Open Water

Europe has given companies more time. The Digital Omnibus on AI, in force since 27 July, moves the AI Act’s high-risk obligations from August 2026 to 2 December 2027 for stand-alone systems, and to 2 August 2028 for AI built into regulated products such as medical devices.

For executives, the temptation is to treat the extension as a reprieve. The obligations themselves are unchanged: risk management, data governance, technical documentation, logging, human oversight and, in some cases, a fundamental rights impact assessment before a system goes live. None of that is a form to fill in the month before the deadline. It is architecture.

The work breaks in the gaps between teams. Legal reads the Act, engineering builds the product, procurement buys the model, and nobody owns the evidence trail that connects them. Retrofitting that trail onto a system already in production is expensive; designing it in costs a fraction.

Economists will recognise a familiar choice between doing something once, properly, inside the firm, and paying repeatedly to patch it later. The practical answer is to treat compliance as a product requirement: an inventory of AI systems and their risk class, logging and documentation built into the delivery pipeline, and a named owner for each system.

It must be said that waiting has some value. Harmonised standards are still being finalised, and building to a moving target carries its own cost. But the direction is clear, and the architecture holds whatever the detail.

Companies that use the extra sixteen months will reach December 2027 with systems they can sell to regulated buyers across the EU. Those that do not will discover that compliance, like a building’s wiring, is cheapest before the walls go up.

The open question is which of the two most AI programmes resemble today.

The Prime Minister has put social care at the centre of his health agenda, arguing that the absence of a preventative “front end” to social care costs the NHS several billion pounds a year. Whatever the eventual funding model, much of that cost already shows up in a familiar place: patients medically ready to leave hospital who cannot, and people admitted who might have been supported at home.

For trust and ICB executives, the operational question arrives well before any national reform. A large share of the flow problem at the hospital door is an information problem at the boundary.

The process explains why. Discharge teams, community services, councils, care providers and family carers each work from their own systems. The patient’s situation is reassembled by phone and email at every handover, and each handover costs days. Economists would call these transaction costs; patients experience them as waiting.

Digital tools can carry information across that boundary: shared care plans, discharge status visible to every party, AI summaries prepared for the receiving team. In West Yorkshire, the ICB has commissioned an app that lets unpaid carers record a contingency plan in the NHS patient record using NHS login, so clinicians can see what should happen if the carer cannot continue.

The practical route is one cohort and one boundary at a time: discharges from a group of wards to care homes, for example. Map the handover, digitise the information that travels, and measure the days recovered.

It must be said that digital tools cannot create care capacity that does not exist. Social care workforce shortages are real. But tools can stop existing capacity being wasted while information catches up.

Our hunch is that the first gains from social care reform will come not from legislation, but from the handover.

Since February 2025, the EU AI Act has required organisations that build or deploy AI to ensure a sufficient level of AI literacy among their staff. This summer’s Digital Omnibus softened the wording without removing the duty. Article 4 now requires providers and deployers to take measures that support the development of AI literacy, rather than to guarantee any particular level of it, and national authorities enforce it.

That makes the duty one of effort, not of result. It also makes it a duty of evidence. When a regulator, customer or auditor asks what an organisation did, the answer will need to be specific: which staff use which systems, what they were taught, and when.

The temptation is a generic e-learning module for everyone. It satisfies a checklist and changes little. The people who matter most are the ones who approve outputs, configure tools and handle exceptions, and what they need differs by role and by system. A claims handler reviewing an AI recommendation needs different knowledge from an engineer adjusting a model’s prompts.

The better approach ties literacy to deployment. Each system goes live with a short, role-specific briefing on what it does, where it fails and when to escalate, and a record that the briefing happened. When the system changes, the briefing changes with it.

It must be said that records alone do not make anyone literate. Training designed for an audit trail can become theatre. The real test is whether people catch the errors the system makes.

Organisations that get this right gain more than compliance. Staff who understand a tool’s limits use it more, and more safely, which is where the return on AI is actually earned.

Our hunch is that the most useful record of AI literacy is a mistake that someone caught.

Almost every company now has access to the same AI. Adoption has stopped being a differentiator; the business model has become one. Nowhere is this sharper than in businesses that bill for effort.

Law, accounting, agencies, consulting and IT services all face the same arithmetic. When the hours shrink, so does the invoice, unless the pricing changes first.

Effort-based pricing was always a solution to a measurement problem. Outcomes are hard to define and verify, so suppliers priced inputs instead. AI removes much of the input, and the measurement problem returns. Pricing on outcomes means agreeing what the outcome is and checking it was delivered, and both carry costs.

For executives running these businesses, the decision is uncomfortable: pass the AI savings on to clients, or keep the margin? Competitors will make that choice for them if they don’t.

The team structure is exposed too. The classic pyramid, where junior hours carry the profit, is hollowing out. Firms must work out how juniors learn when AI does the junior work.

It must be said that outcome pricing shifts risk to the supplier, and many will price that risk generously. Buyers, for their part, can reasonably ask what AI has done to a supplier’s cost base, and who is keeping the difference.

The effect is not confined to firms. Whole regional economies built on labour arbitrage, India’s IT services sector chief among them, face the same question at national scale.

Our hunch is that the next great disruption in professional services will not be a new firm, but an old invoice.

British companies building AI for European customers now work under two different philosophies of regulation. The EU has one horizontal law, the AI Act, with obligations that depend on a system’s risk class. The UK has chosen the opposite route: no AI Act and no AI bill before Parliament, with existing regulators such as the ICO, the FCA and the MHRA applying their own rules to AI within their sectors. The King’s Speech in May announced a Regulating for Growth Bill, built around regulatory sandboxes, rather than an AI law.

For executives, the question is practical. A product sold on both sides of the Channel faces two sets of expectations, and maintaining two compliance designs doubles the cost of every change.

The work tends to break where teams map each rulebook separately, producing parallel documents that drift apart. Engineers then build to whichever requirements were written down most recently.

Economists will recognise a question about where to put fixed costs. A single control framework, designed to the stricter EU standard and then mapped to each UK regulator’s expectations, is built once and reused. Logging, documentation, human oversight and data governance done to the AI Act’s level will usually satisfy a UK regulator asking about fairness, transparency or accountability.

It must be said that building to the stricter rulebook can over-engineer products sold only in the UK. Where a system will never leave the domestic market, lighter controls may be rational.

But few growth companies plan to stay in one market for ever, and buyers increasingly ask for evidence whatever the law requires. Companies that comply once, to the higher bar, sell into both markets on the same product.

The question is whether two rulebooks mean two systems, or one system with two maps.

The NHS Productivity Commission’s technology paper, published in September, puts it plainly: the constraint on technology-led productivity is not a shortage of solutions but adoption, implementation and spread.

Anyone who has run a successful NHS pilot will recognise the problem. The pilot works, the evaluation is positive, and the learning leaves with a slide deck. The next trust starts again.

Transaction cost economics explains much of this. Each of England’s trusts assessing the same tool, writing its own clinical safety case, negotiating its own integration and building its own benefits case multiplies the cost of adoption many times over. The national supplier registry for ambient voice technology, which asks suppliers to evidence clinical safety, technical and data protection standards, is an attempt to pay some of those costs once.

The executive incentive points the same way. Trust leaders are recognised for local innovation; few are rewarded for making someone else’s pilot work in their organisation. Replication has no natural owner.

The fix is in how pilots are designed. A pilot built to be copied ends with a package: the safety case, the integration specification, the training materials and the benefits method, ready for the next organisation to pick up. Reusable integration standards matter as much as the AI itself.

It must be said that some local variation is legitimate. Populations, estates and systems differ. But much variation is simply the cost of starting from scratch.

Suppliers will adapt too. Those arriving with NHS evidence packs and proven integrations will outpace those arriving with demos.

We might ask how many NHS pilots were designed, from the first day, to be copied.

Since 5 February, UK organisations have had more room to let machines decide. The Data (Use and Access) Act replaced the old near-prohibition on solely automated decisions with a permission, subject to safeguards. The exception is special category data, including health information, where tighter rules still apply.

The safeguards are not optional. People must be told when a significant decision about them is automated, and be able to make representations, obtain human intervention and contest the outcome.

The interesting question is what counts as a human. The law asks whether there is meaningful human involvement in a decision. A reviewer who clicks approve on every recommendation is not meaningful involvement; the decision is still, in substance, automated. The Information Commissioner’s Office, reviewing AI in recruitment, found many organisations claiming human review that was, in practice, superficial.

This is a principal and agent problem. Reviewers measured on throughput have every incentive to agree with the machine. Genuine review means giving reviewers the information, time and authority to disagree, and tracking how often they do. An override rate of zero is not a sign of a perfect model.

The technology can help. Systems can show the factors behind a recommendation, route low-confidence cases for closer review, and log every override with its reason. Those logs become the evidence that the safeguards work.

It must be said that more automation, properly safeguarded, can be fairer than tired humans making inconsistent calls. The reform recognises that.

Organisations that design review properly will be able to automate more, and with confidence. Those that rely on rubber stamps carry a risk they cannot see until a complaint arrives.

We might ask how often the humans in the loop have ever said no.

As AI models converge in capability, the choice of model matters less each quarter. What matters more is the harness around it: the data, rules, workflows and checks that turn a general model into something that works for one particular company. It is also what determines whether agents are reliable enough to scale.

This is the old make-or-buy question, and the answer is unusually clear.

Transaction cost economics offers a simple rule: buy what is generic, make what is specific. Models are becoming generic, with a new frontier release every few months. The harness is not. It is valuable only to the company that uses it, and costly if someone else controls it.

That has a practical consequence for executives choosing platforms. A vendor platform bought wholesale often includes the layer that encodes the company’s own processes. Buying it means renting one’s own way of working.

A good harness is not built by engineers alone. The claims handler knows the exceptions; the operations manager knows which rules get bent and why. That knowledge has to be designed in.

Technically, the principle is modularity: the model underneath should be swappable while the data and logic stay put, and owned.

It must be said that building has real costs in talent and maintenance. For genuinely generic tasks, buying is right.

But competitors can buy the same model tomorrow. They cannot buy the harness.

We might ask how many companies, having outsourced their IT in the 2000s and spent years bringing it back in-house, are about to repeat the exercise with their intelligence.

Financial regulators have spent two years asking who would keep the lights on if a critical technology supplier failed. Under the EU’s Digital Operational Resilience Act, in force since January 2025, supervisors designated the first 19 critical technology providers in November 2025, from cloud platforms to data centres, for direct oversight. The UK runs a parallel regime for critical third parties.

The same question now applies to AI models. Many companies have built their AI capability on a single model provider: prompts tuned to its behaviour, evaluations calibrated to its outputs, workflows assuming its pricing. That concentration is rarely a deliberate decision. It accumulates one convenient choice at a time.

This is the classic hold-up problem. The more an organisation invests in assets specific to one supplier, the weaker its position when prices rise, terms change or a model is withdrawn. Resilience rules address the risk at the level of the financial system; inside the firm, it is an executive decision.

The technical answer is portability by design. An abstraction layer lets workflows call different models without rewriting them. Evaluation suites that test outputs against the organisation’s own standards, rather than one model’s quirks, make a switch measurable. A second model, tested on real tasks, turns an exit plan from a document into an option.

It must be said that multi-model architectures cost more to build and run, and a single provider often delivers better performance today. For low-stakes uses, concentration may be a reasonable bet.

For processes that matter, the calculation differs. Customers and regulators will increasingly ask what happens if a supplier fails, and a credible answer takes months to build.

Our hunch is that the companies best placed to negotiate with model providers are the ones that could leave.

NHS England’s trial of Microsoft 365 Copilot across 90 NHS organisations found staff saved an average of 43 minutes a day. A full rollout, it estimates, could save up to 400,000 hours of staff time a month. With access now extending to clinicians and support staff, those minutes are about to multiply.

NHS finance directors know a distinction the headline figures skip: time released is not cash released. Forty-three minutes spread across a ward, a clinic and a back office does not by itself fund a post, open a theatre list or shorten a waiting list.

That leaves trust executives with a choice. Saved time can be banked, through lower bank and agency spend. It can be redeployed into more clinic slots or faster discharge. Or it can be given back to staff as working lives that are less stretched. Each is legitimate. Choosing none means the time simply disappears.

The conversion happens in the process: job plans, rotas and clinic templates. If a template still books the same number of patients, the saved minutes never reach one.

The technology can help show where the time goes. Rostering data, activity data and EPR audit logs can reveal whether released time is turning into capacity, or into emails.

It must be said that giving clinicians their evenings back is not waste. Burnout shows up in sickness absence and turnover, both expensive. But it should be a decision, not a by-product.

Acute productivity rose 2.7% in 2024/25, ahead of the 2% target. Trusts that turn released time into measured capacity will make the case for the next investment.

We might ask how many of those 43 minutes have yet reached a patient.

Explainability is often treated as a philosophical question about how AI models think. For organisations that use AI to make decisions about people, it is becoming an operational one.

Under the EU AI Act, high-risk systems must keep logs that allow their operation to be traced, an obligation that applies from December 2027, and people affected by decisions taken with such systems have a right to a clear explanation of the role the AI played. In the UK, people subject to significant automated decisions can already contest them, which only works if someone can say why the decision was made.

The executive decision is how much to invest in the machinery of explanation before it is demanded. The pressure point is usually a single complaint, audit or regulator’s letter asking what the system did, with which data and which model version, on a given day. Organisations that cannot answer must reconstruct events from scattered records, at great cost and with little confidence.

The work breaks between teams. Data scientists track model versions, operations teams own the workflow, and nobody owns the record linking a decision to both. The solution is a decision trail designed in from the start: the inputs, the model and version, the output, any human review and the final outcome, stored together and retrievable.

It must be said that complete explanations of complex models remain out of reach, and collecting too much data carries its own privacy risk. The goal is accountability, not a full account of every parameter.

Done well, the trail pays for itself. It speeds investigations, supports audits and lets organisations extend AI into decisions they could not otherwise defend.

The test for any AI programme is simple. Could it explain last Tuesday?

An AI agent can follow every instruction, stay within every permission, and still damage the business. It is one of the least discussed risks in enterprise AI, and one of the most predictable.

Each agent is, in effect, a small firm inside the firm. It takes decisions on delegated authority but carries none of the accountability. This is the principal-agent problem in its purest form. Human agents are restrained by reputation and career; software agents have neither. They are perfectly literal and perfectly indifferent.

So a new question lands on the executive’s desk: when an agent acts within its permissions but against what the business intended, whose decision was it? Usually, nobody has been named.

The failure point is rarely the model. It is the hand-off between the business owner who describes an outcome and the engineer who encodes a rule. Intent is lost in between, and the executive who approved the deployment sees neither side.

Crypto learned this first. The 2016 DAO exploit was permitted by the code, yet nobody intended the outcome. “Code is law” proved to be a governance failure rather than a principle.

The remedy is to give each agent a mandate as one would a new hire: narrow scope, measurable outputs, a named owner and a clear way to revoke access, with explainability built in from day one.

It must be said this slows the first deployment. It speeds every one after, because risk teams approve what they can see.

Our hunch is that the firms scaling agents in two years’ time will not be those with the cleverest models, but those that can say, for each agent, who it works for.

When the state uses an algorithm to make or shape decisions about people, the public has an interest in knowing. The UK has turned that principle into practice. The Algorithmic Transparency Recording Standard requires central government departments, and many arm’s-length bodies, to publish a record of the algorithmic tools that significantly influence decisions or interact with the public: what each tool does, why it is used, how it works and how its risks are managed. Around 125 records had been published by the spring. Across the wider public sector, including the NHS, the standard is recommended rather than required.

For technology suppliers, this changes the sale. A public buyer that must publish a record needs information only the supplier holds: the data the tool uses, how it was tested, where it may fail and who oversees it. Suppliers that cannot provide it make the buyer’s job harder, and become harder to buy.

Economists will recognise the information asymmetry at the heart of public procurement. Buyers struggle to judge what they cannot see, so they discount what they cannot verify. Transparency reduces that discount.

The practical answer is to treat the transparency record as part of the product: an up-to-date description of each tool, its data, its testing and its limits, written for a public audience and ready to hand to any buyer.

It must be said that disclosure raises legitimate concerns about commercial confidentiality and security, and the standard allows exemptions. Transparency does not mean publishing source code.

The direction of travel is clear, and it extends to buyers not yet required to publish. Suppliers that are ready will find doors open more easily.

Public money, public algorithms. The question is which suppliers have already written their record.

Who controls NHS data, and where it sits, has become an executive question rather than a technical one. Devolution is back on the agenda, ICB boundaries are to be redrawn to match local government footprints by 2029, and a decision on the national data platform contract is due by December, ahead of a break clause in March 2027.

Greater Manchester offers one precedent. Rather than adopt the nationally mandated platform, its NHS leaders spent six years building their own analytics infrastructure.

For trust and ICB executives, the choice is between national platforms, local builds and combinations of the two. It is a decision with a long life, because data infrastructure is the most organisation-specific asset a health system owns. Transaction cost economics warns that whoever controls a specific asset holds the bargaining power over it for years: buy what is generic, own what is specific.

Ownership alone achieves little, however. Data is only useful with the quality, governance and analytical people to use it. Information governance leads and analysts belong in the design from the start, not the sign-off at the end.

The technical principle is modularity. A local integration and analytics layer built on open standards can work alongside national platforms whichever direction national policy takes, so the organisation is never forced to start again.

It must be said that local builds can fragment. National platforms bring economies of scale and shared standards, and the answer is rarely all one or the other.

Organisations that own their integration layer will adapt to whatever December brings. Those that do not will adapt to whatever their suppliers decide.

Our hunch is that the most consequential NHS data decisions this year will be made not only in Whitehall, but in the architecture choices of individual trusts and systems.

The NHS App is becoming the front door to the NHS. Government plans have patients viewing their single patient record through it, and NHS England is rolling out AI triage in the App to all users by April 2028, after a pilot at one Sussex practice cut phone queues by 29%.

But a front door is not the whole house. Behind it, dozens of partners already provide services: online consultation suppliers, hospital patient portals, messaging services and personal health records, each connected through NHS login and national integration routes.

The economics are those of a platform. The NHS holds the assets nobody else can replicate: identity through NHS login, the record and the audience. Partners supply specialised services around them. Coase would recognise the arrangement: the NHS keeps inside what only it can do well, and buys the rest.

For trust executives, that changes an old decision. Building a standalone patient portal now competes with a front door patients already use. Integrating with it usually wins.

For healthtech suppliers, the map matters. GP online consultation is crowded, with twelve suppliers already live in the App and national AI triage on the way. Care plans, carers’ tools, follow-up after discharge and specialist pathways are far less so.

The route in is demanding. NHS login comes first. Integration requires meeting the App’s standards on accessibility, usability, clinical safety and data privacy, and direct integrations require commissioning through a recognised framework.

It must be said that platforms change their rules, and depending on a single front door carries its own risk. The sensible design works inside the App and outside it.

We might ask which services will still need their own front door in 2028, and which will be glad of someone else’s.

AI changes how fast each stage of product development can run. The companies seeing faster delivery tend to be those that have redesigned the stages themselves.

Economic historians will find this familiar. When factories first swapped steam engines for electric motors, productivity barely moved for decades. The gains came only when factories were redesigned around the new power source. Replacing the engine was easy; rearranging the floor was the hard part.

The executive choice is similar today. A tool is a purchase; a redesign is a reorganisation. Both are needed, and the second carries most of the value.

Speed up coding and the bottleneck moves: to discovery, design, approvals, testing and release. Accelerating one stage piles work up at the next unless the whole flow changes with it.

Teams change shape too. When a product manager can prototype, a designer can ship and an engineer can test ideas with customers in an afternoon, the old hand-offs become the constraint. Smaller, cross-functional teams with fewer hand-offs move faster than larger ones with more specialists.

The technology has to keep pace. If prototypes take days, decisions must take days too. Release pipelines and testing need automating, or they become the new constraint.

It must be said that redesign is disruptive, and best done one product line at a time rather than all at once.

The prize is a shorter idea-to-customer cycle. In markets where features are copied within weeks, cycle time is the moat. The question worth asking is how long an idea waits between stages, and who, exactly, it is waiting for.

Diagnostic AI gets the headlines. The productivity, so far, is often coming from somewhere duller. The Royal College of Radiologists’ latest census found only 11% of departments use AI to draft reports, yet almost a quarter of those report measurable workload reductions, the strongest productivity result of any radiology AI application.

The pattern extends well beyond radiology. Clinical coding, waiting-list validation, correspondence, referral management, contract review and rostering all involve large volumes of structured, repetitive work. The regulatory hurdles are lower than for clinical decision-making, and the payback is quicker.

For chief financial officers and chief operating officers, that makes administrative AI the natural first move. It fits within the financial year, and its benefits can be counted in hours, backlogs and spend rather than inferred from outcomes years later.

The process question is about people. Much of this work is done by administrative teams who know its exceptions better than anyone. Their roles will change, and the redesign works best when they lead it rather than receive it.

The technology question is access. Administrative AI is only as useful as its connection to the patient administration system, the EPR and the finance ledger. Much earlier automation stalled for exactly this reason.

It must be said that administrative AI still handles patient data, and the same governance applies. Lower clinical risk is not the same as no risk.

The prize is capacity released where it is easiest to measure. Waiting-list validation alone can return appointments to patients who need them.

Our hunch is that the first AI to move NHS productivity figures materially will be one patients never see.

Much of the productivity AI creates sits with individuals: the employee who drafts faster, researches faster, finishes earlier. Turning that into a gain for the company is a separate task, and a harder one.

Economists will recognise the pattern. The surplus goes to whoever controls it. An employee who finishes the day’s work by lunch has no reason to announce it, and some reason not to: the reward for surfacing a saving is usually more work. This is not idleness. It is incentive design.

The process adds to it. Individual AI use is largely invisible: personal prompts, private shortcuts, a workflow only one person understands. The knowledge lives in chat histories, and leaves when its owner does.

The technical step is to move from personal assistants to shared systems: prompts and agents owned by the team, documented, versioned and built into the tools everyone already uses.

It must be said that forcing this can backfire. Heavy-handed monitoring simply drives use underground. Clarity works better: a global survey of nearly 12,000 employees and managers this year found that a clear AI strategy improves results even where access to tools is limited. People share what they are rewarded for sharing, so recognition has to follow team-level gains, not just individual ones.

Companies that turn individual gains into shared systems will compound them. Those that don’t will leak capability every time someone resigns.

Our hunch is that the most valuable AI asset in many companies is sitting in a handful of employees’ chat histories. It would be worth knowing whose.

AI models now span a wide price range. Published list prices this summer put a leading frontier model at about $5 per million input tokens and $30 per million output tokens, against $0.20 and $1.25 for a lightweight model. For most companies, deciding which model does which task has quietly become a material cost decision.

Nobody, the old saying went, got fired for buying IBM. The modern equivalent is defaulting to the frontier model for everything. It is safe, defensible, and rather like sending every parcel by courier.

In practice, classifying an email, extracting a date and drafting a legal argument often go to the same model at the same price, because nobody has mapped tasks to models. Teams experiment freely, which is healthy, with no view of cost per outcome, which is not. Some companies have exhausted annual AI budgets within months.

The technical answer is a routing layer that sends routine work to small, cheap models and reserves the expensive ones for genuine reasoning. Unit costs fall sharply. Quality, for most tasks, does not.

Ownership of that layer matters as much as its design. Transaction cost economics has a name for the risk of depending on a single supplier for something a business cannot easily replace: hold-up. A company tied to one AI provider absorbs every price change. One that owns its routing can switch, and can therefore negotiate. With prices moving fast in both directions, the ability to switch is worth more than any discount.

Finance teams scrutinise every other unit cost. We might ask why AI, among the fastest-growing, should be the exception.

Ambient voice technology is spreading faster than almost anything the NHS has adopted. University Hospitals of Leicester and Northamptonshire are deploying it to more than 10,000 clinicians across some 2.5 million outpatient appointments a year. NHS England in the Midlands has procured it for around 70,000 clinicians and 1,239 GP practices.

The listening and transcribing is now the easy part. The decisions that determine value lie elsewhere.

The first is the route in. Every trust that procures alone builds its own clinical safety case, business case and evidence base, and pays for it. The Midlands built those once so other organisations would not have to start from scratch. That is transaction cost economics applied to procurement, and it is the quiet reason regional routes move faster.

The second is the review step. AI does not remove documentation work so much as move it. A recent London study found AI creating new responsibilities around oversight, validation and accountability. Someone must check the note, correct it and sign it. Designed badly, review eats the time the scribe saved.

The third is integration. A transcript is not a record. Value arrives when the draft letter lands in the right field of the EPR, coded and ready to send, rather than in a separate window waiting to be copied across.

It must be said that the market is racing ahead of the rules. The King’s Fund has called for a national strategy to make adoption safe and consistent.

The better measure is not minutes of typing saved but the time from consultation to letter received. Our hunch is that the trusts gaining most will be those that measured the letter, not the transcript.

AI is now in almost every company; profit from it is not. Recent industry research puts regular use at 89% of organisations, with only 37% reporting any positive impact on EBIT and 93% saying they have exceeded their AI budgets. The distance between use and return is where the interesting decisions sit.

The first is ownership. AI spend tends to be approved tool by tool, licence by licence, and no single executive owns the P&L line it is meant to move. When everybody is responsible for AI value, nobody is. Enterprise software went much the same way in the 1990s.

The second is measurement. Usage is easy to count (seats, logins, prompts), so it gets counted. Outcomes are harder, so they often go unmeasured, and a rising adoption chart stands in for a return. Dashboards, it turns out, are not dividends.

The third is placement. Many deployments sit beside core systems rather than inside them. They can summarise a contract but not price one, and draft a claims letter but not settle the claim. Money is made or lost in the transaction, so that is where AI has to reach.

The companies converting use into return tend to share three habits: a named owner for each deployment, one P&L metric agreed before approval, and integration into the systems of record rather than around them.

It must be said that integration is slower and riskier than a standalone pilot. It is also where the returns are.

Investors are beginning to ask about AI margins rather than AI announcements. Expect the question on earnings calls, and expect it to be specific.