Case Study: What Changed When AI Adoption Got a Named Owner

Ask who owns AI adoption in your organisation. If the answer contains the word "everyone", the real answer is nobody¹ — and everything currently working is one resignation away from stopping.1A composite case, anonymised from several Fellow engagements. Numbers are representative of the pattern, not one client's audited results.

The situation

A retail group of about 2,000 people had enviable AI momentum, almost all of it traceable to one senior analyst who had made adoption a personal mission: she maintained the prompt library, ran the informal training, and fielded everyone's questions. Then she resigned. Within a month the library was stale, the sessions stopped, and usage in her division dropped by roughly a third^◦.By the numbers≈⅓drop in division AI usage within one month of the unofficial champion leaving

The mechanics of the collapse were mundane. The prompt library was a personal document shared with about forty colleagues; when her account was deprovisioned, two of the three teams that used it most could not work out where it had gone. Her successor inherited a job description that did not mention AI anywhere, because the work had never been written into one.

Nobody had asked her for a handover, either — not out of carelessness, but because none of what she did appeared on an org chart, in a budget line, or on a scorecard. Her own manager described it in the exit conversation as "something she did on the side". That description was completely accurate, and it was the entire problem.

The readiness assessment, run during the aftermath, made the structural gap explicit: real usage across the scope, and no accountable owner anywhere in it.

Ownership is the variable that shows up in the data

The pattern is visible in the survey data, not just in anecdotes. McKinsey's March 2025 State of AI survey of 1,491 organisations found that CEO oversight of AI governance was the element most correlated with bottom-line impact at large companies — and that only 28% of organisations had it^1◦. The caveat has to travel with the number: this is a correlation with self-reported EBIT impact, not a demonstration of causation, and senior oversight may partly be a marker of organisations that were already serious. But the direction is hard to ignore. Seniority of ownership tracks with outcomes, and most organisations have not assigned any.SourceMcKinsey, "The State of AI", March 2025By the numbers28%large-company respondents with CEO-level oversight of AI governance, McKinsey State of AI, March 2025

The same survey points at why. Of roughly twenty-five attributes tested, fundamental workflow redesign had the single biggest effect on EBIT impact — yet only around 21% of adopters had redesigned any workflows at all (the figure is approximate). That is the number to sit with when your adoption depends on a champion. Redesigning a workflow means changing what other people are accountable for, in what order they do things, and what "finished" means. A volunteer cannot do that, however talented, because it is not a skill problem — it is an authority problem^◦.RelatedWhy the owner needs a baselined number to own

BCG frames the same imbalance as a rule of thumb it calls 10-20-70: about 10% of the value from AI comes from algorithms, 20% from technology and data, and 70% from people, process and culture². Treat it as a consulting heuristic drawn from client experience rather than a measured statistic — but note what it implies. In most organisations the 10 and the 20 have a named owner with a budget, and the 70 does not have anybody at all.SourceBCG's 10-20-70 heuristic — a consulting rule of thumb, not a measured statistic

The consequence shows up at the front line. BCG's AI at Work 2025 study, covering more than 10,600 employees across eleven countries, found regular generative-AI use among frontline staff at 51%, against more than 75% for leaders and managers³. A year later the frontline had closed most of that gap, reaching 74% in the 2026 edition — which moved the problem rather than solving it. In that same 2026 survey a clear AI strategy was worth 25 percentage points of business impact, while better tools alone were worth 5^4◦. Access stopped being the thing that separated organisations. Ownership of the strategy became it^◦. Only about a third of employees said they had been properly trained, and more than half said they would find their own tools if not given what they needed. Enthusiasm at the top plus improvisation at the bottom is exactly the shape of champion-driven adoption, at national scale.SourceBCG, "AI at Work 2025"SourceBCG, "AI at Work 2026: Strategy Matters More Than Tools", June 2026 (via BCG press release)By the numbers25pp vs 5ppbusiness impact from a clear AI strategy versus from better tools alone, BCG AI at Work 2026By the numbers51%frontline employees using generative AI regularly versus more than 75% of leaders and managers, BCG AI at Work 2025

The most instructive public example is Klarna, and it only teaches the right lesson if both halves are told. In February 2024 the company announced that its AI assistant had handled two-thirds of customer-service chats in its first month — 2.3 million conversations, described as the work of about 700 full-time agents — with resolution time falling from 11 minutes to under 2, repeat inquiries down 25%, and a projected $40 million profit improvement for the year. Those are first-party, unaudited figures⁵.SourceKlarna press release, February 2024

Then, in May 2025, Klarna began recruiting human agents again⁶. CEO Sebastian Siemiatkowski told Bloomberg that cost "unfortunately seems to have been a too predominant evaluation factor", and that "what you end up having is lower quality", adding that investing in the quality of human support was the way forward for the company. Read the arc carefully: it was not a retreat from AI². It is a much sharper lesson than that. A single number — cost per resolution — was owned and optimised, and the number that was not owned by anyone accountable, quality, quietly degraded until it had to be bought back. Ownership is not only about who drives the metric up. It is about who is accountable for what the metric does not capture.SourceKlarna's 2025 partial reversal, rehiring human agents — reported by Bloomberg; no public release2Klarna kept the AI assistant and moved to a hybrid model with human agents for customers who want them; no employees were laid off in the original 2024 shift, which relied on reducing outsourced support.

Why "organic" adoption stalls

Champion-driven adoption feels healthy because it is enthusiastic, cheap, and self-starting. It is also structurally fragile^◦. The champion has no mandate, so every improvement is a favour. No budget, so tooling is improvised. No measures, so their impact is invisible until it disappears. And because five teams copy the champion five different ways, practices never converge into anything the organisation owns.DefinitionChampion risk Adoption that depends on a volunteer's personal energy — unbudgeted, unmandated, and gone the day they change roles

The failure is not the champion leaving. It is the organisation having rented its capability from one person's goodwill without noticing.

It also caps how far adoption can go, quite apart from the risk. A champion can teach a tool and share a prompt; those are the parts of the work that require no permission. The step that the McKinsey data associates most strongly with impact — redesigning the workflow itself — is precisely the step a volunteer is not allowed to take. So champion-led adoption reliably produces the same ceiling: individuals working faster inside a process that has not changed at all, which is the same gap MIT's researchers described when they attributed pilot failure to workflow integration rather than model quality⁷.SourceMIT Project NANDA, "The GenAI Divide: State of AI in Business 2025" (via Fortune)

That ceiling explains a result that otherwise looks strange — organisations with high usage and no measurable benefit. Both facts are true at once, and neither is a contradiction. The tools were adopted. The work was not redesigned. Nobody had the standing to do it.

What we did

The fix was deliberately small — the opposite of a transformation office:

1. One named owner. A manager — not the busiest executive, and not a committee — with written decision rights over tooling requests, workflow priorities, and training time^◦.DefinitionDecision rights The specific choices an owner may make alone, written down — which tools get approved, which workflows come first, whose time may be spent on training — as distinct from being "responsible" for an outcome without the authority to change anything
2. Three workflows, one measure each. Store-report drafting, merchandising copy, and supplier-email triage: each with a baseline, a target, and nothing else on the owner's scorecard until those moved.
3. A thirty-minute monthly review with the COO in the room. Not a steering committee — a working session with three numbers on one page. Leadership attendance, more than any mandate document, is what told the organisation this was real^◦.FigureOwnership becomes real the first time leadership sits in the review and asks about the number

The former champion's successors were folded in as workflow leads — the enthusiasm was kept, but it now reported into a structure instead of substituting for one.

Two objections arrived early, and both were reasonable. Store managers heard "named owner" as "a new head-office function that will start asking us for reports", which in that group had a history. The remit was rewritten to three sentences, one of which listed what the owner would not do: no new reporting, no tool audits, no approval step in anyone's existing process. The second objection came from the owner's own manager, who was being asked to give up roughly a day a week of her time. That was resolved by removing an existing reporting duty rather than stacking the new role on top. Nobody has spare capacity, and ownership added to a full job is decoration.

It nearly went wrong twice. The first scorecard draft had six workflows on it, because every function wanted theirs represented; it was cut back by asking which three the COO would actually notice moving. And eleven people accepted the invitation to the first monthly review, which would have turned a working session into a steering committee within two meetings. It went back to four attendees. The COO's presence was the point; the audience was not.

The cost was about four person-weeks across the first quarter: a two-day scoping pass, roughly an afternoon per workflow to set baselines, the owner at about a day a week, then thirty minutes of review plus an hour of preparation each month. There was no new headcount and, deliberately, no new tooling spend in the first quarter — partly to prove the change was structural rather than purchased, and partly because a request for budget would have turned a two-week decision into a two-month one.

What changed

Usage recovered past its previous peak within two months, and — the part that matters — it survived the next departure without a dip, because assets and accountability no longer lived in one head. At re-assessment, adoption ownership moved from the bottom of the profile to its strongest indicator^◦.By the numbers3workflows with a named owner and a monthly-reviewed measure — the entire initial governance structure, on purpose

The durability came from unglamorous things. The prompt library moved to a shared location with a named maintainer and a quarterly review date. The three workflows appeared in onboarding for new starters, which meant the practice was transmitted by the joining process rather than by whoever happened to be enthusiastic. And the monthly page created a written record, so the second owner started from where the first had got to instead of from scratch.

It is worth being honest about what did not change. Total tool spend was flat. Adoption in teams outside the three workflows stayed roughly where it was for two quarters — the structure did not lift the whole organisation, and was never designed to. What it did was make three things reliable and repeatable, which turned out to be the credible basis for asking for a fourth.

What we would do differently

Three things, in ascending order of discomfort.

We would name the owner before the champion leaves rather than after. Everything in this rebuild was available at any point in the preceding year; it took a resignation to make it urgent. The two months of recovery were pure avoidable cost, and the diagnostic question is embarrassingly simple: if this person left on Friday, what stops?

We under-weighted quality, and the Klarna arc is why we now correct for it. Each of the three workflows launched with an efficiency measure, and efficiency measures move in the direction you push them. We now pair every efficiency number with one quality guard — a rework rate, an error count, a spot-check by someone who did not produce the work — precisely because the thing nobody owns is the thing that quietly degrades.

The hardest one: a named owner can become the new single point of failure with a better job title. If the remit, the decision rights, the library and the measures live in that person's head and calendar, the organisation has swapped one dependency for another and given itself a false sense of resolution. What actually prevents it is dull and must be built from day one — the remit written down, a named deputy who attends the review, and assets stored where someone else's login can reach them. Structure is not the same thing as a person with a title, and it is easy to mistake the second for the first.

What to steal

1. This quarter: name one adoption owner — a person with time and decision rights, not a working group.
2. Hand them exactly three workflows with one measure each. Refuse the temptation of a ten-item scope.
3. Put a monthly thirty-minute review on the calendar of the most senior person who will reliably attend. Attendance is the mandate.
4. Pair every efficiency measure with one quality guard, so the unowned number cannot degrade unnoticed.
5. Convert your champions into workflow leads inside the structure — celebrate them, and stop depending on them.

If your readiness report flagged real usage with nobody accountable, treat it as a countdown, not a description — the fix costs one job-description paragraph and a monthly half hour^◦.RelatedThe next step: turning owned workflows into compounding practice