Day 9 · Use it on real work

Research, with sources

Reach a decision you can defend, and say what would change your mind.

14 min readPlus one task, about 10 min

The one idea

Research has two separate jobs: finding material and making sense of it. AI is often useful for the second job. It can compare documents, extract the disagreement and turn a pile of reading into a shape you can think with. Finding trustworthy material is still your responsibility unless you use a search mode that shows its work and you check that work.

A source is not a decorative link at the end of a paragraph. It is evidence that another person can open and inspect. A citation that does not exist, does not support the claim, or points to a page that says something else is worse than no citation because it creates false confidence.

Why it matters

A real decision rarely needs one answer. It needs a set of claims, the evidence behind them, the parts that remain unknown and a condition that would change your mind. That structure protects you from the smooth paragraph that sounds decisive while quietly leaving out the inconvenient fact.

The quality of a research answer tracks the quality of its sources. A model cannot turn a marketing page into an independent test by summarising it elegantly. It cannot repair a study with a weak method by putting the title in bold. Your judgement chooses what enters the pile; the assistant reads and compares the pile faster.

AI-generated illustration

The tools you will use

How to turn a question into defensible research

1. State the decision, not the topic. 'Should I buy this machine for the workshop this year?' is researchable. 'Tell me about machines' is a request for an essay with no test for usefulness.
2. Ask for the questions that would change the decision before asking for an answer. Cost, reliability, compatibility, maintenance and alternatives may matter more than the feature you first noticed.

   I am deciding [decision]. You are a sceptical adviser.
   What are the five questions I should answer before deciding?
   For each, tell me what kind of evidence would settle it.
   Do not recommend anything yet.

3. Gather three to six sources. Start with the official specification, regulator, original research, audited figures or a person with direct experience. Add a critical review or comparison, and record the date. Do not let one search result become the whole evidence base.
4. Give the assistant the sources and constrain it to those sources. Ask it to make a table of claim, source, evidence and limitation. This makes missing coverage visible instead of burying it in prose.
5. Open the important citations. Check that the page exists, that it is the right date and that it says what the report claims. If the source is secondary, follow its link to the original claim where practical.
6. Write your conclusion in your own words, then state what evidence would change it. A decision with no condition for revision is often a preference wearing research clothing.

See it

Take a small purchase decision: whether to move a team from one project-management tool to another. A weak research request says, 'Which is best?' The answer will tend to list features and declare a winner.

A useful request starts by asking what to measure: export compatibility, monthly cost at the actual team size, guest access, mobile reliability, data location and the time needed to migrate. You then gather the vendor's pricing and export documentation, an independent review that tests the mobile app, and a forum discussion from somebody who completed a migration. The assistant can make a comparison table and point out that the vendor's claim about export does not answer whether comments and attachments survive.

Your conclusion might be: 'Stay with the current tool for now because migration would not preserve the fields we need. I would change my mind if the vendor documents a complete export or if our team no longer needs those fields.' That is shorter than a report and more defensible because another person can see the evidence and the condition.

AI-generated illustration

Decide something you are actually unsure about

This is the longest task in the course and the one most likely to be useful in a real week. You do the sourcing; it does the reading. The finished decision note is something you can keep.

1. Pick a real decision you are circling — a purchase, a supplier, a technique, a course or a change at work. Write it as one sentence at the top of a note. Choose something important enough to care about and small enough to research today.
2. Do not ask for the answer. Ask for the questions first. Save the five questions and cross out any that cannot affect your decision. This stops a shiny feature from deciding the shape of the research.

   I am deciding [your decision]. You are a sceptical adviser.
   What are the five questions I should answer before deciding?
   For each one, tell me what kind of evidence would settle it.

3. Now gather sources yourself. Find three to six: official statistics, a specification sheet, a review site, a credible article, a forum where people who have actually done it talk, or the original study. Save the URL, title, author and date in your note.
4. Paste the sources in, or attach the documents, and ask the assistant to compare only them. Require it to mark anything not covered. If you use Deep Research or Perplexity instead, still open the citations that support the decision.

   Here are my sources. For each, summarise what it says about [decision].
   Then tell me where they disagree, and what none of them cover.
   Use only these sources. Cite the source beside each important claim.
   Say clearly if something is not covered.

5. Write your decision in one paragraph, in your own words rather than the tool's. Include the two strongest reasons, one limitation and the source behind the most important claim. If you cannot write it without the tool, you have not decided anything yet.
6. Add the most useful sentence in this whole course: 'The one thing that would change my mind is ____.' Fill it in specifically. Then ask the assistant to argue against you using only the sources above, and record whether its objection changes your conclusion.

   You disagree with my conclusion. Give me the strongest case against it,
   using only the sources above. Name the source for each claim and do not invent
   an objection that the sources do not support.

What you should notice

The final argument is where the value is. You have turned the assistant from an oracle into a sparring partner, and that is a different relationship to the tool than 'give me the answer'. Oracles are either believed or ignored; sparring partners make your reasoning sharper.

Notice also that the quality of your output tracked the quality of your sources, not the cleverness of the AI. If all six pages repeat a vendor's press release, a beautifully balanced table is still a vendor-shaped answer. If one source is an original study and another is a direct user account, the disagreement becomes useful rather than inconvenient.

In research you are still the researcher. The tool is doing the reading, sorting and first comparison. You are choosing the evidence, checking the citations, deciding what matters and owning the conclusion. That division is slower than asking for a verdict and much safer than pretending a verdict is research.

If you have five more minutes

Find one claim that all your sources repeat, then trace who said it first. You may discover they are all citing the same single study or press release. Add that dependency to your decision note; repeated is not the same as independently verified.

Tomorrow

Pictures, and where they are dishonest.

Get the rest by email

One lesson each morning at 07:00 IST, with the task written out in full so you can do it from your inbox.

Start the free 7am AI series

This is the course, not the newsletter. Joining does not subscribe you to the daily Brief.