From AI Experimentation to Enterprise Value

The gap that isn't closing

Enterprise AI has an odd scoreboard. Almost everyone is playing, almost everyone reports individual wins, and the enterprise-level result has barely moved.

McKinsey's August 2026 global survey of 1,719 executives puts numbers on it. Nearly nine in ten organizations use AI regularly in at least one function. Eight in ten respondents say AI has made them personally more productive. Yet only 37% attribute any EBIT impact to AI, essentially unchanged from 2025, and the "high performers" who attribute at least 5% of EBIT to AI remain stuck at about 6% of respondents.

Other data points describe the same pattern from a different angle. Gartner predicted in mid-2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. S&P Global Market Intelligence found that 42% of companies abandoned the majority of their AI initiatives in 2025, up from 17% a year earlier, and that the average organization scrapped 46% of proofs of concept before production.

The tempting conclusion is that the technology is overhyped. That misreads the data. Individual productivity gains are real and widely reported. The problem is that a productivity gain is not a financial result until someone redesigns the work, the budget, or the offer around it. Most AI programs stop one step short of that.

Why a successful pilot proves so little

A pilot is designed to answer one question: can the model do this task acceptably? By 2026, the answer to that question is yes for a very wide range of tasks. That is precisely why it has stopped being informative.

The questions that determine value are different, and pilots rarely test them:

• Does performance hold on the long tail of real cases, not the curated sample?
• Will the people doing the work change how they work, and will it stick?
• Does the cost per transaction stay below the value per transaction at full volume?
• When the system is wrong, who is accountable, and how quickly is it caught?
• Where in the P&L does the gain land, and who has agreed to bank it?

Between the demo and the income statement, value leaks out at four predictable points.

Leak 1: The long tail

Pilots run on representative samples. Production runs on everything else too: the edge cases, the angry customers, the ambiguous documents, the policy exceptions. A system that handles 85% of cases brilliantly can still create net work if the remaining 15% are harder to detect and costlier to fix than before.

Klarna is the most instructive public example, precisely because its early results were genuinely strong. In February 2024 the company reported that its AI assistant handled 2.3 million conversations in its first month, two-thirds of all customer service chats, doing work equivalent to 700 full-time agents and resolving issues in under two minutes rather than eleven. By 2025, its chief executive was publicly describing a renewed push to recruit human agents so that customers could always reach a person. The assistant did not stop working. The organization discovered that throughput was the wrong single measure of success, and that the experience on the hard cases mattered to the brand more than the averages suggested.

Leak 2: Adoption that never becomes redesign

When an AI tool saves an analyst forty minutes a day, those minutes are usually absorbed: a slightly longer lunch, a more thorough review, one more meeting. Nothing is wrong with that, but it is not EBIT.

This is the single strongest differentiator in the McKinsey data. Nearly three-quarters of AI high performers report fundamentally redesigning workflows because of AI. Among everyone else, the figure is about one in four. High performers are not using better models. They are changing what the job is.

Leak 3: Unit economics that invert at scale

Pilot costs are small and often subsidized by a vendor or an innovation budget. Production costs include inference at full volume, retrieval infrastructure, evaluation, monitoring, human review of uncertain outputs, security controls and the engineers who keep it all running. McKinsey found that about one in five organizations is already limiting AI use because of operating costs, including token costs.

The test is simple and frequently skipped: fully loaded cost per completed transaction, compared with the value per transaction, at production volume. If the pilot business case cannot show that number, it is not a business case.

Leak 4: Accountability nobody owns

In Moffatt v. Air Canada (2024), the airline's website chatbot gave a customer incorrect information about bereavement fares. Air Canada argued, in effect, that the chatbot was responsible for its own statements. The British Columbia Civil Resolution Tribunal rejected that argument and held the airline liable for negligent misrepresentation.

The amount at stake was small. The principle is not: an AI system acting in your name is you. Initiatives that reach production without a named owner for errors, a defined escalation path and monitoring that catches failures before customers do eventually lose the trust of the business or the regulator, and get switched off.

What the 6% do differently

The McKinsey high-performer data is worth reading carefully, because it points to management choices rather than technology choices.

Three patterns stand out:

1. They set growth objectives, not only efficiency targets. About 80% of all respondents pursue efficiency. High performers mostly add growth or innovation goals, and are 3.3 times as likely to intend fundamental business transformation within three years.
2. They redesign the work. Roughly 75% versus 25%, as above. This is where individual productivity becomes enterprise capacity.
3. They measure and they lead. High performers are twice as likely to report senior leaders visibly committed to AI initiatives, and twice as likely to have defined processes for measuring impact.

None of these require a frontier model. All of them require executive attention that most AI programs never get, because they are run as technology projects.

Replace the pilot with a value case

The practical fix is to stop approving pilots and start approving value cases. A value case is a pilot with a business owner, a baseline and a stop condition attached. It fits on one page:

The numbers in the right-hand column are illustrative, but the discipline is not. If a sponsor cannot fill in the strong-answer column, the initiative is still an experiment. That is fine, but it should be funded and judged as one, and it should not appear in anyone's value forecast.

For a deeper method on choosing which initiatives deserve a value case at all, see how to select enterprise AI use cases.

Three gates, three questions

Once a value case exists, fund it in tranches. Each gate asks one question, and the evidence required becomes harder to fake as investment grows.

The important design choice is that every gate has three legitimate outcomes: scale, reshape or stop. Organizations that only ever say "continue" at a gate do not have gates. They have status meetings. A healthy portfolio should be stopping or reshaping a visible share of initiatives at each gate, and leadership should treat a well-evidenced stop as a good result, because it releases capacity for the next candidate.

Count the money honestly

The final leak is in the arithmetic itself. Most AI business cases report paper capacity: hours saved multiplied by a loaded cost rate. Very little of that ever becomes banked value, meaning cost actually avoided, revenue actually earned or risk actually reduced.

Consider an illustrative contact centre with 400 agents, where an AI assistant reduces average handling time by 15%.

On paper that is 60 full-time equivalents of capacity. In practice, only three-quarters of eligible interactions use the tool. Some agent time shifts into reviewing and correcting AI drafts. And capacity only turns into money if the workforce plan changes: hiring is paused, attrition is not backfilled, or released time is deliberately pointed at revenue-generating work. In this example, 25 FTE of value is actually banked, about 40% of the headline.

That is still a good result. The point is to forecast it, plan the workforce decision explicitly and have finance agree to the mechanism before the build, not argue about it afterwards.

What to do in the next 90 days

1. Inventory the portfolio and classify every initiative honestly as experiment, value case or production capability. Most organizations find that a large share of what they report as "AI in production" has no measured business outcome.
2. Require a one-page value case for anything seeking production funding, using the seven elements above.
3. Name a business owner, not a technology owner, for every production AI capability, including explicit accountability for wrong outputs.
4. Introduce the three gates with real stop authority, and publish the number of initiatives stopped or reshaped each quarter.
5. Pick one high-volume workflow and redesign it end to end around AI, rather than adding AI to the existing process. Use it to prove the redesign pattern internally.
6. Agree the banking mechanism with finance for every capacity-based business case before build starts.
7. Instrument unit economics in production: cost per completed transaction, human review rate and error rate, reported monthly next to the business outcome.

Initiatives that pass the gates need a durable home: clear decision rights, shared platforms and lifecycle governance. That is the subject of the enterprise AI operating model.

Frequently asked questions

Why do most enterprise AI pilots fail to deliver value?

Pilots usually test whether a model can perform a task, which is now rarely in doubt. They seldom test long-tail performance, real adoption, unit economics at scale, or who owns errors. Value is lost at those points, not at the model.

What share of companies see financial returns from AI?

In McKinsey's 2026 global survey, 37% of respondents attributed any EBIT impact to AI, unchanged from 2025. Only about 6% attributed 5% or more of EBIT to AI.

What do AI high performers do differently?

They pursue growth as well as efficiency, fundamentally redesign workflows rather than layering AI onto existing processes, and have visible senior leadership commitment and defined processes for measuring impact.

How should an AI pilot be evaluated?

Against a pre-agreed value case: a measured baseline, a specific mechanism by which value reaches the P&L, an adoption threshold, fully loaded unit economics, a named owner for errors, and explicit kill criteria.

Sources and further reading

• The state of AI in 2026: On the road to ROI, McKinsey & Company, 25 August 2026. Survey of 1,719 participants, fielded May–June 2026.
• Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, Gartner, 29 July 2024.
• 42% of companies abandoned their AI initiatives in 2025, Cybersecurity Dive, 18 March 2025, reporting S&P Global Market Intelligence survey data.
• Klarna AI assistant handles two-thirds of customer service chats in its first month, Klarna press release, 27 February 2024.
• Klarna changes its AI tune and again recruits humans for customer service, CX Dive, 2025.
• Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, 14 February 2024.