Our AI strategy was written eighteen months ago. What is now wrong with it?

An AI strategy written eighteen months ago is probably not wrong about the tools. The tools have changed, but a strategy that depends on which model it names was never a strategy. What has changed is the set of decisions the document could not have been asked to make, because the questions did not exist in a usable form in early 2025. Five of them do now. Each is a decision rather than an upgrade, and none of them will be found in the vendor section.

One: the systems now act#

In early 2025 the strategy assumed a machine that answers and a person who acts on the answer. Agents invert that. Gartner's June 2025 forecast that over 40 per cent of agentic AI projects will be cancelled by 2027 names inadequate risk controls as one of three reasons, and its estimate that about 130 of the thousands of vendors claiming agentic products are real tells you most of what is in the market is a chatbot with a new name. The 33-author preprint of September 2026 that set out five levels of self-improvement places industrial systems at levels one to four, with 'experts retain control over consequential corrections and deployment'. The strategy needs a sentence it did not have: which actions a system may take without a person, who can stop it, and how the organisation knows it stopped.

Two: the reversals have happened#

The 2024 strategy had Klarna's assistant doing the work of 700 agents as a case study. In May 2025 Klarna's chief executive said cost had been 'a too predominant evaluation factor' and that the result was lower quality, and the company recruited people again. Duolingo declared AI-first in April 2025 and its chief executive walked it back in May. IBM cut a few hundred roles and, by its own account, reinvested and grew. The cases are operator accounts without figures, and their lesson is not that automation fails. It is that a strategy which does not say where a human must remain, and what the freed money is for, finds those answers in public. The record is at what happened to the companies that cut staff for AI.

Three: the regulation is in force#

Article 4 of the EU AI Act, on AI literacy, has applied since February 2025. Article 14, on human oversight of high-risk systems, has applied since August 2026, and Annex III lists what counts as high-risk, including recruitment, promotion, termination and access to education. A strategy written before either applied may name them as coming. It cannot have decided what meaningful oversight looks like in this organisation, who exercises it, or whether they can. The test for that is at meaningful human oversight.

Four: the shadow use is ahead of the plan#

MIT NANDA's 2025 report found workers at over 90 per cent of surveyed companies using personal AI tools for work while only 40 per cent of companies had bought an official one. The figure rests on a small sample and the report says so, but the direction is not in doubt. The strategy from eighteen months ago described a programme the organisation would run. The organisation's people have been running their own, and the strategy has no view on it. The question that follows is at what to do when people work around the AI policy.

Five: the builders have asked for oversight of themselves#

In September 2026 the chief executive of Anthropic asked for outside evaluators embedded inside AI companies with employee-level access, and Microsoft published rules under which its models 'will never resist human interruption, correction, or shutdown'. Neither says where those humans come from or what happens if the ones nominally in charge have stopped practising the judgement the role needs. A strategy written in early 2025 treated the vendor as a settled quantity. The vendors are now saying, in public, that they are not. That does not change what an organisation buys. It changes how much of its own judgement it can afford to hand over.

What to check, in order#

Open the old document and look for one list: the organisation's consequential decisions, each marked inform, recommend, execute or never, with an owner against each, the work the organisation keeps unaided, and how often human reviewers disagree with the machine. If it is there, the strategy has a rules layer and the five changes above are revisions to it. If it is not, the document was a roadmap, and the roadmap was written before the rules. The rewrite does not start with the vendors. It starts with the list, which takes a day with the executive team, and the roadmap is then revised against it, which is the order it should have had the first time. The rules are at rules before tools, the argument for who writes them at AI leadership.

What nobody has measured#

There is no study of AI strategies by vintage and outcome. McKinsey's 2025 survey found the redesign of workflows and the chief executive's ownership of the rules to be the attributes most associated with reported profit, and both are things a strategy either contains or does not, but the survey does not date the strategies. The five changes above are documented events; the claim that a strategy silent on them is the worse for it is an argument from their shape.

Key sources

• Gartner (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Graded entry.
• Duan, Y. et al. (2026). The Last AI Built by Humans. arXiv 2609.11873. Graded entry.
• Siemiatkowski, S. (2025), Klarna, reported by CX Dive from a Bloomberg interview. Graded entry.
• Regulation (EU) 2024/1689, Annex III. Graded entry.
• Challapally, A. et al. (2025). The GenAI Divide. MIT NANDA. Graded entry.
• Amodei, D. (2026). We Must Pace the Frontier. Graded entry.
• Microsoft AI (2026). Humanist AI in practice: a public consultation on our Code of Conduct for MAI models. Graded entry.
• Singla, A. et al. (2025). The state of AI. McKinsey. Graded entry.

Related SuperSkills research#

The rules layer is at rules before tools and the definition at AI leadership. On agents specifically, AI agents and human judgement. On when to reverse a deployment, deployment is not a ratchet. On the September 2026 frontier week, neither hype nor doom.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. He has run, grown, bought and advised businesses with AI in them. Findings are attributed to the studies and statements that produced them and kept separate from the interpretation. This is a living reference, reviewed and updated as significant new evidence appears.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.