03What is actually working

Headlines from four years in

October 2026. Four years since the public arrival of ChatGPT. Long enough for most organisations that would try generative AI to have tried something, and long enough for the results to have stopped being speculation. A sample of what the public record shows:

The same technology, available at roughly the same cost to all comers, has produced a Nobel Prize, a class-action lawsuit, a dramatic public reversal, and the overwhelming majority of deployments producing nothing. The explanation cannot be the technology. If it were, the outcomes would cluster. They do not.

Four years in, three outcomes

The useful way to read the past four years is not as a story of adoption curves or model generations. It is as a triage of outcomes. Of every ten organisations that invested seriously in generative AI since late 2022, roughly nine-and-a-half produced no measurable return. Of the remainder, a small group produced genuine, repeatable value. And across the full population, a visible minority produced active damage: regulatory exposure, reputational harm, operational errors with public consequences. The three buckets are not neatly separated at the edges, but as a first-order description of the terrain they are accurate.

This chapter walks the three buckets briefly, because the specifics are documented well elsewhere, and then spends its weight on the question that actually matters: given that the technology is the same across all three buckets, what explains the split? The answer, when it comes, is not an AI answer.

Projects that led nowhere

The MIT NANDA study from Chapter 02 remains the most defensible estimate of this bucket's size. Roughly 95% of enterprise generative AI deployments, over three hundred surveyed with a defensible methodology, produced no measurable improvement in productivity, revenue or cost that survived scrutiny. The deployments did not necessarily fail loudly. Most absorbed quietly into IT budgets, were quietly re-scoped as "learning investments," or were formally succeeded by the next pilot without anyone writing up the first.

A meaningful fraction of the nothing-bucket deserves a specific caveat. A project started on a 2022-era model, trained on 2022-era data, with 2022-era prompting techniques, looks poor against 2026 standards for the simple reason that the floor has risen. Some of what is counted as "AI failure" in the current surveys is actually AI deployed against the wrong tool, in good faith, before the tool was ready. This is a real explanation, and it is why some organisations are quietly re-running pilots that failed in 2023 and getting different results now. But it is also a convenient explanation, and it is important not to lean on it harder than the evidence supports. A project that failed because the problem was wrong, the leader was wrong, or the measurement was wrong will fail on every model, forever. Only a specific subset of failed projects deserves the generational-reset defence. The rest need a harder diagnosis.

Projects that caused damage

The damage bucket is smaller than the nothing bucket but materially more expensive per occurrence. The pattern repeats across four years and across continents: an organisation deployed generative AI in a decision or content-production context, failed to maintain honest human review, and discovered the cost of that omission in public.

Three illustrative cases, all documented and all from the past three years:

Mobley v. Workday. Filed in 2023, certified as a collective action by a federal court in Northern California in May 2025. The claim is that Workday's AI-driven candidate-screening tooling, used by thousands of employers to filter applicants, systematically discriminated on the basis of age, race and disability. Beyond the direct legal exposure, the case tested whether the AI-vendor-licensor could be held jointly liable with the employer. The court's willingness to certify the collective action was the signal that mattered: AI deployed into regulated decision contexts does not get a technology defence.

Deloitte Australia. In 2025, Deloitte agreed to refund part of a A$440,000 contract to the Australian government after a report delivered under the engagement was found to contain fabricated academic citations and references to papers that did not exist. The content had been produced with generative AI assistance; the human review that should have caught the fabrications did not catch them. For a professional-services firm whose core offering is credibility, the reputational cost of the story exceeded the dollar refund by a wide margin.

Klarna's public reversal. In February 2024, Klarna's CEO publicly announced that an AI customer-service assistant was handling the workload previously requiring seven hundred human agents. By May 2025, the company began rehiring human agents and publicly acknowledged that the fully-automated service had delivered worse outcomes than the human-staffed service it replaced. Klarna did not fail at AI. It succeeded at telling itself a story about AI that the operational data did not support, and had to walk the story back under scrutiny. This bucket contains more Klarnas than it contains Deloittes.

What the damage bucket shares is not a technology failure. In each case the AI worked within its narrow technical specification. What failed was the operating model around the AI: insufficient human verification, mismatched claims against measured outcomes, deployment into decisions that carried legal weight before the governance to carry that weight was in place. The damage was organisational, not algorithmic.

Projects that worked

The success bucket is small as a percentage of attempts, and large in absolute terms because the attempt base is large. Documented wins span sectors and geographies. A representative set:

AlphaFold and the structural biology revolution. Demis Hassabis and John Jumper of Google DeepMind shared the 2024 Nobel Prize in Chemistry for AlphaFold2, which predicted the three-dimensional structure of essentially every known protein and made the dataset public. The downstream effect on structural biology, drug discovery, and life sciences more broadly is measurable in thousands of research programmes and multiple spinouts. This is the clearest example of generative AI producing a scientifically significant result at scale.

The MASAI trial. A randomised controlled trial run in Sweden through the Lund University breast-cancer screening programme, published in The Lancet Oncology in 2023, showed that AI-supported mammogram reading detected roughly 20% more cancers than standard screening, with no increase in false positives, and reduced radiologist workload by approximately 44%. This remains the strongest clinical-trial evidence for AI in a diagnostic setting.

Enterprise code assistance at scale. In 2024, Goldman Sachs rolled out an AI coding assistant to more than 9,000 developers, about 80% of its engineers, and its CIO reported significant efficiency gains. Several other large financial institutions have reported similar results, and the pattern now generalises to any sufficiently large engineering organisation willing to measure the before and after honestly.

European enterprise deployments. Sanofi, BNP Paribas, Air France-KLM and others have publicly discussed internal rollouts of AI assistants for drafting, research, and knowledge retrieval, with measured productivity gains for specific knowledge-worker roles. These are not Nobel Prizes. They are the quiet, well-governed, measurable wins that compound over time and that the surveys often miss because the organisations do not publicise them.

Narrow operational AI. Across industrial manufacturing, retail logistics, and customer-operations triage, specialised AI deployments (often not generative, often not large) continue to produce the steadiest returns. These are not the headline cases. They are the base of the pyramid on which the successful organisations build.

Nothing in this list is magic. None of the successful organisations had a secret model or a secret vendor. What they had was a particular way of running the project. That is what the next section is about.

How to explain the split

The three buckets exist. The technology is the same across all three. Four variables, examined across the documented cases, explain almost all of the variation. None of the four is about AI.

1. Variable 01 How the problem was selected Successful projects started from a narrow, well-defined business problem with a measurable baseline. Failed projects started from a technology ("let's apply AI") and looked for a problem it could solve. Problem-first succeeds; technology-first fails.
2. Variable 02 Who led the project Successful projects were led by domain experts, with AI specialists as tool-holders. Failed projects were led by AI specialists, who built impressive systems that solved the wrong problem beautifully. The lead sets the problem; the specialist solves it.
3. Variable 03 How honestly outcomes were measured Successful projects defined "working" before deployment and measured against the definition after. Failed projects deployed first and found metrics later. Vanity metrics ("users find the tool helpful") are the signature of this failure mode.
4. Variable 04 Whether the organisation would kill it Successful organisations stopped projects that were not returning. Failed organisations defended the sunk cost and kept paying. Cultures that cannot kill their own projects will always have more failures than successes.

These four are not independent. Projects that start technology-first also tend to be led by technologists, measured by vanity metrics, and defended against termination. Projects that start problem-first tend to attract domain-expert leadership, honest measurement, and the willingness to stop when the data says so. The four variables cluster, and together they describe the operational discipline that distinguishes the 5% from the 95%.

None of this is specific to AI. The same four variables explain the ERP migrations that worked and the ones that did not. They explain the cloud migrations that captured the promised savings and the ones that just moved on-premises problems to a different physical location. They explain the digital-transformation programmes that produced measurable change and the ones that produced only PowerPoint. Generative AI did not invent a new failure mode. It exposed existing organisational failure modes with unusual clarity, and at unusual speed.

What this means

Three conclusions follow from the four-years-in evidence. Each does work for a C-level reader deciding how to engage with generative AI from here.

The success profile is reproducible. Nothing in the recipe requires exotic talent, exceptional capital, or lucky timing. It requires the operational discipline the organisation already applies to other serious technology initiatives. A well-run technology function can execute it. This is the central beacon finding of this chapter: the opportunity is accessible, not a lottery.

The dominant failure mode is organisational, not technical. When an AI project fails, the first instinct is often to blame the model, the vendor, or the team. Four years of outcomes say the first thing to examine is the operating model. Was the problem narrow and defined? Did domain experts lead? Was the baseline measured honestly? Could the project have been killed? Four honest answers to those four questions will diagnose most failures faster than any technical review.

"We tried AI and it did not work" is almost never a conclusion about AI. It is a conclusion about how the organisation runs technology initiatives. The useful follow-up question is not "which model should we try next" but "which of the four variables did we get wrong, and are we prepared to get them right this time". If the honest answer is no, the right decision is to stop, not to retry on a newer model.

The operational discipline that produces the 5% outcome, named in six points, is nothing exotic. Any executive who has led a cloud or ERP migration recognises it immediately:

1. 01A narrow, measurable business case. One workflow, one team, one outcome you can measure in hours, euros, or risk reduction.
2. 02A quantified baseline before deployment. You know what "before" looks like precisely, so you can tell what "after" is worth.
3. 03Domain experts in the lead. The person driving is a doctor, a lawyer, an engineer, an operator. AI is a tool in their hands.
4. 04Humans in the loop by design. Not a fallback for immature technology. The operating shape that makes AI safe to deploy in anything that matters.
5. 05Measured iteration with a kill criterion. Instrumented from day one. What works expands, what does not is stopped. Both decisions require evidence.
6. 06Governance from day one. Data ownership, risk policy, compliance mapping, audit trail, human accountability, shadow-IT controls. Not after. Before.

These six points are the exit from Chapter 03 and the entrance to Chapter 04. The rest of the article is about how to operationalise them.