AI Research Agent: Specification

What this agent is

A research agent that takes one written business question, gathers evidence from sources you have allowed it to use, and returns a brief in which every claim carries a citation you can open.

It is not a search box and it is not an oracle. The output is a document with a bibliography, produced to a template, that a named person reviews and signs before it informs a decision. That constraint is the design: an unsigned brief is an unowned brief, and unowned research is how a plausible number ends up in a board pack with nobody able to say where it came from.

The failure it is built to prevent

The dangerous output of a research agent is not an obvious error. It is a fluent, well-structured brief containing one figure no source actually states, sitting next to a link that looks like it supports it. Nobody catches it, because the citation is right there.

Everything in this specification — quote-level extraction, retrievable locators, source deduplication, the sampled human check on citations, and explicit permission to answer "not established" — exists to make that specific failure hard rather than merely unlikely.

Oversight and failure modes

Circular sourcing. Three articles citing the same press release is one source, not three, and it reads on the page as consensus. The agent traces claims back towards their origin where it can, and marks corroboration it cannot verify as unverified rather than counting it as a second source.

A citation is not a verification. A link that exists is not a link that says what the sentence built on it claims. This is why a person opens a sample of citations on every brief, and why that sample is weighted towards the claims the decision rests on rather than drawn at random.

Recency without a date is worthless. Every source carries its publication date and the date it was retrieved, and the brief carries a shelf life on its face. A market figure quoted with no date is not evidence, it is folklore.

Terms of use are a real constraint. Paywalled content, licensed databases and sites whose terms forbid automated access do not become fair game because the retrieval was automated. The source allowlist is a legal artefact as much as a quality control.

The absence of evidence is the most valuable output. An agent that always produces an answer will produce one when none exists. "Not established in the sources available" has to be an acceptable, expected result, and the list of sub-questions it could not answer belongs at the front of the brief rather than in an appendix nobody reaches.

It does not decide. The brief informs a decision made by a person who signs it. An unsigned brief is a draft — including, and especially, when it is inconvenient to treat it as one.

What it might save

Dependent on how much desk research your team does and how deep each question runs, so read this as an estimate for planning. For a team producing a handful of briefs a month — competitor reviews, market scans, supplier assessments — plan for something in the range of four to twelve hours per brief on gathering and first-draft writing.

Note what is deliberately not in that range: verification time. The saving is in collection and drafting. Verification is the part that must not shrink, and a design that appears to save time there is saving it in the one place you cannot afford to.