Practical guide
AI-assisted research plan: free template and guide
An AI-assisted research plan should make the decision, evidence boundary, method, verification work, and ownership visible before anyone starts collecting sources. A prompt such as “research this market” is not a plan. It leaves the model to choose the question, evidence standard, and stopping rule, which makes a polished answer hard to audit. This copyable template uses FIELD: frame the.
An AI-assisted research plan should make the decision, evidence boundary, method, verification work, and ownership visible before anyone starts collecting sources. A prompt such as “research this market” is not a plan. It leaves the model to choose the question, evidence standard, and stopping rule, which makes a polished answer hard to audit. This copyable template uses FIELD: frame the decision, inventory inputs, establish the method, log verification, and define delivery.
Source basis and limits
The UK Government's current AQuA Book treats design documentation as the place to record methodology, inputs, software, and planned verification and validation. It also says assurance should be proportionate to business risk and that analysis should preserve sources, assumptions, decisions, uncertainty, and changes throughout its lifecycle. NIST's Generative AI Profile recommends documenting provenance, source linkage, human oversight, and verification practices for generative-AI outputs. These sources support the planning controls below. They do not make a research result correct or replace a field-specific method, legal review, research ethics process, or accountable approval. AQuA Book NIST AI 600-1
Sources: [1] [2]
Copyable FIELD research plan
### Plan control
| Field | Entry | | --- | --- | | Research title, owner, version, and status | | | Commissioner and decision date | | | Intended audience and permitted use | | | Evidence cut-off | | | Researcher, independent reviewer, approver | | | Applicable policy, ethics, privacy, or professional review | |
### F - Frame the decision
Write the decision before the research question.
| Decision element | Plan entry | | --- | --- | | Decision to be supported | | | Primary question | | | Necessary subquestions | | | Population, geography, and period | | | Material exclusions | | | Precision or confidence needed | | | What this study cannot establish | |
Use a bounded decision statement: By [date], [owner] will decide [action] for [scope]. The research must distinguish [material alternatives] and expose [critical uncertainty].
Reject questions that merely ask for “insights.” Name what evidence could change the decision and what would not.
### I - Inventory inputs and access
| Input ID | Source class | Owner | Access route | Date/version | Expected use | Known concern | | --- | --- | --- | --- | --- | --- | --- | | | Primary / secondary / internal / generated | | | | | |
### E - Establish the method and stopping rule
| Method component | Planned procedure | | --- | --- | | Search venues and query families | | | Inclusion and exclusion rules | | | Screening and extraction process | | | Comparison or synthesis method | | | Treatment of missing and conflicting evidence | | | AI assistance allowed | | | Human checks required | | | Stopping rule | |
The stopping rule should depend on decision coverage, not the model's confidence. For example: stop only when every material subquestion has at least one eligible source, the team has performed a contrary-evidence search, unresolved conflicts are logged, and the reviewer can reproduce the material claims.
Define an AI work protocol. A model may propose queries, classify an approved source set, extract candidate passages with source IDs, or challenge a draft. It must not silently add evidence, invent citations, decide source eligibility, or convert missing data into a conclusion.
### L - Log verification and change control
| Check ID | Object checked | Procedure | Expected result | Evidence retained | Owner | Status | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | |
Plan checks for source identity, quotation accuracy, numerical reproduction, date and scope fit, claim-to-source linkage, contradiction handling, and the distinction between fact and inference. For consequential work, give an independent reviewer the source pack and claim ledger rather than only the final narrative.
Keep a decision log:
| Date | Change or judgment | Reason | Evidence affected | Approved by | | --- | --- | --- | --- | --- | | | | | | |
### D - Define delivery and maintenance
Specify the deliverable, audience, length, accessibility needs, citation format, decision table, uncertainty statement, appendices, and source pack. Define who may release it and what evidence proves approval.
For work that will be reused, add an expiry or review date, a correction route, and a trigger for reopening the research. Triggers can include a new regulation, a changed product version, a source withdrawal, or a material challenge to a headline claim.
- Define which source classes are eligible for each subquestion.
- Record inaccessible, restricted, translated, stale, or machine-generated inputs.
- Decide whether the AI tool may receive each source. Redact or exclude material when authorization is absent.
- Specify how source snapshots, search dates, and stable identifiers will be retained.
- Name contrary evidence that the search must actively seek.
Evidence and claim artifacts
Use a claim ledger for every statement that could alter the decision.
| Claim ID | Proposed claim | Source IDs | Direct observation or inference | Conflict or gap | Disposition | | --- | --- | --- | --- | --- | --- | | | | | | | Keep / narrow / remove / investigate |
Pair it with an unresolved-evidence register:
| Issue | Consequence if wrong | Current evidence | Next check | Owner | Decision deadline | | --- | --- | --- | --- | --- | --- | | | | | | | |
Worked example: fictional service-location study
A programme owner asks a researcher to identify the best city for a new support hub. The initial AI prompt requests a ranking based on labour cost, language availability, and operating conditions. The model produces five cities and cites several consultancy summaries.
FIELD reframes the task. The decision is whether to advance at most two locations into detailed diligence, not to choose a final site. The plan separates labour availability, cost, time-zone coverage, infrastructure, and regulatory constraints. It excludes property negotiation and tax advice. The evidence inventory requires current statistical-agency labour data, official infrastructure sources, internal support-volume records, and dated secondary market context. Consultancy rankings may orient the search but cannot alone support an advancement list.
The method records search families and comparison rules. Costs must use the same currency date and compensation definition. Language indicators must distinguish population fluency from occupation-specific availability. A contrary-evidence search looks for data outages, hiring concentration, and restrictive assumptions hidden in aggregate rankings. AI may normalize an approved table and flag missing cells, but a researcher reopens every source and reproduces material calculations.
The first synthesis favors City A. Verification shows that its low salary figure excludes employer contributions, while City B's figure includes them. The comparison is recalculated on a common basis. The claim ledger narrows “City A has the lowest operating cost” to “City A has the lowest comparable base salary among the three locations with complete official data.” The unresolved register notes that retention and office costs require diligence. The deliverable recommends advancing Cities A and B, explains the incomplete evidence, and expires after six months.
Practitioner QA before starting
Can the commissioner explain the decision in one sentence?
- Would two researchers apply the same inclusion rules?
- Are important missing sources visible before synthesis?
- Does every AI task have an approved input boundary and a human check?
- Can a reviewer reproduce material claims from retained evidence?
- Is there time for challenge and correction before the decision date?
- Does the plan say what the research cannot establish?
Version history
| Version | Date | Change | Evidence affected | Author | Reviewer | Approver | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | |
Two Prune boundary
Two Prune can assess whether a participant frames a research assignment, navigates supplied evidence, uses AI thoughtfully, verifies claims, handles uncertainty, and communicates a decision-useful result in a realistic simulation. It reports observed assessment evidence. It does not conduct the organization's research, approve a research plan, establish research validity, replace specialist review, or make employment decisions.
Sources
GOV.UK: The AQuA Book
- NIST AI 600-1: Generative AI Profile
Sources: [1] [2]
Sources
- 1.AQuA Book · UK Government
- 2.NIST AI 600-1 · doi.org