Researching with AI in a verifiable way

Research starts before the prompt

A good question is not enough. We must define the universe of sources, the cut-off date, the period observed, the geography, the languages, the relevant events and the sufficient evidence. If these are not specified, an AI may cover only the most accessible pages and produce a coherent story that hides the blind spots.

Use two explicit modes, which also apply to the prompt factory:

Mode A: evidence provided. A set of fragments or documents with identifiers is supplied. The assistant works on what is available, points out contradictions and adds no external facts without authorisation.

Mode B: assisted research. The assistant is asked to research, but the prompt instructs it to first check whether any search tools are available. If none exist, the valid output is a search and verification plan, not a list of feigned sources. If they do exist, the URL, issuer, date, relevant fragment, access date, limitations and verification status are documented.

What the evidence file must contain

Do not confuse three objects: a source (site, repository or publication), a document (a specific item with author and date) and a claim (what we say on the basis of the document). Several news items repeating a press release may be three links but a single primary source of origin.

We propose a record with: ID, intelligence question, claim, locatable quotation or verifiable fragment, origin, publication date, access date, type of source, coverage, independent corroboration, reasoned confidence level and gaps. Confidence is a reasoned assessment, not a percentage calculated by the model.

Research protocol in six moves

1. Define. What we want to decide and which period and area affect the question.
2. Plan. Which types of sources should contain proof and which are missing.
3. Consult. Record what was actually seen, not what is presumably available.
4. Cross-check. Look for contrary evidence and detect dependence between sources.
5. Analyse. Separate facts, inferences, hypotheses and unknowns.
6. Deliver. Present conclusions with reference to their identifiers and a list of pending checks.

Common mistakes

Asking for «reliable sources» without defining relevance; taking a link you have not opened as verified; mixing facts from the corpus with external data without labelling; treating the number of headlines as the number of corroborations; not saving the moment of access; forgetting jurisdiction in regulation.

Rule to take away: no link replaces the explicit relationship claim → evidence → limit of that evidence.

Situation: the convincing answer we cannot publish

The management of Company A, an industrial manufacturer, asks: «Are there enough indications to invest in industrial thermal storage in the next financial year?». A person consults their AI assistant. The assistant returns:

«Industrial thermal storage is already widely deployed in Europe. Three competitors launched commercial solutions in 2026, production costs have fallen by 40% and regulation will require its adoption. Recommendation: immediate investment».

Warning: the paragraph above is a simulated response, not a real statement about technology, companies or regulation. It must not be indexed or shared outside the context of this case.

The problem is that a fluent response mixes five kinds of claims: adoption, competitors, costs, legal obligation and recommendation. None comes with a verifiable primary document or defines a date, jurisdiction or scope.

The first task is not to ask again, but to break it down

Rule: the fact that the AI can provide a link does not in itself prove that the link supports the claim. Check the page, date, scope and whether the quotation matches the text.

Technique: verifiable chain of evidence

For each claim, keep the following fields:

1. Identifier: A01…; avoids losing the link with the report.
2. Exact claim: copy only the proposition that needs proof.
3. Locatable source: URL or document and section; if none exists, «no evidence».
4. Nature of the source: official register, company document, research, secondary news, commentary.
5. Publication date and access date: distinguish historical facts from current validity.
6. Degree of support: direct, partial, contradictory or absent. Do not replace it with an invented percentage.
7. Consequence for the decision: what we would do differently depending on whether it is confirmed or ruled out.
8. Reviewer and review date.

What a responsible correction looks like

Instead of «we must invest immediately», the didactic briefing should say:

«The signals in the file suggest an emerging interest in industrial thermal storage, but the available evidence is not enough to establish commercial adoption, cost reduction or regulatory obligation. We recommend a limited technical assessment, two interviews with potential users and monitoring of official publications before approving an irreversible investment».

Important: this text is also a simulated recommendation, conditional on fictional evidence. It illustrates an argumentative structure; it does not represent a market conclusion.

Professional prompt for you to take away

Act as a claim auditor in a competitive intelligence process.
GOAL: review a draft intended for [recipient] to support the decision [decision].
MATERIAL: [paste text and available references].

DO NOT research or cite sources that you have not actually consulted.
If you lack browsing or access to a source, say so expressly.

1. Extract atomic claims and number them.
2. Classify each one: fact, estimate, interpretation, hypothesis, recommendation.
3. For each claim, link the evidence provided or mark «not verified».
4. Point out omissions of period, scope, unit, method and primary source.
5. Distinguish disagreement between sources from absence of evidence.
6. Write a defensible version, keeping relevant uncertainties.
7. Propose specific checks and a human owner for each.
8. Do not turn correlations into causality or repeated mentions into independent sources.

RETURN:
A) claim-evidence matrix; B) pending critical points;
C) revised text; D) conditional decision options;
E) verification plan with owner and date.

When not to delegate to AI

Do not let an assistant decide regulatory compliance on its own, replace legal advice, automatically send high-impact conclusions or feed confidential information into unauthorised systems. Data processing depends on the provider, the configuration and the contractual terms; it cannot be deduced from the use of MCP or from a prompt.

Assessment rubric (0–2 points per criterion)

A total of 8–10 indicates that the exercise is well solved; this is our own teaching rubric, not a normative instrument.