An audience description produced quickly is not yet an audience insight fit for a commercial decision. The missing step is often evidence: who was observed, when the information was collected and whether it supports the conclusion the strategy team wants to use.
November 13 reporting by Marketing Dive revisits GWI’s Artificial Confidence research. GWI reports that 58% of respondents had made a significant decision using an AI insight later found wrong or misleading; 36% said they verify outputs most of the time. The useful management question is what gets checked before a generated description becomes a brief.
Read the sample before borrowing the headline
GWI’s release describes 1,883 senior marketing, insights and agency professionals in the UK, US and Singapore, at companies with at least 500 employees or £20 million annual turnover. It specifies July 2026 fieldwork. These are reported professional experiences, not a controlled trial of every model or a representative estimate for all small businesses.
The findings therefore do not mean that 58% of AI answers are false. Nor do they quantify the average loss from one inaccurate audience profile. GWI also sells research, so its commercial interest belongs in the interpretation. The study is a useful signal of a verification problem, with boundaries that should travel with the headline.
Keep generated hypotheses separate from observed evidence
A model can propose that a customer segment values convenience, reassurance or status. Mark that as a hypothesis until the team can attach relevant evidence. For each consequential claim, record the population, market, collection date, original source and the reasoning connecting observation to conclusion. Preserve disagreement instead of smoothing it into a persuasive persona.
Checking with another model may identify inconsistencies, but agreement between two generated answers is not an independently observed customer response. For a claim that could change targeting, positioning or an offer, look for suitable human research or relevant first-party evidence. If the evidence is absent, narrow the decision or commission the missing investigation.
Put review and repair inside the productivity calculation
Our proposed operating measure is the cost of an accepted, usable insight. Include drafting, source retrieval, editorial or analytical review and rework after errors are found. Compare that total with a defined earlier process for a similar task. A shorter first draft can coexist with a more expensive finished decision input.
Keep quality visible beside speed. Record how many proposed claims were accepted, corrected, rejected or left unresolved, and why. The aim is not an arbitrary pass-rate target that encourages reviewers to approve weak material. It is to learn which tasks need better inputs, which need a smaller scope and where automation actually removes work rather than moving it downstream.
Match the evidence requirement to the decision
Exploratory ideas for a workshop and a segment definition that will direct a substantial campaign should not carry the same review burden. Name a decision owner and the evidence needed for the intended use. Allow a team to brainstorm with provisional material while making its provisional status explicit; require stronger support before committing money or changing a customer promise.
At the next strategy review, bring one AI-assisted audience claim, its supporting evidence and the full checking effort. Ask whether the claim changed a decision and whether the source justified that change. GWI’s research supports taking this problem seriously. The practical response is a traceable path from observation to action, with the cost of that path counted honestly.
Sources
- AI-derived audience insights hurt results: What the numbers say
- GWI: Artificial Confidence — The AI Trust Gap
- GWI: 58% of Business Leaders Have Made a Major Decision on AI Insight That Proved Wrong
