Advisory, Not Automatic: A Boardroom Protocol for Machine Recommendations

Artificial intelligence rarely arrives in the UK boardroom as a robot. It arrives as a number: a churn forecast, a supplier risk score, a capital allocation ranking, a suggested strategic option. The presentation is usually tidy, the language confident, and the provenance thin. That combination is the governance risk. The danger is not that boards believe machines are infallible; it is that algorithmic outputs acquire the appearance of neutrality and quietly displace the scrutiny that a human recommendation would attract.

As 2026 unfolds, more mid-market and listed boards are receiving AI-generated material without a corresponding discipline for examining it. The answer is not to ban the tools or to require directors to become data scientists. It is to insist that machine advice enters the boardroom on the same terms as any other adviser: visible, challengeable and owned.

The governance gap is not technical

Directors do not need to understand transformer architecture. They do need to know what question was asked, what data was used, what assumptions were made, how confident the model is, and who is accountable for the recommendation. Under the Companies Act 2006, directors must exercise reasonable care, skill and diligence and promote the success of the company. A recommendation that cannot be attributed to a named owner or interrogated on its inputs is difficult to reconcile with that duty.

The UK Corporate Governance Code expects robust challenge and a board that can explain its decisions. If an AI output cannot be explained in plain English, the board cannot challenge it in any meaningful way. The gap is therefore not a technical one. It is a governance one, and it sits squarely with the board rather than the technology function.

Treat AI advice as a witness, not a verdict

Every material AI-generated recommendation should arrive with a short provenance note. One page is usually enough. It should record the source data and period, the model or tool and its version, the business owner, the known limitations, the confidence level, and the conditions under which the recommendation would change. Think of it as a witness statement rather than a verdict.

The board's questions should follow naturally. What would make this wrong? Which inputs are stale or incomplete? What alternative did the model reject, and why? What would management have recommended without the model? That last question is often the most revealing. The difference between the human baseline and the algorithmic recommendation is where insight sits.

Name the human owner

No AI recommendation should be presented as coming from "the model". Every material output needs a named executive or function owner who is prepared to stand behind it, explain its limits, or explain why they have diverged from it. This is not about blame. It is about accountability.

Board minutes should record the owner, the material caveats and the rationale for the decision. Where the board accepts a recommendation that the owner has qualified, that qualification should be minuted. Where the board rejects the recommendation, the reason should be recorded. Over time, these records become a useful account of how judgement was exercised rather than merely what was decided.

Build a challenge slot into the agenda

AI-generated material should not be buried in papers marked for noting. Give it a fixed slot on the agenda, with time for challenge. For high-materiality recommendations, appoint a designated sceptic or ask management to present a red-team view. Require a counterfactual: how does the recommendation change if a key assumption moves by ten or twenty per cent? Ask for a range rather than a single figure.

Boards should also watch for automation bias, the human tendency to accept computer-generated suggestions uncritically, particularly when the output is precise. Tidy numbers create false confidence. A single-point forecast is often less decision-ready than a distribution, a set of scenarios or a clearly stated confidence interval.

Keep an advice register

A simple register of material algorithmic advice can do for AI what an assumption register does for strategy. Record the date, the decision, the tool or model, the owner, the confidence level, the board's action and a review date. Review the register after six or twelve months. Was the advice right? Did the board over-rely on it? Did the model consistently understate risk or overstate returns?

This is not bureaucracy for its own sake. It creates institutional memory and makes patterns visible. It also gives the board evidence with which to challenge repeated claims of model accuracy.

Look through the tool, not at it

The Oculus test is simple: can a non-executive director, reading the paper, see the chain from data to recommendation to decision? If not, the board is not governing AI; it is receiving it. Clarity is not the same as certainty. A good AI paper makes uncertainty visible rather than hiding it behind a clean interface.

For third-party tools, procurement and assurance should match materiality. That means appropriate explainability, audit rights, incident notification and data provenance. The board does not need the contract detail, but it should see the assurance. Internal audit or the risk function can provide independent review, and regulated firms should align the approach with existing model risk expectations.

AI can sharpen board judgement. It can widen the evidence base, surface patterns and test assumptions at speed. But it will only do so if it is treated as an adviser with a name badge, a track record and a challengeable case. The alternative is a quiet transfer of judgement to systems that no one can cross-examine. UK boards should adopt the protocol now, before the recommendations become too smooth to question.