Operational AI · Leadership and accountability
My leadership team wants to ignore AI. How do I get them involved?
Ignoring AI is not the same as controlling it. Leaders do not need to become AI enthusiasts, but they do need to know where it influences work, who checks it and who remains accountable.
On 7 May 2026, I spoke to MBA candidates at the University of Sussex about Leading when AI is in the room. A question after the talk captured a problem many organisations face: “My leadership team wants to ignore AI. How do I get them involved?”
My starting point would not be a presentation about how powerful AI has become. It would be a conversation about the work leaders already own and the evidence they need to govern it.
Non-engagement does not equal non-use.
1. Make existing AI use visible
People may already be using AI to draft reports, summarise documents, prepare briefings or test ideas, with or without a formal programme. Some use may be low risk. Other use may expose sensitive material, omit critical context or influence consequential decisions.
In a US survey, Bick, Blandin and Deming reported that by late 2024 nearly 40% of people aged 18 to 64 had used generative AI, while 23% of employed respondents had used it for work during the preceding week. These are US population survey findings, not estimates of use within any particular organisation. They nevertheless challenge the assumption that an organisation can simply opt out.
The practical questions are straightforward: Which tools are people using? What information is entered? What outputs are retained? Which recommendations or decisions might be influenced? And who can see that activity?
Where use is hidden, policy alone provides little assurance. The first task is to establish a truthful picture without treating every employee experiment as misconduct.
2. Replace the adoption debate with operational responsibility
A reluctant leadership team should not be asked to decide whether it is for or against AI in general. It should be asked what controls a specific task requires. A crisis briefing, contractor risk assessment, board paper and routine email do not have the same consequences if information is wrong or missing.
| Unhelpful debate | Useful leadership question |
|---|---|
| Should we adopt AI? | Where is AI already being used, and for what? |
| Should we ban AI? | Which uses, data and decisions need boundaries? |
| Can we trust AI? | What evidence and checks are required for this task? |
| Will AI save time? | What outcome would improve, and how will we measure it? |
For an incident report, for example, the person reporting the event should remain responsible for their account, the reviewer for checking completeness and discrepancies, and the authorised manager for decisions or escalation. AI may help surface missing information or retrieve procedures, but it should not quietly become the source of an unverified statement.
3. Protect human judgement, not just model accuracy
Hallucinations matter. AI can produce plausible but incorrect claims, omit qualifying detail or offer conclusions that the evidence does not support. But the deeper leadership problem is what happens when people begin treating confident output as authority.
Steven D. Shaw and Gideon Nave examine the idea of cognitive surrender, where users defer cognitive effort or judgement to an AI system. The concept is a useful warning: human approval is not meaningful if the person approving has stopped checking.
Leaders should therefore specify what a reviewer must actually do. Can they inspect the source? Identify unsupported assertions? Challenge exclusions and assumptions? Explain the final recommendation independently? A nominal human-in-the-loop step is not enough.
For more on the underlying human-factors evidence, see What human factors research already tells us about AI overreliance.
4. Test claimed benefits against real work
AI might reduce time, improve consistency, surface omissions or make information easier to find. These are different benefits, and they should not be treated as interchangeable.
The No man's hand study of AI-assisted police report writing is instructive. In a pre-registered randomised controlled trial involving 85 officers and 755 reports, the researchers did not find a statistically significant reduction in report-writing time. The finding concerns time in that study, not every potential effect of AI-assisted reporting: other possible benefits, including quality, were not established by this result.
In security, safety and resilience work, the strongest improvement may instead be a more complete handover, fewer missed issues during document review, faster access to the relevant procedure or better-supported escalation. Each needs an appropriate baseline and measure.
Measure the outcome the workflow is meant to improve, not the impressiveness of the demonstration.
Choose a bounded task, use approved material, compare AI-assisted and existing approaches, record errors and review effort, and decide in advance what result would justify wider use. Where stakes are high, include the cost of failure and the ability to detect it.
5. Move from unmanaged use to approved operational use
AI is best governed as technology within an actual workflow. The controls should become more demanding as its influence on decisions, actions or sensitive information increases.
Individual experimentation with limited organisational visibility. Inputs, checks and downstream influence may be unclear.
One defined workflow, authorised information, named reviewers, test cases, documented results and explicit stop conditions.
Defined sources and permissions, verification steps, accountable owners, monitoring, escalation and periodic review.
The purpose is not to rush every experiment into production. It is to make the progression deliberate and reversible, including the option not to proceed.
Start with one leadership conversation
A first meeting does not need an AI strategy, a procurement programme or a long demonstration. It needs agreement on five questions in a practical order.
- Existing use
- Human ownership
- Consequence of error
- Bounded test
- Evidence threshold
- Where is AI already being used? Include formal tools and informal practice.
- Which outputs and decisions must remain human-owned? Name the person or role responsible.
- Where would a confident but wrong answer cause harm? Consider operational, legal, financial, safety and reputational consequences.
- What is one task we can test safely? Define the inputs, boundaries, review method and stopping conditions.
- What evidence would permit wider use? Agree on quality, speed, review effort, exceptions and accountability before running the test.
These questions are more useful than asking whether a leadership team is enthusiastic or sceptical. They make the conversation specific enough to govern.
Leadership means preventing assistance from becoming dependency
AI can help people understand information, test options, prepare material and carry out defined tasks. It cannot take organisational responsibility for what follows. In consequential operational work, someone must still establish the evidence, recognise uncertainty, challenge the output and authorise action.
Leaders do not need to accept AI uncritically. Nor can they rely on ignoring it. Their responsibility is to make its role visible, place boundaries around it, and insist on evidence before trusting it more.
When AI can inform decisions but not carry responsibility, leadership means preventing assistance from becoming dependency.
Research and further reading
- Bick, Blandin and Deming, The Rapid Adoption of Generative AI. US survey evidence on adoption, used here as background rather than an organisation-specific estimate.
- Shaw and Nave, research on cognitive surrender. Discussion of how reliance on AI may affect users' independent judgement.
- Adams and colleagues, No man's hand: artificial intelligence does not improve police report writing speed. Randomised study of report-writing duration.
- What human factors research already tells us about AI overreliance. Related SIRV analysis.
- How SIRV measures impact. Approach to workflow evidence and assessment.
Discuss a controlled AI workflow
If your team is evaluating AI for security, safety or resilience operations, start with one defined task, its evidence, its review requirements and the outcome you need to measure.
Discuss a workflow
