. The Five Eyes advisory says the threat environment has changed materially in the months since, and the assessment needs to reflect current threat capability rather than historical assumptions. An outdated risk assessment is not a minor administrative gap at this point. It is a governance exposure.

The third is using the advisory to reframe the conversation at board level. Six cybersecurity agencies from five countries issued this statement with an explicit focus on business leadership. That gives transformation leaders the instrument they need to move boards that have been treating AI security as an implementation detail. The Five Eyes advisory makes it a governance question. Use it as one.

The AI deployment decisions taken in the last 18 months created an attack surface. Most enterprise risk registers have not yet priced what that surface is worth to an adversary with AI-powered attack tools that are months from bypassing prevailing defences. That gap needs to close, and it closes with a current assessment, not one accurate at the time of go-live.

The EU AI Deadline Your Compliance Team Probably Missed

Posted on by

The EU AI Act enforcement date most organisations have been tracking is not 2 August 2026. They have been watching the high-risk provisions, the conformity assessments, the prohibited applications. Those timelines stretch into 2027 and beyond, and enterprise compliance teams have planned accordingly.

Article 50 has a different clock. It takes effect in 31 days, it applies to a far wider population of organisations than most realise, and for most of its obligations there is no grace period.

 

Not the Regulation You Were Watching

For the past two years, enterprise AI governance conversations have centred on the Act’s high-risk classifications. Which systems require conformity assessments? Which use cases are prohibited outright? The questions were legitimate, and the extended timelines attached to those provisions created a reasonable sense of runway.

That runway does not apply to Article 50.

Article 50 covers transparency obligations, and it lands on 2 August 2026. It requires any organisation deploying customer-facing AI systems to disclose to users that they are interacting with an AI. It requires providers of generative content tools to implement machine-readable marking on AI-generated outputs. Operators running emotion recognition or biometric categorisation systems must notify the individuals affected. And for any new system entering the EU market on or after 2 August, compliance is required from day one.

One aspect of the regulation that most compliance programmes have not fully processed: Article 50 is not jurisdictional. Article 50 follows the user, not the provider. That is how the Act defines its own scope. A company headquartered in Dubai, Singapore, or New York that deploys AI-generated content visible to EU users is in scope. Where the output lands determines the obligation. The practical consequence is that Article 50 applies to any organisation with a customer base that includes EU residents, regardless of where that organisation is incorporated or where its AI systems are built and operated.

The organisations that will be caught short are not the ones building prohibited systems. They are the ones that assumed the regulation was still in the planning stage, or that it would only apply to organisations based in Europe.

 

The GDPR Comparison That Matters

GDPR was announced in 2016 and took effect in 2018. Two years of awareness campaigns, legal seminars, board-level briefings, and vendor remediation work. The compliance industry built an entire ecosystem around it. Privacy officers were hired. Data mapping exercises ran for months. By the time enforcement began, organisations at least understood what was expected of them, even if some were still catching up.

GDPR also reached beyond EU borders from the start. Any organisation processing the personal data of EU residents was in scope, regardless of where it was based. Article 50 operates on the same principle: it reaches wherever EU residents are on the receiving end of AI-generated content or AI-driven interactions.

Article 50 does not have that context. Most enterprise compliance functions have been tracking the Act’s overall timeline without separating out which provisions take effect when. The transparency obligations were not deferred. They were always scheduled for August 2026. But because the high-risk provisions dominated the conversation, the transparency rules arrived quietly, and they arrive soon.

Thirty-one days is not a planning horizon. It is an implementation sprint, or it is already a compliance gap.

 

What Article 50 Actually Requires

The obligations are more specific than the general framing of “AI transparency” suggests, and that specificity matters for scoping the work.

The most broadly applicable obligation is disclosure. If a user is interacting with a chatbot, a virtual assistant, or any automated system capable of conversation or personalised response generation, they must be told. The requirement is not a buried terms-and-conditions clause. It is a functional disclosure at the point of interaction. This applies from 2 August, to all systems, with no transitional provisions.

Generative content carries a second obligation. Organisations using generative AI to produce content distributed in EU-market contexts must ensure outputs carry machine-readable markers indicating AI generation. This applies to text, images, audio, and video. The AI Omnibus agreement provisionally agreed in May 2026 and expected to be formally adopted before 2 August extends this specific requirement to 2 December 2026 for systems already on the market before 2 August. For any new system entering the market from that date, the obligation is immediate. The extension is not a signal to deprioritise: December 2026 is not far away, and the technical implementation is not trivial.

Emotion recognition and biometric categorisation carry a third obligation, active from 2 August with no transitional period. Individuals must be informed when these systems are operating on them.

None of these obligations are complex in isolation. The difficulty is that most organisations have not mapped which of their current systems fall within scope, and that mapping exercise takes longer than 31 days when it is starting from scratch.

 

What to Do in the Next 31 Days

Non-compliance carries fines of up to €15 million or 3% of global annual turnover, whichever is higher. This is not a planning conversation. It is a board conversation.

Article 50 requires operational change: disclosure mechanisms built into interfaces, technical markers implemented in content pipelines, notification processes embedded in operational workflows. A policy document does not close this gap.

The practical starting point is a scoping exercise, and it needs to happen this week, not at the end of July. Three questions define the scope: Which customer-facing systems use AI in any form of interaction or response generation? Which content production workflows use generative AI to produce material distributed in EU-market contexts? Are any systems using emotion recognition or biometric categorisation?

If the answer to any of those questions is yes and the disclosure or notification mechanism is not already live, that is an Article 50 compliance gap.

Once the scope is clear, triage by exposure. Not every system carries the same risk. Externally facing consumer products in regulated sectors carry a different risk profile than internal productivity tools. Sequence the remediation by audience, jurisdiction, and volume of interaction.

Confirming the mechanisms actually work is where most programmes get caught. A disclosure notice that technically exists but is not surfaced at the point of interaction does not satisfy the requirement. The same applies to machine-readable markers that are added to some content outputs but not systematically applied across all generative workflows. Implementation is not the same as compliance.

 

31 Days Is Not a Problem. 32 Days Is.

There is still time to close this gap for organisations that act now. August 2026 is not GDPR day one, when regulators were finding their feet. It is an enforcement event in a regulatory framework that has had two years of published timelines. Regulators will not be looking the other way.

The organisations that treated the high-risk provisions as the whole story now have 31 days to correct that assumption. Wherever they are based.