Deep-research tools in ChatGPT, Gemini and Claude can now search across the web, files and connected services, then return a cited report. ChatGPT also lets you review the research plan and choose which sites it can use. That is a real improvement. It still does not guarantee that the work answers the question you actually need to settle.

If you ask for "research on AI customer service tools", the model has no finish line. It can collect features, prices and opinions indefinitely. The result may look thorough while quietly avoiding the trade-off that matters to you.

Start with a decision, a date and a constraint

Turn the topic into a choice. For example: "Should our 12-person support team trial an AI assistant before November, if we cannot send customer data outside our approved systems and the trial must save at least five hours a week?"

That question tells the AI what good research must do. It needs to identify suitable options, check the data rules, estimate the likely time saving and expose where the evidence is weak. A general market summary cannot answer it.

Make it separate evidence from judgement

Most research failures are not caused by one completely invented fact. They happen when a real fact, an assumption and a recommendation are blended into one confident paragraph. Ask for those parts separately.

  • Verified facts: dated claims with a direct link to the original source.
  • Interpretation: what those facts appear to mean for this decision.
  • Assumptions: things the recommendation depends on but the research has not proved.
  • Unknowns: information that only you, a supplier or a small test can provide.
  • Counter-case: the strongest credible reason to choose the opposite option.

This does two useful things. It makes the report easier to challenge, and it stops a neat citation from giving an unsupported recommendation borrowed authority.

Give sources an order

Tell the tool which evidence should win when sources disagree. Start with official documentation, contracts, filings, published prices or your own records. Then use reputable independent analysis. Treat social posts, summaries and search snippets as leads to check, not evidence to rely on. Include an "as of" date so old information does not quietly become current advice.

Use this decision brief

Use Deep Research if your tool has it. Before it starts, inspect the proposed plan and remove any sources or lines of enquiry that do not help the decision.

Copy-paste prompt

I need to decide: [one specific choice].

I need to decide by: [date].

The result must respect these constraints: [budget, policy, time, location or capacity].

Prioritise these sources: [official sites, uploaded files or named records]. Use information current as of [date].

Return a one-page decision brief with a recommendation, five decisive facts with direct citations, assumptions, unknowns, the strongest counter-case and what new evidence would change the recommendation.

End with the smallest useful next test. Do not hide uncertainty by giving me more background.

Read the last two sections first

When the report arrives, start with the counter-case and the evidence that would change the recommendation. If those sections are vague, the tool has probably argued for its first answer rather than tested it. Send it back before you spend time reading the rest.

The best use of AI research is not to make uncertainty disappear. It is to show you which uncertainty matters, what is already known and what small piece of evidence you should gather next. That gives you something much more valuable than a comprehensive report: a decision you can explain.

Also in this issue

The news I picked out on 29 September 2026. Some of it may have moved on since.

  • Working with knowledgeGoogle lets you talk to Gmail, Docs and Keep

    Google is rolling out voice modes for Gmail, Docs and Keep. The linked explanation shows how Gmail Live, Docs Live and Keep Live work, so you can judge whether voice notes fit your next brief or meeting record.

  • Tools worth knowingFrankenWhisper transcribes Mac recordings without uploading them

    FrankenWhisper is a free Mac app that turns recordings, audio files and videos into text on your device. Its listing explains speaker labels, export formats and caption tools, so you can judge whether it suits interviews or creative reviews.

  • Worth being careful aboutXero tightens controls after influencer claims Claude cut fees

    Xero says it is tightening controls after an influencer promoted connecting its accounting platform to Claude and claimed it cut fees. The linked note explains what happened, so read it before linking client billing data to an AI tool.