AI desk research for a client brief, step by step
5 minutes read
Desk research is where most consultancies first try AI on client work, and where a fluent answer does the most damage. Here is a method that keeps the research tied to the brief and to sources you chose.
AI desk research works on a client brief when two things are fixed before the first query: the questions the brief needs answered, and the sources the answers are allowed to come from. Skip either one and the tool will still produce a tidy summary. It will just be a summary of whatever was easiest to find, and nobody on the team can say where a given line came from.
Most of the method below is ordinary research discipline. AI changes how fast a consultant gets to a first pass. The work still has to be scoped and sourced by a person, and read with some suspicion.
Turn the brief into research questions first
A client brief is not a research question. It usually states a decision the client needs to make and leaves the consultant to work out what has to be known to make it. Do that step yourself, in writing, before any tool is involved.
Three to six questions is normal for a first pass. Each one should be answerable from evidence and narrow enough that you would recognise a wrong answer. "What is happening in the market" fails that test. "Which three competitors changed their distribution model in the last two years, and what did each one say about why" passes it.
Writing the questions down also gives the reviewer something to check the output against. A reviewer who only has the finished summary can judge whether it reads well. A reviewer with the question list can see which questions went unanswered.
Choose the sources before the first query
Decide what counts as a source for this brief, then run the research inside that set. The open web is the default for most AI tools, and it should not be the default for client work. The longer argument is in how to choose approved sources for AI.
For desk research on a brief, it helps to sort sources by where they came from, because each kind does a different job.
| Source type | What it covers | What to check before using it |
|---|---|---|
| Internal | The firm's prior projects, methods, and research notes | That the material is current and was approved for reuse |
| Client | Documents the client shared for this engagement | That it is used only on this client's work |
| Public | Reports, filings, statistics, and publications you trust | That the publisher is one you would cite in the deliverable |
Keep the label attached to each source all the way through. A finding that rests on the client's own data is a different kind of statement from one that rests on a public report, and the person reading the draft needs to know which is which.
Run one question at a time
Ask each research question separately and keep the output separate. A single prompt that asks for "a market overview" invites the model to smooth over the places where the evidence is thin. One question per run makes the thin places obvious, because a question with little behind it comes back short.
Watch the steps as they run if the tool shows them. Which sources were opened, and which were passed over, tells you more about the quality of an answer than the answer does.
Read each finding by how well it is supported
Sort every finding by how much evidence is behind it before it goes anywhere near a draft.
- Supported: a source in your set states it, and you can open the passage.
- Inferred: no source states it, but it follows from two or more that you can open.
- Uncertain: the sources disagree, are out of date, or do not cover it.
Inferred findings are often the most useful part of desk research, because joining two sources is exactly what a consultant is paid to do. They need the label because the reviewer has to decide whether the inference is sound. The source cannot decide that for them. How to tell a sourced claim from an assumption goes through the checks for telling the two apart.
Uncertain findings are not failures. A short list of what the available sources could not answer is a legitimate research output, and it is often the start of the primary research plan.
Hand the reviewer the evidence with the draft
The research is done when a second person can check it without redoing it. That means the draft arrives with the question list, the sources used and their labels, and the status of each finding. A reviewer who has all of that can go straight to the inferred and uncertain findings, where their judgement is needed, and skip re-verifying what is plainly supported.
A reviewer who gets none of it can only trust the draft or repeat the research, and neither of those counts as reviewing it. The wider case for building AI work this way is in the guide to reliable AI client work.
How this looks in Hebno
Hebno runs research from a brief described in chat and shows each step and source as it goes. Each source is labelled internal, client, or public, and each finding is marked supported, inferred, or uncertain, so the sorting described above arrives with the draft. Citations stay attached while you edit, and you can finish the document in Word or PowerPoint. The work feature page shows a run from brief to draft.
Hebno does not decide that a finding is correct. The labels show a reviewer where to look, and the reviewer still makes the call.
