The owner I am thinking of found out from the wrong direction. Her project manager had spent two days executing a course of action that, on her reading of the AI output the manager had used, was clearly a recommendation, one option among others, with an obvious trade-off she would have wanted to discuss. To the project manager, looking at the same paragraph, the same words read as a settled call. He had taken “the most effective approach is to consolidate the two suppliers” as a decision and acted on it. Neither of them was wrong about what the page said. The page said both things at once, in the same grammatical mood, and they had each chosen which version to read.
That is what this post is about. AI tools produce a single fluent block of text that mixes facts (what is true), recommendations (what to do) and decisions (what has been chosen) into one continuous flow. The blur is fluent enough to pass review unnoticed. The owners whose teams can separate the three on the page have an evaluation discipline that costs almost nothing. The owners whose teams cannot are running on outputs that quietly promote suggestions into commitments behind their back.
What are facts, recommendations and decisions in an AI output?
Three different kinds of statement, each requiring a different review move. Facts are statements about what is true, verifiable against data, regulators, or the firm’s records. Recommendations are suggestions about what to do, conditional and challengeable, framed as “if your goal is X, consider Y”. Decisions are choices that have been made by someone with the authority to make them, and they carry accountability for the outcome.
The distinction sounds academic until a paragraph lands on a desk that contains all three with no labels between them. A typical AI response to “what should we do about onboarding complaints?” will report that three customers complained this week (fact), suggest a progress indicator would reduce drop-off (recommendation), and use the phrase “we will implement the indicator by end of June” (decision-shaped language with no owner) inside the same paragraph. The work the team has to do is to read those three sentences as three different objects, not as one position the AI is taking.
Why do AI tools blur the categories by default?
Because language models are trained to produce coherent, persuasive text, not text with visible seams, hedging or explicit category markers. When a model generates output about a business question, it integrates facts, analysis and prescriptive language into one narrative paragraph because that is what its training data taught it good writing looks like. The markers that would signal “this is suggestion, that is verified fact” are smoothed away for the sake of readability.
There is a second mechanism on top of the first. Research from Stanford on eleven leading AI systems, reported by the Associated Press, found the systems affirmed user actions forty-nine per cent more often than human peers did, including in scenarios involving questionable choices. The tendency means recommendations routinely arrive in the grammatical mood of prescription, “the best approach would be” rather than “one option to consider is”, and the framing itself converts a suggestion into something that reads as already authorised. MIT Sloan’s work on AI hallucinations adds the third layer, that confidence and accuracy in language-model output are not the same thing, and a recommendation presented alongside accurate facts inherits the apparent certainty of those facts.
Where will you actually meet this blur in daily work?
At the points where the team is short on time and AI output is doing the structural work, briefing documents, meeting prep packs, proposal drafts, internal SOPs, summaries of customer feedback feeding a roadmap. The exposure is highest wherever a busy operator is reading AI output to extract “what should we do”, because the model has already arranged the page to answer that question in a way that sounds settled.
The Air Canada case from 2024 is the cautionary version of the same pattern, on the customer-facing side. The airline’s virtual assistant told a bereaved passenger it could refund a bereavement-fare discount retroactively within ninety days. The passenger booked two expensive tickets and discovered the policy did not exist. The Civil Resolution Tribunal ordered Air Canada to pay damages, finding the chatbot’s confident assertion had crossed from advice into representation of fact, and the passenger had been entitled to rely on it. The boundary was there, between what the model was suggesting and what the firm had actually decided as policy, and it was not maintained. The principle applies inside the firm as much as outside it. A recommendation dressed as a decision in a project-manager brief is the same boundary failure, just with smaller and quieter consequences.
When should you stop and ask “who actually decided this?”
When the AI output contains a phrase that sounds settled and you cannot name the person who settled it. The yellow-flag phrases are predictable, “the best course of action”, “we should implement”, “the preferred method”, “implement by Friday”. Each reads as a prescription, none carries an owner unless the team adds one. A team that pauses at those phrases and asks “who decided this?” catches the blur before it travels.
The check is fast and the language reset is mechanical. If the AI output reads “implement weekly stand-ups”, the team treats it as a recommendation and asks the model to reframe in advisory terms, “you could consider implementing weekly stand-ups, with the trade-off that synchronous meetings reduce flexibility for distributed teams”. Twenty seconds of editing returns the output to the register it should have been in. The decision itself, whether to actually run the stand-ups, is then made by the person who carries the consequences, in the room, with the trade-off visible. That is where decisions live in a small firm, and it is the boundary the AI output should not have crossed in the first place.
The structural prevention is even smaller. A standing prompt template that asks the AI for three labelled sections, facts with their data and sources, recommendations with at least two options and their trade-offs, and open questions still needing human input, returns output already sorted into the three categories on arrival. The cost is one line in the prompt. The benefit is that the team spends its review minutes on verification, challenge and authority confirmation, not on untangling what the AI meant to say.
How does this connect to wider AI governance and review disciplines?
Cleanly, and at three points. Facts feed into the verification disciplines already on this site, the spot-check, the cross-reference against source data, the factual-accuracy pass. Recommendations feed into the challenge disciplines, the sparring-partner role for AI on hard decisions, the brand-voice and quality-signal passes. Decisions feed into the Article 22 GDPR rule on meaningful human review of significant automated decisions, which the ICO has clarified for UK deployers.
The Boston Consulting Group’s OVIS framework, separating Owner, Veto, Influence and Support roles in a decision, is the structural backbone on the decision side, every decision an AI influences needs a named owner who is accountable for the outcome. The CTO Advisor’s auditable-authority piece adds the audit-trail layer, the firm should be able to point to where the AI advised and where the human decided, as two separate events on the record. None of this requires a governance function or a compliance budget at a fifteen-person firm. It requires the team to be able to tell the three categories apart on the page, which is the move this cluster keeps coming back to.
The separation of facts, recommendations and decisions is the small structural move that lets the team use AI output without it quietly promoting itself into authority it does not have. It sits upstream of any decision-rights framework you may layer on top, and downstream of nothing. If you want a sounding board on where that blur is already travelling through your own decision documents, book a conversation.
