When AI Research Keeps Growing, Write Down the One Decision It Must Support

Larry

Maya asks AI to compare three workshop venues near Taipei for a team of six, with just three conditions: budget, travel time, and wheelchair access. Half an hour later, the AI’s report has added accommodation, catering, rain plans, sightseeing routes, and seventeen more candidate venues. The report is thick, yet it has no table that lets her pick one of the original three.

There’s no shortage of data. What’s missing is any statement of when the research counts as done. AI is very good at finding adjacent questions; as long as the instruction says “be thorough” or “find as much as possible,” every new source can sprout a few more branches. In the end it looks like a lot of work got done, while the decision it was meant to support gets buried. Research is finished when it can support a specific decision, not when it has enough sources.

Source count is not progress

The number of sources is easy to count, so it’s easy to treat as progress. But twenty articles about venues may never answer “can a wheelchair get from the entrance to the main room?”, while a single official price list may be enough to confirm cost.

AI research tools already hint at this. Google Gemini’s Deep Research drafts a research plan before it starts, which users can change with “Edit plan,” and lets users choose which sources to include. The direction and scope of research were always meant to be set by a person before it begins.

Write down who the research helps decide what

The research topic shouldn’t be “research places near Taipei suitable for team events.” It should be “help Maya choose, by Wednesday, one of three venues that have already quoted, with total rent within budget, no more than 50 minutes from Taipei Main Station, and wheelchair access from the entrance to the main activity space.”

That sentence fixes four things at once: who decides, by when, among which options, and which conditions are non-negotiable. Accommodation and food may well matter, but if they won’t change this choice of venue, they stay out of the main comparison for now. Important and in scope right now are two different things.

Government evidence reviews work the same way. A 2025 rapid evidence assessment by the UK Department for Culture, Media and Sport set out its research questions, search strategy, and inclusion and exclusion criteria before screening material, built a longlist, then filtered it to a shortlist against those criteria. Everyday research doesn’t need to copy the procedure, but it can borrow the core: decide what counts as evidence before starting to collect.

Four things that let research wrap up

First, support only one decision. If one report has to choose a venue, book hotels, arrange transport, and design activities, it’s really four jobs.

Second, use the same evidence for every option. The venue example needs only a few fields: price including tax and the quote date, door-to-door travel time, specific evidence of accessible entrances and restrooms, and cancellation terms. Don’t let venue A have twelve fields while venue B has only marketing copy, then use page count to pretend they’re comparable.

Third, agree on stop conditions in advance. “Stop after ten sources” isn’t a good one. A practical version: every required field for all three options has a source you can open and check; conflicting numbers have been flagged; missing data comes with a note on whether it would change the choice; and if another round of searching only repeats what’s known, stop.

Fourth, put new findings into a deferred list. For each deferred item, note why, when it should reopen, and who picks it up, such as “once a venue is chosen and an overnight stay is needed, compare nearby hotels; handled by the office manager.” Questions moved out don’t disappear; they just stop blocking this delivery.

A stop rule doesn’t mean the truth is all in

Stop conditions are a rule for bringing work to a close, not a guarantee that the truth has been exhausted. Research can be complete with gaps remaining, as long as the gaps’ effect on the decision is visible. When two reliable sources conflict, don’t pick whichever number looks tidier to deliver a neat conclusion. Put both values and their dates side by side and say whether the difference changes the ranking. If it does, mark it “needs human confirmation before deciding.”

Also, “search more” isn’t the only way to verify. If the venue’s website and a booking platform show different prices, the next step may be having someone ask for a quote directly, not having AI read twenty more blog posts. The narrower the research scope, the less likely it is to read web pages containing malicious instructions. OWASP calls this indirect prompt injection: an attacker hides instructions in external content the AI will read, such as a web page or document. Research AI should treat web pages as material to analyze, not commands that can rewrite its task.

As a next step, take a research task that’s ballooning and rewrite its instructions into five fields: the decision, a cap on options, required evidence, stop conditions, and a deferred list. Ask AI to deliver only the comparison table, source links, conflicts and gaps, and the deferred list. If you’re worried the research sources themselves are contaminated, see A Source Cited in an AI Research Report Shows Where It Read Something, Not That It’s True If the research has turned into a long, multi-round automated task, for managing its cost see Long AI Tasks Spend Most of Their Budget on Retries and Rereading Context.

AI handoff card

I have an AI research task that keeps growing and never wraps up, and I want to narrow it into something deliverable.
The method comes from this article: https://boosterminiclass.com/en/posts/ai-research-stop-rule-before-scope-creep/
Its view: research is done when it can support a specific decision, not when it has enough sources. Write down that decision (who, when, among which options), the same set of evidence every option needs, and stop conditions, and put other adjacent questions into a deferred list.

Ask me 3 questions first:
1. Who will make what decision, and by when, based on this research?
2. Which options are currently on the table, and which conditions are non-negotiable?
3. Which questions has the research expanded into that I never meant to cover?

Based on my answers, give me:
- A one-sentence research goal: who, when, among which options, by which conditions;
- The same set of evidence fields every option must have;
- Stop conditions: what counts as deliverable, and which gaps must stop the work first;
- A deferred list, with a reopening condition and an owner for each item;
- A one-line verdict: deliverable now / missing one piece of evidence / should stop first.
Don't run new searches, change files, or make the final choice for me; the decision belongs to the person responsible.