AI · MBA

An orange fraction mark over a sparse grid, hinting at survivors versus the missing rest

What the winning 28% do

Gartner asked the winners why they won, and the winners answered: we plugged AI into the workflows and systems we already had, and the business stood fully behind us. It sounds like a recipe. It is a list compiled by the ones who came back from the sortie. The decision that hangs on it comes round every quarter, when you pick the one use case to launch first, and you make that call on data with every casualty cut out of it.

What that survey actually counted

Start with the denominator, because everything else rests on it. Gartner's press release of 7 April 2026 says that 28% of AI use cases in IT infrastructure and operations (I&O) are fully successful and meet return-on-investment expectations. One more figure is stated outright there. 20% of use cases fail flatly. The unit is a single use case in a single area of IT. Not a company and not a project. The remaining 52% or so sit in an intermediate state that the source itself never names, so any sentence of the form "72% went down" adds content to this survey that the survey does not contain. I took that figure apart separately and I will not repeat that work here.

The same survey carries a second pair of numbers that write-ups usually lose. 77% of the organisations surveyed delivered at least one successful use case, and 57% reported at least one failure. They add up to more than a hundred, so the two sets have to overlap heavily. A company with nine failed deployments and one that worked lands inside that 77% on exactly the same footing as a company that hit every time.

Name what kind of number this is before you build anything on it. "Meets ROI expectations" is a respondent's declaration about their own result, not an audited account, so the 28% measures how satisfied I&O leaders are with what they themselves commissioned. The survey covered 782 respondents in November and December 2025. One intermediary gives 783; I am not rounding that discrepancy away into "roughly eight hundred", because a difference of one respondent says something about the care taken in the relaying. Gartner's domain returns a 403 to automated retrieval and the sampling methodology stayed inside the paid report. I confirm the content of the release against two independent trade write-ups and give it here in that form.

One sentence in that release is what interests me. The leaders inside the 77% who delivered at least one successful use case attribute the success above all to embedding AI into existing workflows and systems, and to full backing from business leadership. That is a post hoc attribution made by winners about themselves. Nobody checked how many of the casualties had those same two traits.

Set it against the other side of the same survey, because both lists are of one species. The losers gave reasons too. Most often they pointed to unrealistic expectations of the technology. Skills gaps, and data quality and availability, drew 38% of mentions each. Gartner's Melanie Freeze attributes the high failure rate to precisely those unrealistic expectations of the technology: leaders assumed that AI would immediately automate complex tasks or repair long-standing operational problems. Winners explain themselves with embedding and a sponsor, losers with expectations and data. Both lists are self-report.

One more warning. Gartner has an earlier and smaller wave inside the same research programme: 253 respondents, May to July 2025, four countries. That wave speaks to something else — that 54% of I&O leaders deploy AI mainly to cut costs, for instance. Lining it up against the survey of 782 as evolution over time would be an abuse, because it is a different sample, a different window and a different set of questions.

On top of that comes a distribution that will return at the end of this piece. 53% of the reported successes cluster in one place: generative AI in IT service management (ITSM). One niche, more than half the wins.

Anatomy of a survivor

Let us pull out of this survey and its neighbours five traits that describe a use case that worked. Before I list them, a label, because otherwise this essay does the very thing it is about to accuse others of. These five are not a model from the literature. They have no author, no year, and nobody has tested them. This is my working construct, assembled out of other people's findings, and you should treat it as a hypothesis to be knocked down.

The first trait is narrow scope: the use case does one thing, and that thing can be described in a single sentence with no "and" in it. The second is an owner with their own P&L, meaning a person whose numbers get worse if this does not work. The third is embedding into an existing step rather than a new slot in the user's calendar. The fourth is a measurable before and after, where the number existed before launch and somebody wrote it down together with its definition. The fifth is reversibility: you can switch it off within a week and go back to the old step without explaining a failure to anyone.

The fourth and the fifth do not come from Gartner's survey. I am adding them myself, because without them the first three can neither be checked nor undone, and the whole list then reduces to a story about somebody else's good fortune. Reversibility has one property that none of the other four traits has. You control it entirely yourself.

Behind the first three stands an observation from an independent study. The MIT Project NANDA report laid deployments out on a matrix of scope against execution complexity. Narrow and simple produced quick wins, such as expense categorisation or contract review. Broad and complex ended in failure, the textbook case being full orchestration of a procurement process. Document handling and code generation for repetitive tasks won. So did summarising and routing voice calls. What fell over were deployments that required complex internal logic, or optimisation against a firm's own heuristics.

Rigour about that source is a separate matter. It is a preliminary report, and its survey component is 153 responses collected at four industry conferences, plus 52 interviews and a review of more than three hundred publicly documented initiatives. The authors write themselves that the figures are directionally accurate on the basis of individual interviews. The definition of success there is soft: a deployment counts as successful when users or executives noticed a clear and durable effect on productivity or on the result. The same document gives two different values for the same budget allocation on facing pages, once around 70%, once around 50%. I take the direction from it gladly. I stake nothing on its numbers.

The plane comes home holed where a hit does not kill

Now I have to do to those five what I do to other people's numbers. In 1943 the Statistical Research Group at Columbia University issued eight memoranda by Abraham Wald on estimating an aircraft's vulnerability from the damage to the machines that came back from a sortie. The fifth describes a method for inferring how vulnerable a part is from the distribution of hits among the survivors. The logic runs against intuition. The densely holed areas on returning aircraft are the areas where a hit was not fatal, and the clean areas are the ones that, once hit, stopped the machine coming home. The armour goes where the data shows no holes.

More ornament than fact has grown around this story, and an essay about other people's numbers cannot trade in ornament. Wald did not draw a bomber silhouette speckled with red dots; that graphic is a modern illustration fitted to the anecdote afterwards. The memorandum contains no instruction to armour the engines and no count of crews saved. What it contains is a recursive method for estimating the probability that a hit on a given part brings the plane down, resting on strong assumptions about how hits are distributed. The original publication was classified, the reprint came out in 1980 as CRC 432 at the Center for Naval Analyses, and the canonical scholarly treatment was given by Marc Mangel and Francisco Samaniego in 1984.

Now carry that over to the survey. Gartner questioned the use cases that came back. Embedding into an existing workflow and backing from the board are holes in the fuselage of the survivors. To call them a cause of success, we would have to know how many use cases that were embedded in an existing workflow and covered by a sponsor went down anyway. That number is not in the release. Nobody counted it.

Jan Denrell showed in 2003 why this is not a methodological quibble but a defect in the whole business of learning by observation. The organisations we can see at any given moment are survivors of a selection that cut away a large part of the starting population, and the press and the business books add a filter of their own, because they write about successes. In the sample available for imitation, failures are systematically under-represented. His formal conclusion is sharper than the warning itself. Risky practices, meaning those with high variance in outcomes, can look positively correlated with performance in a sample of survivors while having nothing to do with it in the full population.

The second objection hits causation from another direction. Phil Rosenzweig described the halo effect: judgements about an organisation's attributes are contaminated by knowledge of its financial results, so what we call a cause of success is often a consequence of it. In the version that cuts deepest for this essay it reads like this: embedding into an existing workflow may be a correlate of organisational maturity. A company that has a documented and measured workflow you can plug anything into is already a company that knows how to deploy. An imitator without such a workflow has nowhere to plug in, and copies the form without the condition that was holding it up. I will say it plainly: I cannot settle this with the data I have, because separating organisational maturity from embedding itself would take a sample of deployments from the same company, described from both sides of the outcome. Nobody has published that sample.

So what survives for my five? The strongest honest version of the thesis is this: embedding into an existing step is the most frequently cited correlate of success in a sample of winners, and it has independent support in the adoption literature. But it has not been shown to be a cause. The rest of this essay works under that caveat. I do not lift it anywhere.

What the habit you demand of the user costs

A correlate with no derivation behind it is a superstition, so let us take this one apart from the ground up. Why does a swapped line inside an existing procedure beat a new and better screen alongside it?

The oldest part of the answer is forty-four years old. Louis Tornatzky and Katherine Klein pooled the results of seventy-five studies of innovation adoption in 1982 and tested which characteristics recur most consistently. Compatibility came out at the top, understood as the fit between an innovation and the adopter's existing values, experience and current needs. Relative advantage and complexity held up beside it. The entire discussion of AI deployment is four decades younger than that finding.

Habit is not a metaphor in this discussion either; it is a construct inside a model. Viswanath Venkatesh, together with James Thong and Xin Xu, added it to technology acceptance theory in 2012 and measured what then happens to the model's explanatory power. Variance explained in behavioural intention rises from 56% to 74%, and in actual use from 40% to 52%. The study ran in two stages, covered 1,512 mobile internet users, and the data on real usage was collected four months after the first survey. One caveat I am not lifting: the model was validated on consumers, so carrying it over to an employee inside a company procedure is my extrapolation.

That makes the implication for a deployment simple. If habit accounts for a good share of whether anyone genuinely uses a tool, then a deployment that demands a new habit starts in deficit and has to buy that habit first.

What does it cost in time? Phillippa Lally and her team asked ninety-six volunteers to repeat a chosen behaviour every day in the same context for twelve weeks, and then fitted an automaticity curve to each of their results. The median came out at 66 days, and the spread ran from 18 to 254. The popular version, "66 days to a habit", loses both the spread and the fact that the median comes from the subset the model actually fitted, roughly half the participants. I take an order of magnitude from this. I do not treat 66 days as a constant. A quarter is the realistic cost of one new reflex in one person.

On the attention side the cost is countable. Rohan Narayana Murty, with two co-authors, observed twenty teams in three Fortune 500 companies for up to five weeks. That covered a hundred and thirty-seven users. It came out at roughly 1,200 toggles between applications a day and about four hours a week on reorientation alone, some 9% of working time. 65% of the toggles came less than eleven seconds after the previous one. The caveat matters more than the number here: this measures the cost of switching context, not the cost of learning a new tool. The bridge between the two is mine, not the authors'.

Put it into one sentence. A new screen demands a habit whose cost is counted in quarters, out of an attention budget that is already spent, at a value felt as a fraction of the value claimed. That last clause is a frame rather than a measurement: John Gourville assembled it in 2006 out of two opposing overvaluations, because the user overvalues what they already have and the maker overvalues their own novelty. A swapped line inside an existing procedure demands nothing from that list. The step is already in the calendar, somebody already performs it, and what changes is the thing they perform it with.

Independent confirmation came from the NANDA sample. Sponsors and users at fifty-two organisations rated how often various barriers block a pilot from scaling, and resistance to adopting a new tool came first. The authors note that the ratings reflect reported frequency, not a measured impact of the barrier. The report also lists what executives actually look for in a vendor, and one item says it outright: minimal disruption to current tools. The quote from the interviews is simpler still: "If it doesn't plug into Salesforce or into our systems, nobody will use it."

One item on that list aims at the same spot from another side. Executives want a vendor with a deep understanding of their own workflow, because the interviews keep returning to a complaint that most vendors do not understand how approvals work in that particular company, or where the data travels. The complaint is about the sign-off chain. Embedding into an existing step is in practice a question about who approves and in what order, so the answer to it is settled in a department that has nothing to do with a language model.

The nominal sponsor and the arithmetic that exposes him

The second trait on the winners' list is harder, because the word "sponsor" has worn out. In project documents it means the person who approved the budget. The operationally useful definition is a different one: the sponsor is whoever's own result depends on whether this use case works. Sometimes those are two different people. Then the project has a signature and no owner.

The arithmetic alone makes the distinction. In July 2014 the Project Management Institute surveyed 897 project management practitioners and 232 executives who had sponsored projects in the preceding two years, at organisations with revenues from a hundred million dollars up. The share of projects with an actively engaged sponsor came to 63% in 2012, and did not move in the following year or the one after that. Three measurements gave the same value, while the rhetoric about the role of sponsorship kept growing over that time.

The two figures beside it are more interesting. Sponsors ran an average of three projects at once and gave an average of thirteen hours a week to each, on top of their normal work. That is thirty-nine hours a week bolted onto a full-time job. Nobody has to tell you that a sponsor is nominal: it is enough to ask how many initiatives are on their list today and when they last looked at the numbers on them. One in three failed projects missed its goals because of a poorly engaged sponsor, and overloaded sponsors showed up almost twice as often in low-performing organisations.

The report is twelve years old and grows out of a self-report survey run by an organisation that sells certificates in project management. I am not offering it as a neutral source. What has value here is the arithmetic of overload itself; the success rate from that report I do not use.

Let me also say where this thesis has no support. In the NANDA sample the absence of an executive sponsor came only fourth of five barriers, behind resistance to a new tool, concerns about the quality of model output, and poor user experience. Anyone selling sponsorship as the first cause of success is doing so against that sample. My thesis stands on a definition and on nothing else. I ask whose result hurts when this use case stops.

There is one more observation there, and it fits the definition better than it fits the ranking. The strongest deployments in that sample were reported by line managers and budget holders rather than by a central AI lab; the people whose process hurts named the problem, tested the tools and ran the rollout, while executive accountability sat separately and higher up. That is a qualitative observation from fifty-two interviews. Nobody has tested it.

I write about this from a seat where it shows up with unusual clarity. In my salaried job I work on risk and resilience. I defend other people's budget submissions far more often than I decide them. From that side of the table a nominal sponsor has one recognisable reflex. He asks about risk only once somebody else has named it in a document. The vantage point is biased in a predictable direction. I say so now, because my entire recommendation grows out of it.

You know a nominal sponsor by three answers. Asked about his own number, he answers with somebody else's number or with the annual target of a whole division. Asked when the first review falls, he points to the calendar of a programme he does not run himself. Asked what he will do if nothing has moved after a quarter, he talks about communication and cultural change.

Ask these five questions before you pick the first use case

So far this has been about what I claim and with what confidence. Now for what to do on Monday morning.

QuestionWhat the answer has to containThe answer that disqualifies
Which existing step does this replace?The name of the step in the procedure and the person who performs it today"It will be a new tool in a new tab"
Whose result gets worse if this does not work?A first and last name, and a specific line in their P&LThe name of a committee, or the annual target of a whole division
How narrow is it really?One sentence with no "and" in itA description of an end-to-end process
What was that number last quarter?A value, a measurement date and a definition on paper"We will start measuring after launch"
How do you switch it off?A procedure for returning to the old step within a weekData migration and decommissioning of the old system

The first question is the hardest to cheat, because it demands a name. A step that nobody performs today is not a step to be replaced; it is a new slot in somebody's day, and it costs what a habit costs. If the answer is "we will improve the whole process", count how many new reflexes you are ordering, and in how many people.

The last question looks like a formality and is not one. A deployment that cannot be undone within a week stops being a test the moment it goes live and turns into a commitment that nobody wants to revoke in public. You buy reversibility before launch or not at all. Write the rollback procedure down before anyone touches the configuration, because after go-live nobody will write it.

Let me run this once on my own patch, so as not to leave it at abstraction. I run a studio of one. I have as many candidates for a first use case as I have repeatable steps. I took one: preparing the descriptions and metadata for every piece I publish. The step has existed for a long time, it has an owner, it takes about forty minutes per piece and it comes round several times a month.

Four of the questions in the table this candidate passed immediately. I am not opening a new tab, because I am swapping the tool inside a step I perform anyway. The scope fits into one sentence. If it does not work, my own numbers get worse and there is nobody to reassign them to. Going back to the old way takes as long as opening an editor. The question that fell over was the one about the baseline, and in the stupidest way available: I never wrote the first measurement down, so a month later all I remembered was that it had gone faster. The second attempt started with a sheet of paper.

Before you take this list as your own, read it once more against my own interest. 53% of the reported successes sit in ITSM, and IT service management is ritualised to the limit: ticket, categorisation, assignment, escalation, closure. There is a ready-made step to replace there and a ready-made counter. So it is possible that the 53% is telling us about something other than embedding. ITSM simply has the best-defined before-and-after measure in all of IT, which makes it the easiest place to prove anything at all. That weakens my derivation, and I am leaving the sentence in the text.

The baseline nobody has

The fourth trait is the easiest to miss and the cheapest to fix. A measurable before and after means that the number existed before launch, that somebody wrote it down, and that its definition lies on paper. It sounds trivial. The scale of its absence is not.

Between 25 June and 29 July 2025 McKinsey surveyed 1,993 respondents from a hundred and five countries, weighting the data by each country's share of global GDP; 38% worked at organisations with revenues above a billion dollars. Fewer than one organisation in five tracks defined metrics for its generative AI solutions. 39% attribute any effect at all on the operating result to AI, and most of them speak of an effect below 5%. More than 80% see no discernible impact at the level of the whole enterprise.

Read the first of those figures again, because it explains the rest. Four organisations in five have nothing to compare against after six months, so the question of whether the deployment paid off has no right answer there and no wrong one. It has a narrated one.

This source needs care in both directions. The claim that tracking metrics has the largest effect on the bottom line of the twelve practices studied is a correlation out of a self-report survey. An organisation that knows how to measure probably knows how to do several other things as well. McKinsey's domain returns a 403 to automated retrieval, so I give the figures as they appear in secondary write-ups.

The procedure itself is cheap and takes one afternoon. You pick a step, not a tool, and you write down four things about it: how long it takes, how many times a month it occurs, how often it needs rework, and who pays for that rework. Add the date of the measurement and a one-sentence definition of each of those numbers, so that six months on nobody moves the boundary of what counts. A tool metric, meaning the number of queries, the seats purchased or the tokens consumed, describes usage and nothing beyond it. The definition on paper matters more than the value itself, because six months on nobody remembers whether "ticket handling time" ran from arrival or from assignment to a human being. The boundary moves on its own.

Without that sheet of paper a pilot has no way of ending, and I have written separately about pilots that never end. A use case with no baseline does not die. It persists.

Embedding buys adoption, redesign buys value

One tension is left, and passing over it would be a cheat. In that McKinsey survey, of the twenty-five attributes tested, the largest effect on an organisation's ability to see the result in its operating performance came from redesigning workflows. That means rebuilding them. Only about 21% of organisations using generative AI have redesigned even part of their workflows, and among the highest performers about 55% have, against about 20% of the rest.

Under that figure sits the same weakness as under the previous ones. It is a correlation out of a self-report survey, computed and weighted by a consultancy that sells process redesign. There was no experiment here.

These two findings stand opposed only in appearance, because they are measured at different levels. Gartner asks whether a single use case hit its objective. McKinsey asks whether it shows up in the operating result of the whole company. Embedding into an existing step buys adoption, meaning it gets somebody to use the thing at all. Redesign buys value, meaning it makes the usage countable higher up than the level of one department. You can set them in order: first you swap the step and buy the habit, and only then, with that habit in hand, do you rebuild the workflow around it. That sequence is my hypothesis. Neither of those studies tested it. Worse: if redesigning the workflow were a precondition of adoption rather than a consequence of it, the order would be reversed and every recommendation in this essay would be sending the reader to the most expensive place available.

The thesis comes out of this narrower. It concerns the first use case and that one phase in which you buy the right to a second. The five traits of a survivor do not say which deployment will earn the most; they say which one has a chance of entering use before anyone gets round to counting the earnings. To that is added the limitation you already know about. It is a list compiled by the ones who came back, so its honest work consists in rejecting candidates. It will not point out the winner.

Take the candidate you have lined up for next quarter and run it through the five questions in the table. If it fails on two, put it down and reach for the next one; that much this list can do. If it passes all five, you still know nothing certain about its chances, because nobody counted the casualties with the same traits. One thing is left that you hold entirely in your own hands, and it is the last question. Can you switch this off within a week and go back to the old step? The rest you are buying on credit.

Sources

  1. primaryAbraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Statistical Research Group, Columbia University, 1943 — eight memoranda; the original publication was classified, reprinted as CRC 432, Center for Naval Analyses, July 1980, held in DTIC as ADA091073.
  2. primaryMarc Mangel, Francisco J. Samaniego, Abraham Wald's Work on Aircraft Survivability, Journal of the American Statistical Association 79(386), 1984, pp. 259–267 — the canonical account of the method and a correction of the popular version of the anecdote.
  3. primaryJan Denrell, Vicarious Learning, Undersampling of Failure, and the Myths of Management, Organization Science 14(3), 2003, pp. 227–243 — a formal model and a theoretical argument, not an empirical study of specific firms.
  4. primaryLouis G. Tornatzky, Katherine J. Klein, Innovation Characteristics and Innovation Adoption-Implementation: A Meta-Analysis of Findings, IEEE Transactions on Engineering Management EM-29(1), 1982, pp. 28–45 — a meta-analysis of 75 studies.
  5. primaryViswanath Venkatesh, James Y. L. Thong, Xin Xu, Consumer Acceptance and Use of Information Technology: Extending the Unified Theory of Acceptance and Use of Technology, MIS Quarterly 36(1), 2012, pp. 157–178 — n=1,512, a consumer context; carrying the habit construct over to an organisational context is this essay's extrapolation.
  6. primaryPhillippa Lally, Cornelia H. M. van Jaarsveld, Henry W. W. Potts, Jane Wardle, How are habits formed: Modelling habit formation in the real world, European Journal of Social Psychology 40(6), 2010, pp. 998–1009 — n=96; the median of 66 days comes from the subset to which the curve was fitted.
  7. primaryGartner, press release Gartner Says AI Projects in I&O Stall Ahead of Meaningful ROI Returns, 7 April 2026 — a survey of 782 I&O leaders, November–December 2025; the sampling methodology sits in the paid report, the domain returns a 403 to automated retrieval, the content confirmed in two independent trade write-ups (CIO.com, National Technology), one intermediary gives 783 respondents.
  8. primaryAditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025 — a preliminary report; 52 interviews, 153 survey responses collected at four conferences, a review of more than 300 publicly documented initiatives; the document gives two different values for the same budget allocation on facing pages.
  9. primaryProject Management Institute, Pulse of the Profession In-Depth Report: Executive Sponsor Engagement, October 2014 — a self-report survey run by an organisation that sells certificates in project management; n=897 practitioners and 232 executives.
  10. primaryMcKinsey & Company / QuantumBlack, The State of AI: Global Survey, 2025 — n=1,993 from 105 countries, in the field 25 June–29 July 2025, data weighted by each country's share of global GDP; the domain returns a 403, figures given from secondary write-ups.
  11. secondaryRohan Narayana Murty, Sandeep Dadlani, Rajath B. Das, How Much Time and Energy Do We Waste Toggling Between Applications?, Harvard Business Review, August 2022 — a management publication, methodology described briefly and behind a paywall; the study measures the cost of switching context, not the cost of adopting a tool.
  12. secondaryPhil Rosenzweig, The Halo Effect… and the Eight Other Business Delusions That Deceive Managers, Free Press, 2007 — a methodological critique of the literature on winners; cited as an argument, not as data.
  13. secondaryJohn T. Gourville, Eager Sellers and Stony Buyers: Understanding the Psychology of New-Product Adoption, Harvard Business Review, June 2006 — a management article, not a study; the gap between the value seen by the maker and the value seen by the user is a conceptual derivation from two roughly threefold multipliers, not a measured coefficient.