When the Summary Outruns the Source: Why UK Insight Decks Need a Provenance Page
The convenience is real. The risk is quiet.
A UK insight team can now paste a 60-page report into an AI assistant and receive a board-ready summary before the coffee goes cold. That is a genuine productivity gain. It is also a new point of failure. The summary sounds certain because the tool is designed to produce fluent prose, not because it has verified the sample, the fieldwork dates or the question wording. When the summary travels from the insight team to the finance director to the chief executive, the original caveats often fall away. What remains is a confident sentence on a slide.
The problem is not that AI summarises badly. The problem is that summarisation compresses provenance. A report might say that 68 per cent of respondents agreed with a statement, but the useful decision context sits around that number: who was asked, how many, when, in what mode, with what prompt, and what the research could not cover. An AI summary can preserve the headline and lose the chain of evidence. In a board pack, that is a material risk.
What a provenance page should record
A provenance page is a one-page record attached to any insight deck that relies on external or AI-summarised evidence. It should be short enough to be read and specific enough to be challenged. At minimum, it should cover five fields:
- Source: the publisher, commissioner or data owner, plus a link or file reference.
- Sample: the population, achieved base, recruitment method and any weighting.
- Date: fieldwork dates, publication date and the date the summary was produced.
- Method: data collection mode, question design, analysis approach and whether the work was primary, secondary or modelled.
- Known gaps: coverage limits, low bases, non-response, missing segments and any caveats that could change the decision.
For AI-generated summaries, add two more fields: the model or tool used, and the prompt or instruction that shaped the output. That second field matters. A prompt that asks for three reasons to enter the market will produce a different summary from one that asks for three reasons to enter and three reasons to wait. The provenance page should make that editorial choice visible.
Why UK boards should care now
UK organisations are under pressure to move faster with less external spend. Research budgets are being scrutinised, procurement cycles are longer, and internal teams are being asked to do more with existing evidence. AI tools look like a way to close the gap. They can be. But they also make it easier to present a secondary source as if it were primary evidence, or to treat a small qualitative study as if it were a nationally representative survey.
The governance question is straightforward: can the organisation trace a board-level claim back to its source in under five minutes? If not, the insight deck is asking the board to trust the summary rather than test the evidence. That is a weak position for any decision involving budget, market entry, pricing or customer strategy.
A provenance page also helps procurement. When an external agency supplies a report, the page gives the buyer a consistent checklist for handover. When an internal team uses a published survey, the page records the licence terms and the limits of reuse. When an AI tool is used, the page creates an audit trail that can be reviewed later if the decision is challenged.
How to challenge an AI-generated summary
The challenge should happen before the deck is signed off, not in the meeting. A practical sequence is:
- Ask for the source sentence behind each headline number. If the AI cannot point to it, the number is not yet evidence.
- Check the base. A finding based on 42 respondents is not the same as one based on 420, even if both are expressed as percentages.
- Check the date. A 2023 survey may still be useful, but it should not be presented as current market conditions in 2026 without a caveat.
- Check the question. Agreement with a prompted statement is not the same as unprompted demand.
- Check the gap. Ask what the research does not tell you. If the answer is nothing, the summary is probably too confident.
These questions are not hostile to AI. They are the same questions a good research buyer would ask of any supplier. The difference is that AI makes the summary so easy to produce that the questions can be skipped.
A discipline, not a block
A provenance page should not become a bureaucratic appendix that nobody reads. It is a decision guardrail. It tells the board where the evidence is strong, where it is thin and where the insight team has made a judgement. That is more useful than a false sense of certainty.
The practical rule is simple: no provenance, no headline. If the page cannot be completed, the slide should be labelled as indicative, or held back until the source is confirmed. That rule will slow some decks down. It will also stop some weak evidence from shaping a budget.
AI can help UK teams read more widely and summarise faster. It cannot manufacture the sample, the fieldwork or the caveats. A provenance page keeps the summary honest by keeping the chain of evidence visible. In 2026, that is not a nice-to-have. It is basic research hygiene.
