AI and fact-checking: producing reliable client notes

Large language models have transformed the productivity of intelligence and consulting teams. An analyst who used to spend six hours writing a synthesis note now spends two, and the result is often more structured than what they would have produced alone. But this acceleration has a documented downside: LLMs produce false statements with the same confidence as true ones. In a consulting firm where credibility is the primary asset, a client note containing an incorrect date, a misattributed role or an invented statistic can cause lasting damage.

This article explores the mechanics of hallucinations, the risk zones specific to strategic intelligence, and the verification protocol to put in place so that AI remains an accelerator without becoming a source of reputational risk.

Understanding hallucinations: why LLMs invent

An LLM does not "lie" in the sense of seeking to deceive. It predicts the next most likely token given the context. When the correct answer is not in its training data, or when the data is ambiguous, it generates a plausible answer rather than admitting ignorance. This property is useful for creativity and synthesis; it is dangerous for factual accuracy.

In strategic intelligence, hallucinations manifest in three recurring patterns.

Role and person confusions. An LLM may attribute to a former official the statement made by their successor, merge the attributions of two executives with similar names, or invent an appointment that never happened but is consistent with an actor's likely trajectory. These errors are particularly insidious because they fit into a coherent narrative.

Approximate dates and figures. The model knows that an agreement was signed but not exactly when. It produces a plausible date. It knows the order of magnitude of a turnover figure but not the exact number: it generates one that falls within the expected range. In a client note, a date shifted by six months or a figure overestimated by 20% can completely change the interpretation.

Ghost citations. This is the most documented and most damaging form: the LLM generates a plausible quotation attributed to a real source (a report, an executive, a regulator) that does not exist or was never produced. The quotation is stylistically credible, the source is reputable, but the text is invented.

Priority risk zones to verify

Not all elements in a note carry the same risk. An effective verification protocol prioritises accordingly.

Red zone, systematic verification required:

• Precise figures (percentages, amounts, volumes, headcounts)
• Dates of events, signings, official publications
• Titles, roles and reporting lines of named individuals
• Direct quotations attributed to a named source
• References to official documents (reports, regulatory texts, decisions)

Orange zone, check on decision-critical points:

• Multi-year event timelines
• Causalities between events ("it was law X that led company Y to...")
• Competitive positioning comparisons between actors
• Quantified trends presented as established

Green zone, proofreading sufficient:

• Argumentative structure and formatting
• Reformulations of clearly identified source passages
• Tone and editorial register

An analyst who applies this triage can check a 6-page note in 20 to 30 minutes rather than two hours, without sacrificing reliability on the points that matter.

The verification protocol in practice

Verification is not a proofreading pass: it is a distinct process, with its own steps and rules.

Step 1: identify verifiable claims. Read through the note and mark every factual claim: date, figure, role, quotation, reference to a document. This is the most time-consuming step but it defines the scope of the work that follows.

Step 2: source each claim. For each marked element, identify the primary source that confirms it. Strict rule: a secondary source (an article that cites a report) is not sufficient if the figure or date is central to the argument. The primary source can be an official press release, a regulatory document, a press file from the organisation concerned.

Step 3: reconcile divergences. When verification reveals a discrepancy, shifted date, different figure, inaccurate title, correct the claim and note the exact source as a reference. Do not attempt to "average" two conflicting sources: escalate the ambiguity rather than mask it.

Step 4: final validation by the responsible author. A note signed by the firm commits its reputation. Final validation belongs to a senior consultant or partner, even when the drafting was AI-assisted.

Labelling AI-generated content: stakes and practices

The labelling question is as much ethical as regulatory. Should you tell the client that a note was produced with AI assistance?

The regulatory context. The European AI Act, progressively in force since 2024, imposes transparency requirements on certain AI systems. For B2B consulting notes, direct obligations remain limited, but the notion of professional good faith applies. Misleading a client about the provenance of a deliverable may constitute a breach of the duty of advice.

What market practices show. Among firms that have adopted AI in their workflow, three postures coexist. The first, total non-disclosure, is increasingly untenable as clients become more informed about the tools. The second, disclosure at the method level (in general terms or at the first briefing), has become the prevailing norm. The third, note-by-note labelling, is practised by firms that have made it a differentiation argument on rigour.

The pragmatic recommendation. Mention the use of AI at the method level, without necessarily labelling each document. Be transparent about the verification protocol in place. Never imply that AI "thinks for" the consultant: it accelerates, structures, reformulates. Analysis and validation remain human.

A specific case: quotations and data. If a note includes a direct quotation drawn from a document, that quotation must be verified and its source indicated, whether or not drafting was AI-assisted. AI labelling does not replace citation of sources.

What the AI Act concretely changes for consulting firms

The AI Act classifies AI systems by risk level. Intelligence monitoring and document synthesis tools used by consulting firms fall into the "limited risk" category for the majority of use cases. Direct obligations are light: no algorithm audit, no registration in a national database.

But two obligations merit attention.

Decision traceability. When a recommendation relies directly on an AI-produced analysis, the firm must be able to document the reasoning. "The AI said so" is not an acceptable justification in an audit or litigation context.

Source bias risk. An LLM trained on an anglophone corpus will carry geographic and linguistic coverage biases. For intelligence covering markets poorly represented in English, Central Europe, sub-Saharan Africa, Southeast Asia, this bias can introduce significant blind spots. Firms operating in these geographies must factor this into their protocol.

How Sentinel Briefing handles reliability

Sentinel Briefing is built around a simple principle: AI filters and structures, the consultant validates. The system retains the primary sources associated with each signal, allowing the analyst to trace directly back to the source document for verification.

Notes produced are labelled "AI-assisted, sources verified by the team" in exports, with the list of primary sources in an appendix. This editorial choice allows firm users to distribute their deliverables with the transparency that protects their reputation, without adding to their production process.

Reliability is not an option in consulting. It is the precondition for everything else.