Category Archives: AI

Pre-Mortem: A Billion Workers Scored in Secret. Is It Legal?

Posted on by

On 20 January 2026, two job applicants filed a class action against Eightfold AI Inc. in a California state court. The complaint alleged that the company had scraped personal data on over one billion workers, scored every candidate on a zero-to-five scale, and discarded low-ranked applicants before any human saw their application. The legal basis is the Fair Credit Reporting Act (FCRA). The plaintiffs’ central claim is not that the algorithm was biased. It is that the algorithm existed in secret.

This is the fourteenth piece in the Pre-Mortem series. Five questions, applied to the public record, before the outcome is known.

 

The Bet

Eightfold AI and the companies deploying its platform are betting that an AI system which aggregates third-party data, including social media profiles, location data, and online tracking cookies, to score individuals for employment purposes does not meet the legal definition of a Consumer Reporting Agency under the Fair Credit Reporting Act. The complaint names Microsoft, Morgan Stanley, Starbucks, BNY, PayPal, Chevron, and Bayer as companies using Eightfold in their hiring process. Co-Founder and CEO Ashutosh Garg responded with a public statement on responsible AI, noting that the platform undergoes third-party bias audits and that data comes from candidates or employers, not third-party scraping. The bet is not about whether the algorithm is accurate. It is about jurisdiction: whether the FCRA, written before algorithmic hiring existed at this scale, reaches far enough to cover what Eightfold built.

 

The Assumption

The single belief the whole framework rests on: that an AI platform scoring candidates for employers is categorically different from a consumer reporting agency, because the platform does not produce a consumer report in the form the FCRA contemplates. Eightfold filed a 35-page motion to dismiss arguing precisely that. The hearing was held on 4 August 2026 before U.S. District Judge Yvonne Gonzalez Rogers in Oakland. No ruling has been published. If the assumption is wrong, the compliance obligations the FCRA places on consumer reporting agencies, including disclosure, consent, and accuracy mechanisms, apply to every AI hiring platform operating on third-party data at comparable scale.

 

The Sequence

20 January 2026. Class action filed by former EEOC Chair Jenny R. Yang and the nonprofit Towards Justice. The complaint: Eightfold AI functioned as an unregistered consumer reporting agency across a dataset of over one billion workers.

18 June 2026. Plaintiffs’ opposition to Eightfold’s motion to dismiss filed.

22 June 2026. In the parallel Mobley v. Workday case, a federal judge denied Workday’s motion to dismiss claims of race, age, and disability discrimination through AI hiring tools.

9 July 2026. Eightfold reply brief filed.

4 August 2026. Motion to dismiss argued in Oakland before Judge Yvonne Gonzalez Rogers. No ruling published as of 16 August 2026.

13 August 2026. Eightfold AI named “Agentic AI HR Solution of the Year” at the HR Tech Breakthrough Awards.

 

The Pager

Kistler et al. v. Eightfold AI Inc., No. 3:26-cv-01768 names Eightfold AI as defendant. No talent acquisition leader or CHRO at Microsoft, Morgan Stanley, Starbucks, or any other company deploying the platform has been named as a defendant, and no deploying company has publicly committed to disclosing the tool’s existence to applicants. The pager sits with the vendor. The question of who carries it at the companies deploying the platform remains unanswered.

Garg’s public statement on responsible AI is a creditable position. It does not address what obligations the companies using Eightfold carry, or what those companies owe to the candidates who may have been scored and discarded before a human saw their application.

 

The Proof

The motion to dismiss ruling is the first proof point. A denial advances the FCRA question to discovery and the merits. It would be the first federal answer on whether AI candidate scoring constitutes consumer reporting. A grant sends the question back to the FTC and Congress, where progress has not matched the scale of deployment. The outcome measure worth watching is not which side wins the motion. It is whether any major Eightfold client commits to applicant disclosure before the court decides whether disclosure is legally required.

 

Verdict

If Judge Gonzalez Rogers denies the motion to dismiss, the case advances and the FCRA question gets its first federal answer in the context of AI hiring tools. That ruling will matter to every organisation using algorithmic screening, not only Eightfold’s clients. A denial does not mean Eightfold loses; it means the question gets answered in a setting with evidence, argument, and binding precedent. If the motion is granted, the accountability gap returns to regulatory and legislative channels, where the pace has not matched the scale of the deployment. What would change this assessment is action of a different kind: a major employer publicly committing to applicant disclosure before the court makes the decision for them.

Prompt Injection Is a Governance Failure Wearing a Technical Costume.

Posted on by

Every prompt injection headline reads like a technical failure. A model got tricked. A filter didn’t catch it. The vendor needs to patch something.

That framing is comfortable, and it is wrong. The technical trick is real. The governance failure sitting underneath it is the actual story, and it is the one almost nobody in the room wants to own.

 

Why the Trick Works in the First Place

The mechanism is architectural, not a bug in the usual sense. Large language models treat the system prompt, the user’s request, and any text retrieved from an external source as a single stream of tokens. There is no reliable internal boundary between an instruction and a piece of data. A hostile sentence buried in a document, a web page or a support ticket can carry the same authority as a command typed directly by a trusted user, because the model was never built to tell the difference.

OWASP’s 2026 State of Agentic AI Security and Governance report found prompt injection now maps to six of its ten top categories for agentic applications, up from a mostly theoretical concern in the 2025 edition to one backed by documented breaches and tracked vulnerabilities. Coding agents dominate the attack data, and only 37% of organisations report having a policy in place to even detect unauthorised AI deployments running inside their own environment.

 

The Failure Is a Control Boundary, Not a Model Flaw

This is where the governance framing actually matters. Prompt injection succeeds because enterprise workflows assume the model can reliably tell trusted instruction apart from hostile text, an assumption that fails the moment one interface carries user intent, retrieved content and tool-facing control signals in the same session. Most organisations respond by treating guardrails as a static filter list rather than a real system of content separation, monitoring and authorisation. A filter can catch a known bad phrase. It cannot answer the actual governance question, which is who controls what the system is allowed to do once it has been steered.

Security researcher Simon Willison’s “lethal trifecta” names the actual risk condition plainly: an AI agent with access to private data, exposure to untrusted content, and the ability to communicate externally, all three at once, is where exfiltration happens. Meta’s own internal guidance treats those three properties as a budget rather than a checklist. Combine all three and the agent needs a human in the loop before it acts, not after.

 

Why This Keeps Getting Treated as IT’s Problem Alone

Handing this to the security team as a patching exercise misses what the data is actually showing. A model update might close one exploit path. It will not answer who approved an agent’s access to a customer database, why that same agent can also send emails externally, or what happens the day it does both at once because nobody ever wrote down that it should not be allowed to. Those are ownership questions, not model questions, and ownership questions do not get solved by a vendor release note.

 

What Governance-First Actually Requires

Start by classifying every channel an agent reads from according to trust level, and keep untrusted content out of instruction scope entirely rather than hoping the model sorts it out at runtime. Quarantine tool access behind explicit policy gates, so an agent combining private data access, untrusted content and external communication needs sign-off before it can act, not a retrospective audit after it already has. Treat a pattern of near-miss prompts as an abuse signal worth escalating, not a string of isolated one-off incidents each closed out individually.

All of it is the same governance discipline organisations already apply to identity and access management, pointed at a new kind of actor that happens to run on language instead of credentials, not a new technology purchase.

 

Who Approved This, and Did They Know What They Were Approving

Before the next prompt injection incident gets logged as a technical exploit, ask who actually approved the access the exploit relied on.

If nobody can answer that cleanly, the model was never the vulnerability. The governance around it was.

Pre-Mortem: The Accountability Question the Mills Review Left Open

Posted on by

On 6 July 2026, the Financial Conduct Authority published the Mills Review, its examination of how AI will reshape retail financial services in the UK. The review covers seven recommendations across the regulatory perimeter, oversight architecture, and the transition to autonomous decision-making. It names the accountability gap at the centre of autonomous AI trading. It does not close it.

This is the thirteenth piece in the Pre-Mortem series. Five questions, applied to the public record, before the outcome is known.

 

The Bet

UK firms deploying autonomous trading AI are betting that the Senior Managers and Certification Regime (SMCR), the framework that holds named executives personally accountable for conduct failures in their area of responsibility, covers their position through general senior manager oversight. The FCA has been clear that delegating a decision to an algorithm does not transfer senior manager liability to the algorithm. The bet is that this principle, correctly stated and on the public record, can be demonstrated in practice before an enforcement case defines what demonstrating it actually requires.

 

The Assumption

Seven recommendations. One question still without an answer:

When an autonomous trading system executes a decision at machine speed, without pausing for human approval of the individual trade, which specific senior manager function is accountable if that decision causes a customer loss, and what does demonstrating adequate oversight of a system like that actually require?

The Mills Review acknowledged the problem directly. Without guidance, the review found, the combination of greater opacity in AI-mediated decisions and factors such as model drift makes it harder for the regulator to identify a de facto responsible individual, or for senior managers to evidence meaningful human control. Stakeholder feedback throughout the review called for clearer guidance on what constitutes the “reasonable steps” expected of senior managers. The review recommends the FCA develop it. The FCA has not yet published it. Every firm currently deploying autonomous trading AI is operating on the assumption that its existing accountability structure covers the gap. That assumption has not been tested in an enforcement case.

 

The Sequence

December 2019. SMCR extended to all FCA solo-regulated firms, completing its rollout across financial services.

27 January 2026. The FCA launched the Mills Review, acknowledging that AI in retail financial services had developed faster than the regulatory frameworks designed to govern it.

24 February 2026. Call for input closed.

6 July 2026. The review published seven recommendations. The FCA committed to adapting its regulatory frameworks as the transition to autonomous models continues. No guidance named a specific senior manager function as accountable for autonomous trading decisions. No guidance defined what “reasonable steps” requires for a system executing at machine speed without human review of individual decisions.

The capability reached the market before SMCR was tested against it. The review arrived after the capability. The guidance has not arrived yet.

 

The Pager

The FCA has confirmed there will be no dedicated Senior Manager Function for AI, and that accountability falls on existing functions. That is a clear policy position and it deserves credit for being stated plainly. The Treasury Select Committee has urged the FCA to publish guidance specifying the level of assurance expected of senior managers for AI-related harm. The Mills Review carried that request forward into its recommendations. The harder question is the one seven recommendations did not answer: when an autonomous trading system causes a customer loss, which specific function holder carries the call?

 

The Proof

There are no enforcement cases. The first case will establish what “reasonable steps” means in an AI trading context. The Mills Review is a process measure: it produced recommendations. The outcome measure worth watching is whether the FCA’s follow-on guidance names a specific function and defines the oversight standard in operational terms rather than principles alone. A principle restated is not a gap closed.

 

Verdict

If the FCA’s follow-on guidance names the senior manager function accountable for autonomous trading AI and defines what “reasonable steps” requires at the operational level, UK financial services will have resolved an accountability gap that every other major jurisdiction is still navigating. The review’s existence, the named individual who led it, and the seven published recommendations are genuine evidence that the FCA identified the problem and moved on it. Without operational guidance, the gap stays open. The first enforcement case will write the rule in the least comfortable setting available. That is a considerably worse way to write it.

The $5.5 Trillion Bill for Doing Nothing About AI Skills

Posted on by

IDC put a number on the cost of enterprises not knowing what to do about AI skills: $5.5 trillion. Not by 2030. By the end of 2026.

That is not a distant workforce-planning problem. It is the price tag on decisions organisations are making, or more often deferring, this quarter.

The number comes from IDC’s survey of enterprise IT leaders across the US and Canada. It measures a specific kind of pain: product delays, lost competitiveness, and business walking out the door because the people needed to build and run AI capability were not there when the work needed them.

An IDC Spotlight Paper distributed via workforce-skills platform Workera applies the same figure directly to the AI skills conversation, and the application largely holds up. Most of what IDC’s original survey describes as a broader tech talent shortage is, in practice, an AI capability shortage wearing a wider label.

Three other figures make the same point from different angles. The World Economic Forum’s Future of Jobs Report 2025, surveying over a thousand employers and 14 million workers across 55 economies, found that 59% of the global workforce will need reskilling or upskilling by 2030, and that 11% of that group are unlikely to receive it. Indeed’s hiring data shows the share of job postings with “AI” in the title has more than tripled since 2022, from 2.6% to 8.3%. And PwC’s 2026 Global AI Jobs Barometer, built on more than a billion job advertisements, found that AI-skilled workers now command a 62% wage premium over comparable peers, up from 57% the year before and roughly 25% two report cycles before that.

These numbers describe a market that has already repriced itself, not a future state, while most enterprise workforce plans are still budgeted as though it hasn’t.

 

The Premium Is the Market Telling You Something

Wages move slowly almost everywhere except where genuine scarcity exists. A skill premium that has climbed from roughly 25% to 62% across three consecutive PwC survey cycles is not a normal labour-market signal. When a specific skill commands 62% more pay than the equivalent role without it, and that gap is still widening year over year, that is the market pricing in a shortage faster than most HR functions can respond to it, not a talent management curiosity.

Second Talent’s 2026 research puts a shape on that shortage: roughly 1.6 million open AI-related roles globally against around 518,000 candidates qualified to fill them, a demand-to-supply ratio a little over three to one. Every organisation competing for AI capability right now is competing inside that gap, and every quarter spent treating reskilling as a training-budget line item rather than a capital allocation decision is a quarter spent losing that competition to whoever moved first.

 

Why This Belongs in the Risk Register, Not the Learning and Development Plan

Most organisations still route AI reskilling through the same governance as any other training initiative: an L&D budget line, a completion metric, a once-a-year review. That treatment made sense when the skills in question were incremental. It does not hold up against a $5.5 trillion cost estimate and a wage market moving by double digits year over year.

A capital risk gets tracked differently to a training initiative. It gets a named owner, a quantified exposure, and a review cadence tied to the business calendar rather than the HR calendar. Few organisations apply that discipline to AI skills, because the function historically responsible for skills, HR, was never built to run risk registers, and the function that runs risk registers, finance and the PMO, was never asked to own workforce capability.

That gap in ownership is the reason the $5.5 trillion figure keeps compounding instead of shrinking, not a technicality.

 

What Actually Changes the Trajectory

Closing this gap requires three specific shifts most organisations have not made, not a bigger training budget.

Put a named executive owner on AI capability risk, distinct from whoever owns general L&D, with the same reporting rigour as any other material risk on the register. Measure the capability gap in the terms the wage market already uses: roles you cannot fill, roles you are overpaying to fill, and work you are declining because you lack the people to do it, not completion percentages on a training platform. And treat the reskilling decision as time-sensitive capital allocation, where every quarter of delay is a quarter in which the 62% premium, and the competitors already paying it, get further ahead.

 

The Window Is a Cost Curve, Not a Deadline

There is no single date after which the AI skills gap becomes unrecoverable. What exists instead is a cost curve that gets steeper the longer it is ignored, priced daily by a labour market that has already decided what AI capability is worth.

The organisations that treat this as a 2027 problem will be paying 2026 prices for it well into the decade. The ones already moving are the ones setting the price.

Pre-Mortem: The Liability Chain Medicare’s AI Prior Auth Model Has Not Drawn

Posted on by

On 1 January 2026, the Centers for Medicare and Medicaid Services in USA launched the WISeR model in six states, introducing prior authorisation to procedures that traditional Medicare had always provided without it. Contracted companies now assess medical necessity using AI. Human clinicians are required to sign off on any denial. The Senate voted 46-50 in July 2026 to keep the programme running. One question has not been answered.

This is the twelfth piece in the Pre-Mortem series. Five questions, applied to the public record, before a programme has had the chance to succeed or fail.

 

The Bet

CMS is wagering that AI-assisted prior authorisation reduces unnecessary Medicare spend without producing the patient-safety incident that forces a political reversal. If WISeR delivers measurable waste reduction without a documented causal chain from AI denial to patient harm, it becomes the template for prior authorisation across Medicare nationally. If it produces that chain, a documented line from AI recommendation to denial to patient harm, it does not just end WISeR. It becomes the reference point that makes AI prior auth politically untouchable in federal health programmes for a generation.

 

The Assumption

CMS has answered every operational question about WISeR except this one:

When an AI recommendation leads a contracted clinician to deny care and a patient is harmed as a result, where does liability sit?

The model design places a human clinician between the AI output and the denial decision. That establishes a paper trail. It does not establish a liability framework. Contractors earn between 10 and 20 per cent of the savings generated by denials and lose that payment when a denial is overturned on appeal. That is a commercial penalty, not a clinical one. The Federal Tort Claims Act does not cover contracted entities. No federal court has tested whether a contracted clinician reviewing AI recommendations at volume carries the same duty of care as a treating physician making an independent clinical judgement.

The assumption doing all the work in this model is that the human review layer is accountability enough. That assumption has not been tested.

 

The Sequence

1 July 2025. CMS published the WISeR notice in the Federal Register and did not submit it to Congress under the Congressional Review Act. That omission would matter later.

1 January 2026. WISeR launched in New Jersey, Ohio, Oklahoma, Texas, Arizona, and Washington.

17 March 2026. The Washington Post published an exclusive: Medicare’s new AI gatekeeper was delaying care for seniors. The University of Washington’s medical system had nearly 100 patients waiting for epidural injections. In Arizona, Phoenix pain specialist Dr Matthew Crooks told Medscape that every epidural injection submitted in the first three months had been denied and described the system as completely nonfunctional and unsustainable. In Texas, initial AI approval rates ran at 62 per cent, against a 92 per cent national approval rate across Medicare Advantage.

25 March 2026. The Electronic Frontier Foundation filed a FOIA lawsuit against CMS in federal court in California, seeking records on WISeR’s AI algorithms, training data, bias safeguards, and the financial incentives paid to contractors. The suit confirmed that CMS had not made its AI methodology or vendor compensation structure publicly available seven weeks after launch.

6 April 2026. CMS published a Federal Register notice delaying prior authorisation implementation for certain services within the model to allow additional time for operational readiness. CMS also issued a corrective action order against one of its AI contractors. Both confirmed that the model’s operational design had not performed as intended in the first quarter.

12 May 2026. The Government Accountability Office issued its determination: WISeR met the Administrative Procedure Act definition of a rule and was subject to the Congressional Review Act. CMS had not made the required submission to Congress before the model took effect.

20 May 2026. Senator Ron Wyden and Representatives Suzan DelBene and Greg Landsman introduced resolutions of disapproval in both chambers, seeking to repeal WISeR under the CRA.

6 July 2026. Gold carding launched in Washington state. Providers achieving a 90 per cent affirmation rate across a minimum of ten prior authorisation requests become exempt from further review for covered services. Quarterly rollout to the remaining five states is planned.

16 July 2026. The Senate voted 46-50 against advancing the disapproval resolution. Party line. WISeR survived. The liability question the GAO had exposed survived with it.

The Pager

Dr Mehmet Oz, Administrator of the Centers for Medicare and Medicaid Services.

The message: WISeR’s accountability chain has not been drawn. The model places a contracted clinician between an AI denial recommendation and a Medicare beneficiary, but no published document establishes where negligence sits when a patient is harmed following an AI-assisted denial. The Federal Tort Claims Act does not cover contractors. Contractors point to the human clinician. Clinicians are reviewing AI output under volume pressure with no published duty-of-care standard for that specific context. When the first federal lawsuit tests this configuration, and one will, CMS will need a published framework, not a contract clause. That framework is easier to write before litigation than after.

 

The Proof

Gold carding is the model’s self-correction mechanism. If quarterly rollout reaches all six states and the 90 per cent affirmation threshold functions as a genuine quality signal, the AI layer contracts over time as trust is established. Proven providers exit prior auth. New entrants face the review. The model becomes calibrated rather than blanket.

If gold carding stalls or rollout criteria are applied inconsistently across jurisdictions, the AI layer expands without a release valve. Prior auth burden accumulates regardless of provider track record. The model becomes a cost-reduction instrument with no exit for providers who have earned one.

The proof of the bet is not the aggregate savings figure. It is whether WISeR, by the end of 2026, has published a liability framework and delivered gold carding in all six states. Without both, the model is running on the same untested assumption it started with.

 

Verdict

If CMS publishes a liability framework for AI-assisted denials before a federal case forces the question, and gold carding delivers consistent rollout across all six states, WISeR will be the strongest government evidence yet that AI-assisted utilisation review can reduce Medicare waste without a patient-safety crisis. The accountability design would become the reference for every federal health programme that follows.

Without the liability framework, WISeR accumulates its risk quietly. Not through a single dramatic incident, but through the gap between AI recommendation volume and human review capacity, compounded by an accountability vacuum no published document has yet closed. That gap does not stay open indefinitely.

Pre-Mortem: The US Government’s 3,611 AI Use Cases

Posted on by

On 3 April 2025, the White House issued OMB Memorandum M-25-21, directing every major federal agency to appoint a Chief AI Officer, expand the use of artificial intelligence across government operations, and manage risk proportional to each system’s impact on citizens. Twelve months later, the Federal Agency AI Use Case Inventory records 3,611 AI use cases across 56 agencies, more than double the prior year’s total. A May 2026 survey of more than 200 technology executives across civilian and defence agencies found 53% are actively planning agentic AI pilots. Only 8% of those agencies have incident response frameworks in place.

This is the eleventh piece in the Pre-Mortem series. Five questions, applied to the public record, before a programme has had the chance to succeed or fail.

 

The Bet

The US government is betting that embedding AI across 3,611 federal workflows covering benefits decisions, immigration adjudications, healthcare determinations, and law enforcement, will make government faster and more efficient before the accountability architecture governing those decisions is clarified. OMB M-25-21 requires Chief AI Officers, public AI inventories, and risk management proportional to impact. The hard compliance deadline for role-specific AI training arrives in September 2026. If that architecture catches up to the deployment before a consequential wrong decision reaches a citizen with no named relief, the bet holds.

 

The Assumption

The expansion’s credibility turns on one unanswered question: whether the Federal Tort Claims Act, designed to govern negligent acts by human federal employees, applies without amendment to decisions made by AI agents running inside federal systems. The same May 2026 survey found only 44% of agencies include vendor liability clauses in AI contracts, and only 29% have documented kill-switch procedures. The legal architecture governing accountability in federal government was designed for humans acting on behalf of the state. No court has ruled on whether it extends to the agents they built.

 

The Sequence

In 2024, federal agencies reported 1,757 AI use cases. By 2025, that figure had grown to 3,611. In March 2026, the Department of Veterans Affairs expanded AI use in claims processing, with 215 of its 367 AI systems classified as high-impact, covering benefit eligibility, healthcare access, and fraud detection. In May 2026, the majority of agencies were planning agentic pilots, with only 20% having defined pre-deployment testing policies. The AI reached citizens before the accountability reached the AI.

 

The Pager

Russell Vought, Director of the Office of Management and Budget, carries the M-25-21 mandate at the centre of the federal AI expansion. Every covered agency has designated a Chief AI Officer responsible for inventory, risk management, and AI governance at agency level. The VA alone runs 215 high-impact AI systems. No single published document names what relief is available to a veteran whose claim was influenced by one of those systems, which official carries accountability for that decision, or whether the Federal Tort Claims Act applies when the acting party is software, not a civil servant.

 

The Proof

The measure that would settle this is a published legal standard: a named accountability chain clarifying who carries liability when a federal AI agent makes a consequential wrong decision, whether government, vendor, or joint, and whether the Federal Tort Claims Act applies or new legislation is required. No such standard has been published. OMB M-25-21 requires risk management proportional to impact. It does not name the relief available to a citizen when that risk management fails, nor the date by which that question must be answered.

 

Verdict

If OMB publishes, before the September 2026 training compliance deadline, a named accountability standard for AI-driven decisions in high-impact federal systems, covering who carries liability when the AI is wrong and what legal remedy a citizen holds, the expansion will stand as the most deliberate attempt the US federal government has made to govern AI before it reaches citizens at scale. Without that, the US government has put AI into 3,611 workflows and left the question of who carries the call when the AI gets it wrong to be answered in court, by accident, or not at all.

Two-Thirds See AI Working. The CFO Still Can’t Prove It.

Posted on by

Two-thirds of organisations say AI is delivering real productivity gains. Ask the CFO whether that translates into a return they can defend, and the answer changes completely.

 

Two Surveys, Same Landscape, Different Question

Deloitte’s 2026 State of AI in the Enterprise report, surveying 3,235 senior leaders across 24 countries, found 66% of organisations reporting real productivity and efficiency gains from AI, and 53% reporting genuinely better insights and decision-making. That is not a marginal result. Two-thirds of a very large sample are seeing the operational benefit show up in how the work actually gets done.

EY’s 2026 Global DNA of the CFO Survey, covering 1,610 CFOs, finance directors, and heads of finance at organisations with over a billion dollars in revenue, was fielded in the same window and asked a different question: not whether AI is helping, but whether finance can prove it in the language capital allocation actually requires. The answer is uncomfortable. Just 12% of CFOs say their finance transformation outcomes exceeded expectations over the past two years. Only 21% describe their function’s AI readiness as leading or advanced. And 71% say plainly that traditional metrics are not enough to evaluate initiatives that combine people and technology.

 

This Is Not Two Surveys Disagreeing

Read carelessly, that looks like a contradiction: one report says AI is working, the other says it isn’t. It isn’t a contradiction. It is two different functions answering two different questions, and the gap between the answers is the actual story.

The operational teams reporting gains in the Deloitte data are measuring what they can see directly: faster cycle times, fewer manual steps, sharper analysis. None of that is fabricated, and none of it is trivial. But productivity gain and return on invested capital are not the same measurement, and the EY data shows finance has not built the bridge between them. Sixty-one per cent of CFOs cite data quality and bias as their top challenge in securing further AI investment, which is a polite way of saying the numbers underneath the business case are not yet reliable enough to defend in a capital allocation meeting.

A separate July 2026 survey of 1,505 senior finance leaders across the US, UK, Australia, and India found the same pattern from a slightly different angle: 92% feel active pressure to prove AI investment is paying off, while only half report their AI agents have actually achieved a measurable return. The pressure to prove value is running well ahead of the organisation’s actual ability to measure it.

 

The Real Governance Gap

None of this is a technology problem. The Deloitte numbers show the technology is doing what it was bought to do, in a majority of cases, across a very large sample. It is not really a capital problem either. Boards are still funding AI investment at pace, and EY’s own data shows CFOs largely expect that to continue.

It is a measurement problem, and measurement problems are governance problems wearing a data costume. Somebody has to own translating “the team says this is working” into “here is the return, measured the way capital allocation actually requires it to be measured,” before the investment decision, not after it, using metrics that were agreed while everyone could still agree on them.

Most organisations have not assigned that ownership to anyone specific. It sits, by default, somewhere between IT, who built the thing, and finance, who has to defend the number nobody built the measurement framework to produce.

 

What Actually Closes the Gap

Fixing this does not start with better AI. It starts with defining, before the next AI investment gets approved, exactly what “return” means for that specific initiative, in terms finance and the operational team both sign off on before deployment, not after. It means putting one named owner against that measurement framework, not a committee, and not IT by default because they happened to build the system. And it means accepting that a lot of the value AI is already producing is real but currently invisible in the metric finance is required to report against, which is a reason to fix the metric, not to distrust the value.

 

The Question Worth Asking Before the Next AI Business Case

The next time someone asks whether an AI investment delivered a return, the sharper question is whether anyone defined, in advance, what return was actually supposed to look like, and in whose language it would need to be proven. Most organisations running significant AI programmes right now cannot answer that question. That is the real gap the numbers are describing, and it is entirely fixable, starting with the next business case, not the last one.

Healthcare AI Enters Its Accountability Phase

Posted on by

Healthcare AI has stopped being an experiment. Holland & Knight, the US law firm, put it plainly in its mid-2026 healthcare report: the sector has entered a “recalibration phase,” where capital discipline and demonstrable return on investment have replaced the growth-first logic that funded the last five years of digital health.

That is a legal and investment framing, not a clinical one. But the clinical evidence backing it up is now specific enough to name.

Kaiser Permanente’s Permanente Medical Group rolled out ambient AI scribing to 7,260 physicians across more than 2.5 million patient encounters between October 2023 and December 2024. The result, confirmed by Kaiser’s own Division of Research: nearly 16,000 clinician-hours of documentation time saved. Not a pilot cohort. Not a vendor’s projection. A production deployment, measured after the fact, across a workforce large enough that the number means something.

Ambient documentation is also the part of healthcare AI with the least room left to argue about. A 2026 survey of 120 US health systems, run by the healthcare research firm Eliciting Insights, found clinical note-taking and ambient listening tools now sit at 68% adoption, up 62% year on year. Among the health systems able to quantify results, 61% report at least a 2x return specifically from ambient listening tools.

 

The $3.20 Figure Is Real, and Older Than It Looks

The oft-quoted “$3.20 return for every $1 invested in healthcare AI” is genuine, but it is worth knowing where it actually comes from before repeating it in a board pack. It traces to a Microsoft-sponsored IDC study published in late 2023 and reported in early 2024, not a fresh 2026 finding. It has simply become the industry’s standing benchmark figure, cited so often across 2025 and 2026 coverage that it now reads as current data. It is not wrong. It is just two years old and vendor-commissioned, which matters if you are the one deciding how much weight to put on it.

The Kaiser and adoption figures matter more, precisely because they are recent, specific, and independently reported rather than recycled.

 

What Actually Produced the Return

This ROI happened because of a specific programme design, not because someone bought a good tool, one that most other sectors experimenting with AI have not adopted.

Kaiser did not deploy ambient scribing and then discover the workflow around it. Clinical documentation workflow got redesigned first, and the AI tool was the mechanism, not the starting point. Accountability for the outcome, hours saved, adoption sustained, clinician trust maintained, was established before rollout, not retrofitted afterwards to justify the spend. And the whole exercise operated under exactly the capital discipline Holland & Knight describes: prove the return, or the funding does not continue.

That sequence, workflow redesign first, accountability from day one, capital discipline over growth optimism, is the actual explanation for why healthcare produced verifiable ROI while most other sectors are still producing pilot decks.

 

Healthcare Is Now the Benchmark, Not the Exception

Treat healthcare’s result as evidence that AI works and you will draw the wrong lesson. The technology was never really in question. What was in question, and what most other sectors are still failing to answer, is whether the organisation deploying it redesigned anything before switching it on.

Healthcare had no choice but to answer that question properly. Clinical documentation errors have consequences that show up in patient outcomes and malpractice exposure, not just quarterly numbers, so the sector could not afford the deploy-first governance-later approach that has quietly become normal everywhere else.

That is what other sectors should actually be benchmarking against: not whether their AI produces a return, but whether their programme was ever designed to make one provable.

 

The Question Worth Asking Before the Next AI Business Case

Before signing off the next AI investment, the question is not whether AI delivers value. Healthcare has already answered that question, under specific and now well-documented conditions.

The real question is whether your programme has been designed to match those conditions, workflow redesign before deployment, accountability defined from the outset, capital discipline over growth optimism, or whether it has been designed the way most digital health investment was designed before 2026: fund it, hope the outcomes show up eventually, and find out later whether anyone was ever going to check.

Healthcare already found out. That is the whole difference.

Pre-Mortem: The EU AI Act’s Accountability Gap

Posted on by


On 2 August 2026, the EU AI Act gives the EU AI Office the power to fine the developers of general-purpose AI models up to three per cent of global annual turnover, demand documentation, and commission independent access to source code. Three weeks before that date, the high-risk AI compliance deadline moved from August 2026 to December 2027, enacted as binding law on 29 June. The two facts share a date. They do not share a plan.

This is the ninth piece in the Pre-Mortem series. Five questions, applied to the public record, before a programme has had the chance to succeed or fail.

 

The Bet

The EU is betting that extending the deadline for high-risk AI compliance by 16 months, agreed in May 2026 and enacted on 29 June, produces better enforcement outcomes than a met deadline inside a half-prepared enforcement architecture. The logic holds. As of August 2026, only nine of 27 member states have shown advanced public implementation of enforcement infrastructure. Germany has designated the Bundesnetzagentur as its market surveillance authority and adopted draft transposition legislation. Spain built the AESIA, a dedicated supervisory agency, from scratch. Ireland deployed fifteen coordinated authorities under a central National AI Office. Those are genuine structural commitments. Eighteen member states have not reached that point. The extension gives them time. Whether they use it is the bet.

 

The Assumption

The entire framework rests on this: that national competent authorities, operating under 27 different legal frameworks, will converge on consistent enforcement before December 2027. The AI Act is a directly applicable regulation. Its enforcement infrastructure is not. The regulation sets the rules uniformly across the bloc. The authorities responsible for applying them have been built at very different speeds, under very different political conditions. That divergence is the risk the extension is buying time to close. There is no public commitment that the time is sufficient.

 

The Sequence

The AI Act entered into force in August 2024. Member states were required to designate their national competent authorities by August 2025. At least twelve missed that deadline. Seven months later, in May 2026, the Council and Parliament agreed to simplify the rules as part of the Digital Omnibus package. On 29 June, the high-risk AI deadline moved. What remains in force on 2 August is a narrower set: general-purpose AI model obligations and transparency requirements for new deployments. The high-risk AI rules, the Act’s original centre of gravity, are no longer in that set. Governance was adjusted to fit the readiness gap. That is not the order in which enforcement architecture is supposed to be built.

 

The Pager

Lucilla Sioli, Director of the EU AI Office, carries accountability for general-purpose AI enforcement from 2 August. For high-risk AI systems, including credit-scoring models, recruitment tools, and systems used in border control, healthcare, and law enforcement, accountability rests with national competent authorities. In 17 of 27 member states, no public designation exists. The Act names the category. Seventeen member states have yet to name the person.

 

The Proof

The measure that would settle this in 2028 is year-one enforcement consistency: the share of member states that have conducted at least one formal high-risk AI enforcement action, under the same evidentiary standard, in the first twelve months after the December 2027 deadline. No EU institution has publicly committed to publishing that figure. The AI Office’s annual progress reporting is the closest mechanism on the public record. It tracks activity. No published mechanism commits to measuring whether enforcement actions are consistent across member states.

 

Verdict

If the Commission designates a public accountability owner in each member state before December 2026 and commits to publishing year-one enforcement data by name, the 16-month extension holds up as a governance decision made under realistic conditions. Without that, a framework that took two years to reach enforcement hands itself an extension with nobody carrying it.

Deploy Now, Govern Later as a Strategy Just Expired

Posted on by

.

Nearly three-quarters of companies are planning to deploy agentic AI within two years. Only 21% of them report having a mature model to govern it.

That gap, not a headline percentage on its own, is the structural condition enterprise AI now operates under, according to Deloitte’s 2026 State of AI in the Enterprise report. The same report found that sanctioned AI tool access has grown 50% in a single year, from under 40% to around 60% of workers. Read those two findings together and the picture is not ambiguous: deployment is accelerating faster than governance can follow, and roughly four in ten workers still operate without any sanctioned AI tool at all, which is exactly the population most likely to reach for something unapproved.

Framing this as a planning problem, something to address in the next cycle, stopped being an accurate read of the situation in the first week of July 2026.

 

Why the Timing Changed, Not the Substance

The EU AI Act’s Article 50 transparency obligations take effect on 2 August 2026. The Five Eyes intelligence alliance issued a joint statement on 29 June warning that frontier AI could transform both cyber offence and defence “in months, not years,” with attackers already moving from initial access to data theft in under 72 minutes. Available coverage of the statement does not indicate that it singles out enterprise AI tools by name as an attack surface. What both point to is the same underlying condition: the assumptions organisations built their cyber-risk models on are ageing out faster than those models are being revised, and AI deployment is a large part of why.

None of these three developments is new information arriving out of nowhere. Article 50 was always coming, and its requirements have been public for months. What changed is the simultaneity: a regulatory deadline with a fixed date, an intelligence community warning about compressed attack timelines, and a governance maturity figure that puts a number on the gap between what is deployed and what is actually controlled. Those pressures used to arrive on separate timelines. In July 2026 they are concurrent.

 

The Decision Behind the Gap Was Rational

The governance gap did not happen through neglect. Most organisations deploying agentic AI without a mature governance model made a deliberate trade-off: move now, build the governance model once the technology and the internal use cases stabilise. That calculation made sense through most of 2025. Early movers captured a real advantage, and governance frameworks built around technology that was still changing weekly risked being obsolete before they were finished.

That trade-off does not survive contact with August 2026 intact. The regulatory deadline is fixed. The security environment has compressed. And the governance figure, 21% with a mature model against a 75% deployment intention, is no longer a benchmark to compare against competitors. It is a description of where the exposure actually sits inside your own organisation.

 

What Actually Needs to Happen Now

For transformation leaders, this does not resolve into a disclosure for the board. It resolves into a specific, immediate piece of work: an accurate inventory of what AI is actually running across the organisation, not what was approved on a policy document, but what is deployed and in active use. The distance between those two lists is precisely the exposure that Article 50 and the current threat environment are now positioned to surface.

That inventory has to happen before the governance model gets built, not alongside it. You cannot govern a system whose actual footprint your organisation has not yet measured, and by the time an auditor, a regulator, or an attacker measures it for you, the cost of closing the gap has already changed.

The deployment number will keep climbing. The governance number moves only when someone decides to move it. Right now, for most organisations, no one has.