Agent-built market research reports that survive review

An agent can read more sources in an afternoon than an analyst gets through in a week, and it can lay the results out in a tidy report with headings, tables and footnotes. That is the problem as much as the promise. A polished market report whose numbers cannot be traced back to a page is worse than a rough one, because it travels further before anyone checks it. The work that makes agent-built research useful happens before and after the agent runs: framing the question, deciding which sources count, fixing a structure that keeps evidence attached, and reviewing before anything is shared. The same discipline carries over when a one-off report turns into a recurring intelligence feed.

Frame a decision, not a topic

“Research the heat pump market” is a topic. An agent given a topic produces a survey: everything it can find, weighted by how easy it was to find. A useful question is tied to a decision someone will make, and it sets the boundaries the agent should stay inside.

A short brief written before the run does most of the work:

• Decision: whether to offer a self-serve plan for small installation firms.
• Question: how many such firms exist in the target countries, how they buy software today, and who already serves them.
• Scope: two named countries, residential installers only, sources from the last two years.
• Out of scope: commercial contractors, hardware manufacturers, pricing advice.
• Audience: the product lead, who will read the summary and the competitor table first.

Then write down what counts as done. Acceptance criteria for a research report are mostly about evidence, for example: every market figure names its publisher and the year it describes; every competitor row links to the page where its positioning was read; any claim no source supports appears under open questions, not in the summary. Writing these first, as acceptance criteria before the agent starts argues, gives the reviewer something concrete to check against.

Approve the sources before the agent reads them

Market research is unusually exposed to weak sources. Searches for market sizes return undated press releases from report resellers, figures copied from one article to the next, and generated listicles. An agent that treats all of them as equal will happily average them.

Decide up front which sources may answer, and in what role:

• Primary: statistics offices, regulators, company filings, a vendor's own product and pricing pages, trade association data.
• Secondary: analyst reports you are licensed to use, reputable trade press, interviews you ran yourself.
• Context only: blogs and aggregators, which may point to a primary source but should not be cited in its place.
• Excluded: undated pages, paywalled content you do not license, and anything that cannot be opened again later.

In GoCux this list belongs to a research plan, alongside the question and the ordered steps. Saving the plan creates a draft; a person approves it before anything runs, and the run cannot reach beyond what was approved. An empty allowlist means nothing qualifies as a source. If the question shifts halfway through, the right move is a revised plan rather than a quietly widened one. GoCux also does not browse the open web by itself: web pages come in through an authorized browsing or export step and keep their original address and retrieval date. Deciding which sources may answer goes deeper into building the list.

A report structure that keeps evidence attached

Agree the skeleton before the run, so the agent fills slots instead of inventing a format each time. A structure that works for most market questions:

1. Answer to the decision question, in one paragraph, with the main uncertainty stated plainly.
2. Market definition: who is in and who is out, so readers know what the numbers describe.
3. Size and direction: each estimate with its publisher, reference year and method, plus a short note where estimates disagree.
4. Customers and buying behaviour: segments, how they buy, what they use today.
5. Competitors: a table with one row per company, each row carrying its evidence.
6. Trends and risks, each tied to a source.
7. Open questions: what the sources did not answer.
8. Source list with retrieval dates.

The competitor table is where reports most often drift away from evidence, so give it explicit columns:

Look at the second row. “Not published”, with a reason, is a finding; a guessed price is not. GoCux keeps these apart: a claim backed by a quotation and a question nothing answered are both retained, in separate places. Its Extraction pulls values out of material you supply, each with the passage behind it, and leaves a missing value marked unknown. FindAll builds candidate tables in which every row carries its evidence and coverage gaps are reported rather than smoothed over. Even then, a citation only points at evidence; someone still has to read it. Citations that hold up covers what a reviewer should look for.

What agents do well here, and what stays with people

Agents are strong at gathering and tabulating: working through an approved source list, pulling figures into a consistent shape, building the first version of the competitor table, drafting sections against the skeleton, and flagging where two sources disagree. People are better at framing, at judging which of two conflicting estimates fits the market definition, and at writing the recommendation.

A few failure patterns are worth checking for every time:

• Averaging incompatible estimates. Two market sizes built on different definitions cannot be averaged. The report should show both and explain the difference.
• Circular sourcing. Five articles quoting the same figure are one source. Trace numbers back to where they first appeared.
• Stale figures presented as current. A number from an old report needs its reference year beside it.
• Vendor claims stated as fact. “The leading platform for installers” is what the vendor says about itself.
• Numbers with no source at all. If a figure has no passage behind it, treat it as invented until shown otherwise.

One cross-check is cheap and revealing: ask the agent for a rough bottom-up estimate (number of firms times a plausible spend per firm, each input sourced) and compare it with the published top-down figures. A large gap does not tell you which side is wrong, but it tells the reviewer where to look.

From a report to a market intelligence tool

Once the same questions come back every month, such as competitor pricing, new entrants, funding rounds and regulatory moves, it is time to turn the report into a pipeline. Teams usually assemble one from a handful of parts:

• Inputs: a search or news API, company data providers, filings, and your own notes from sales calls.
• A tracked entity list: the companies, products and regulators you follow, with stable identifiers.
• An extraction schema: the fields you want per entity, each stored with the passage and address it came from and the time it was read.
• Change detection against a baseline, so the system reports what moved rather than everything it saw.
• A review step before anything reaches a dashboard, a digest or a colleague's inbox.

Two design rules save a lot of trouble later. Store the evidence with each value, not just the value, so any number on a dashboard can be traced. And keep “nothing changed” distinct from “could not check”; a source that failed to load should not look like a quiet week. In GoCux, Pulse handles the recurring part: a question asked again over sources you pick, with its cadence, timezone and definition of a material change approved as an exact version, and each run reviewed one change at a time. A detected change is information; it does not act for you. For neighbouring workflows, see automating competitive intelligence and market mapping with AI.

Review before the report travels

A report is most dangerous right after it is finished and right before it is forwarded. A short review pass catches most problems:

1. Open a sample of citations, including every figure in the summary, and confirm each passage says what the report says.
2. Check the reference year on every size and growth figure.
3. Confirm the open questions section exists and is honest; an empty one in a broad report is a warning sign.
4. Read the summary last and make sure it claims nothing the body does not support.
5. Check that the source list matches the approved plan.

Sharing is a consequential step in its own right. In GoCux, sending something out pauses so a person can approve, edit or reject it, and an export waits with the files in scope listed. Artifact Studio keeps deliverables versioned with their evidence links, so a correction becomes a new version instead of overwriting what people have already read.

Checklist for your next agent-built report

• The brief names the decision, scope, exclusions and audience.
• Acceptance criteria are written down, and most of them concern evidence.
• Sources are approved by role before the run, with exclusions stated.
• The report skeleton is fixed, including an open questions section.
• Every figure carries its publisher, reference year and a supporting passage.
• Conflicting estimates sit side by side instead of being averaged.
• Recurring questions run against a baseline, with “could not check” kept apart from “no change”.
• A person reviews citations and the summary before the report is shared.