AI Research Workflow: Keep Only Evidence That Changes the Decision
An AI research workflow reviewed on 2026-09-03 had one hard constraint: no verified performance, cost, revenue, user, conversion, or timing evidence was available. The honest answer is therefore a method, not a claim that the method produces better results. Write one decision question, find one source that supports it, one that challenges it, and one that exposes what remains unknown. Then keep only the sources that could change the decision. The finished artifact is a source map, not a long bookmark list.
Start with the evidence boundary
Research becomes difficult when gathering begins before the decision is clear. Search results accumulate, summaries grow, and the original question quietly disappears.
For this playbook, the reviewed conditions are deliberately narrow:
| Evidence field | Recorded condition |
|---|---|
| Reviewed date | 2026-09-03 |
| Audience | Beginners conducting free web research |
| Scope | One question and one decision |
| Required evidence | Support, opposition, and uncertainty |
| Verified performance evidence | None supplied |
| Decision boundary | No claims about speed, accuracy, cost, traffic, revenue, users, or conversion |
That last row matters. This article does not prove that a source map saves time or improves decisions. It shows how to build one while keeping unsupported claims out of the result.
The three-line answer is:
Define the decision before searching.
Collect evidence by role, not by volume.
Remove any source that cannot affect the conclusion.
A source earns its place by changing the decision, not by looking authoritative in a tab.
Turn the topic into a decision question
A broad topic such as “AI research” is not yet researchable. It names an area but does not say what must be decided.
Rewrite the topic as a question with a visible choice. A useful decision question contains:
- The action under consideration
- The condition under which it would be chosen
- The evidence that could rule it out
- The uncertainty that would prevent a confident answer
Use this copyable frame:
Should I [take an action] for [specific situation], given [important constraint]?
I will choose it if [supporting condition] is credible.
I will reject it if [opposing condition] is credible.
I will keep the decision open if [critical unknown] remains unresolved.
This framing prevents a common failure: asking an AI system to “research everything” and receiving a polished tour of the topic. A tour may be informative, but it does not identify which evidence controls the decision.
Before searching, write a provisional answer in one sentence. It can be wrong. Its purpose is to give the evidence something to challenge.
Give every source a job
The source map needs three evidence roles.
Supporting evidence strengthens the provisional answer. It should establish a relevant fact, condition, mechanism, or documented observation.
Opposing evidence weakens the answer or supports a credible alternative. It is not a token objection added for balance. It must create a real reason to reconsider.
Unconfirmed evidence identifies a missing fact that matters. This category is not a drawer for weak sources. It records what could not be established and explains why the uncertainty affects the decision.
Search each role separately. A single open-ended query tends to reward repeated versions of the same position. Role-based searches force the research workflow to look for disagreement and missing information.
Useful query patterns include:
[decision question] evidence[provisional answer] limitations[provisional answer] criticism[alternative] comparison[critical claim] official documentation[critical claim] independent analysis[critical unknown] data
AI can help extract claims, summarize passages, or propose counterarguments. It should not silently decide that a source is credible. Open the original page. Check who produced it, what it actually supports, when it was published or updated, and whether its conditions match the question.
Agreement is not corroboration when every result repeats the same unsupported claim.
Build the source map while reading
Do not take general notes first and organize them later. Add each candidate directly to the map.
| Role | Source and date | Exact claim supported | Conditions or scope | Effect on decision | Keep? |
|---|---|---|---|---|---|
| Support | Strengthens because… | Yes / No | |||
| Opposition | Weakens because… | Yes / No | |||
| Unconfirmed | Blocks confidence because… | Yes / No |
The “exact claim supported” column should contain your paraphrase of what the source establishes. It should not contain the conclusion you hope to reach.
The “conditions or scope” column catches hidden mismatches. Evidence about a large organization may not answer a solo operator’s question. A product demonstration may show that something can work without showing that it works reliably. A recent page may still rely on older underlying data.
The decisive column is “effect on decision.” Complete it with a causal sentence:
If this evidence is credible and applicable, the decision changes because…
If that sentence cannot be completed, the source is probably background reading. Remove it from the final map, even if it was interesting.
Prune links without losing uncertainty
A crowded map creates the appearance of diligence while making review harder. Pruning is part of the research, not a cosmetic cleanup.
Remove a source when it:
- Repeats a claim already supported by stronger evidence
- Discusses the topic without addressing the decision
- Cannot be traced to an original claim or artifact
- Uses conditions that do not match the stated scope
- Adds detail but cannot change confidence or choice
- Makes a stronger claim than its evidence supports
Keep uncertainty when it is decision-relevant. “I could not verify this” is useful when the missing fact could reverse the recommendation. It is not useful when the missing detail would have no practical effect.
This is also where attractive AI summaries can fail. A fluent synthesis may blur the difference between a source’s statement and an inference drawn from it. Preserve that distinction in the map:
- Observed evidence: what the source directly establishes
- Inference: what follows only when evidence and conditions are combined
- Recommendation: what action makes sense within the remaining uncertainty
Uncertainty belongs in the artifact when resolving it could reverse the choice.
Make the decision visible
The final answer should be traceable without reopening every link. Use this decision block:
Decision: Choose / reject / defer.
Observed evidence: [What the retained sources directly establish.]
Opposing evidence: [What materially weakens the choice.]
Unconfirmed: [What remains unresolved.]
Inference: [What the combined evidence suggests.]
Stop rule: Reopen the decision if [specific missing or contradictory evidence] appears.
The explicit decision for this playbook is: use a source map as the required output of beginner AI research, but do not claim that it improves performance without verified comparative evidence. Its immediate value is inspectability. A reviewer can see why each source survived and where the conclusion remains exposed.
Copy the final review checklist
Before finishing, verify the artifact:
- The question asks for one decision, not a topic summary.
- A provisional answer was written before collection.
- The map contains support, opposition, and uncertainty.
- Every retained source has a visible date or an explicit missing-date note.
- Every claim is narrower than or equal to its source.
- Source conditions match the decision conditions.
- Observation, inference, and recommendation are separated.
- Each retained source could change confidence or choice.
- Repeated, decorative, and untraceable links were removed.
- The final answer says choose, reject, or defer.
- A stop rule explains when to revisit the decision.
- Missing performance evidence is disclosed rather than estimated.
The method still has limits. A small map can miss important evidence. Free search can surface biased, stale, duplicated, or inaccessible material. Three evidence roles do not guarantee truth, and AI assistance does not replace checking the original source. The map is a reviewable record of reasoning—not proof that the research is complete.
Related build logs
- For AI Customer Service Beginners, Keep the Workflow Draft-Only
- An AI Review Workflow for Beginners: Draft, Approve, or Stop
For a beginner AI research workflow, keep one supporting source, one opposing source, and one critical uncertainty—then delete every link that cannot change the decision.