AI Research Report Template: 12 Fields a Reviewer Can Check

Build a research record before writing the narrative

The research question determines what evidence is relevant. A claim is a statement that a reviewer can examine, such as a measured count or a product's documented eligibility condition. An implication explains why a finding may matter. A recommendation proposes an action and requires a responsible owner. These are related, but they are not interchangeable.

Create a small source register and a claim ledger. A source record identifies the document, version, publisher, retrieval status and relevant location. A claim record identifies the exact wording, support, transformation and review result. One source may support several claims; several links may ultimately depend on one source. Preserve those relationships instead of counting every link as independent evidence.

The 2024 paper On the Evaluation of Machine-Generated Reports examines report coverage and supporting citations. Its evaluation framework assumes the document collection is truthful; independently establishing source truth is outside that framework. Citation support and source reliability therefore need separate consideration. Paper and publication record.

Record only material needed for the approved purpose. A source locator can be a page and table number, a section heading, a timestamp in an authorized recording, or a stable row ID. A whole document attached without a locator may be technically available but still costly to verify. Keep the relevant qualifications with an excerpt rather than stripping them away.

Twelve fields in an AI research report template

Use these fields as a working contract, not twelve paragraphs of filler. The question, scope and source rules come before collection. Draft the executive answer after examining the evidence, then put it near the front of the reader-facing report. The order in which a report is read does not have to be the order in which it is researched.

1. Research question

Write one answerable, neutral question with the subject, comparison, outcome and relevant boundaries. “Should this forty-person team pilot an inline weekly briefing?” is more useful than “Prove that AI reports improve productivity.” The first leaves room for an unfavorable or incomplete answer; the second embeds the desired conclusion.

Define terms that affect inclusion or arithmetic. Does a respondent mean anyone who opened the form, or one eligible person with a valid completed response? Does a report mean an attachment, an inline summary or a generated analysis with sources? If the definition changes during research, document the change and reconsider the earlier evidence.

List adjacent questions that will not be answered. A preference survey about format does not establish writing-time savings, correctness, accessibility or willingness to pay. Those questions may become follow-up work, but they should not enter the conclusion through an ambiguous heading.

2. Decision and audience

Name the decision owner, intended readers, deadline and consequence of being wrong. A researcher preparing a reversible trial needs a different evidence threshold from a team approving a company-wide system that processes confidential material. Neither threshold can be inferred from the desired report length.

Specify what the owner will receive: an executive answer, a comparison, open questions and a proposed next step. Senior readers may need fewer details on the first page, but they still need the caveat that changes the decision. Do not hide a weak sample or an inaccessible source in an appendix simply to make the summary cleaner.

Record the distinction between research acceptance and action approval. A reviewer may accept a report as an accurate description of limited evidence while the business owner declines to act. A completed document is not permission to buy, publish, contact people or change a production process.

3. Scope and as-of date

Identify the entities, products, versions, populations, regions, languages and periods included. Record a final evidence cutoff and the retrieval date of important sources. A report published recently can describe an old sample; a recently retrieved page can contain an old statement. Publication, observation, retrieval and effective dates have different meanings.

Name exclusions and access limits. If only public English-language sources were searched, say so. If a subscription database or earlier product version was unavailable, do not imply it was inspected. A narrow but accurately described search is more useful than a claim of global coverage with no record behind it.

Decide what makes the report stale. A revised eligibility rule, corrected dataset or changed decision scope may require immediate review, while an unchanged background definition may not. The as-of date is a boundary on what was checked, not a promise that everything relevant had been discovered by then.

4. Source inclusion and exclusion rules

Set rules before selecting evidence. Consider relevance, original publication, identifiable authorship, method, currency, independence, conflicts of interest and permitted use. Official documentation can establish what a vendor says it offers; it does not necessarily establish independent performance. Secondary analysis can provide context without becoming primary evidence for every number it repeats.

Keep exclusions with reasons such as wrong population, insufficient method, superseded version, inaccessible text or duplicate origin. Do not remove unfavorable findings by quietly tightening criteria after seeing them. If criteria genuinely need to change, record why and apply the change consistently to favorable and unfavorable sources.

When checking scholarly material, inspect corrections and publication status. Crossref's Crossmark can expose registered updates, corrections and retractions where implemented. Its presence is not itself a guarantee of quality or completeness. Crossmark documentation. A document without that control still needs an appropriate status check; absence alone does not make it invalid.

5. Method and search record

Save the sites or databases used, query text, date, filters, screening process and extraction method. Preserve result counts when they are actually available and relevant; do not invent counts to make a research log look complete. Record the queries that produced no useful evidence as well as the productive ones.

Search snippets and generated search answers are discovery aids. Open the underlying document before citing it as read, and verify the relevant passage rather than trusting a familiar title. A page that returns successfully may still be a login screen, an error page or a different document after a redirect.

For a formal systematic review, use an appropriate reporting standard and specialist methods. PRISMA provides reporting guidance, checklists and flow diagrams for systematic reviews. Filling this business template does not establish PRISMA compliance. PRISMA 2020 resources. Keep the report's method label proportional to the work actually performed.

6. Executive answer

Write the narrowest useful answer supported by the evidence, then add the strongest reason, the most important counterpoint and the remaining decision-relevant uncertainty. For the fictional example, “A small pilot is worth considering; preference among nonrespondents and any time benefit remain unknown” is more defensible than “The organization overwhelmingly wants AI reports.”

Link material sentences to claim IDs or supporting records. Replace vague confidence adjectives with reasons: the count reconciles to the register, the population is limited, and another important outcome was not measured. A model's confident wording is not a confidence interval or an independent review.

Write this section last in the research process. Then check it again after detailed editing. Summaries often retain an earlier strong claim even after the body has been corrected. The headline, deck, table and recommendation must not tell different versions of the same result.

7. Evidence table and claim ledger

Give each important claim a stable ID. Record the exact proposed sentence, source ID, supporting location, evidence type, population, units, period, calculations, limitations and reviewer disposition. Distinguish a direct observation from a derived number and from an interpretation. Keep enough context for another person to reproduce the distinction.

A citation passes more than one check: the document exists; the team accessed the relevant version; the locator leads to the evidence; and the evidence supports this wording at this strength. “Associated with” must not become “caused,” and “eight respondents” must not become “most employees.” Multiple citations do not repair a sentence if none supports its scope.

Quote only what is necessary and permitted. Paraphrases still require faithful attribution. If an AI tool extracts a table, verify headers, footnotes, merged cells, denominators and units against the original. A syntactically valid CSV can preserve an extraction mistake perfectly.

8. Conflicting and missing evidence

Put important disagreement beside the finding it limits. First ask whether sources use the same population, definition, intervention, comparison and time period. Different answers under different conditions are not automatically a contradiction. Do not average incompatible figures merely because they share a percentage sign.

Show whether a conflict is resolved, narrows the conclusion or remains open. If two pages repeat a single vendor release, record their common origin. They are two documents, not two independent studies. If the underlying methods document cannot be opened, the proper label is unverified method, not verified evidence because several summaries agree.

Missing evidence matters even when it is not opposing evidence. Twenty-eight people who did not respond to a survey have unknown preferences; silence is neither support nor rejection. Keep the gap visible in the answer and specify what further research could change the decision.

9. Analysis and implications

Show the reasoning between evidence and meaning. For every calculation, record the formula, numerator, denominator, exclusions, rounding and units. Explain whether figures describe a sample, the defined population or a hypothetical scenario. Recompute totals from the underlying records instead of copying a model's arithmetic.

Separate description from causation and transferability. A vendor's result in another setting does not estimate your team's savings without a justified comparison. A preference for inline text does not demonstrate that an AI-generated report is correct or faster to review. Keep alternative explanations visible, including selection effects and differences in task difficulty.

Use implications to explain relevance, not to skip a missing study. “The response suggests a format to test” is a bounded inference. “This proves the new workflow increases productivity” would require evidence that a preference question did not collect.

10. Uncertainty and limitations

Name the limits that affect action: access gaps, search coverage, selection bias, measurement error, small or unrepresentative samples, language exclusions, outdated versions, AI extraction mistakes and incomplete review. Tie each important limitation to the claim it affects and to a possible way of reducing uncertainty.

Use reasoned labels only when the team has defined what they mean. You might be confident that eight valid responses selected an option while remaining uncertain about the preference of the wider group. A single “high confidence” badge would hide that difference. Do not assign a numerical probability because the model can produce one.

Avoid treating the presence of a limitations section as permission to overstate the summary. If the caveat removes the basis for an action, change the recommendation. If additional evidence cannot be obtained before the deadline, present the uncertainty to the owner rather than filling the gap with a plausible answer.

11. Recommendations and accountable owner

Offer choices with evidence, prerequisites, cost or effort considerations, risks, reversibility and a validation step. Assign an owner and decision date. A recommendation should say what would make the team stop, revise or expand the action, not only what success might look like.

For a briefing-format pilot, specify the intended comparison and observable outcomes before starting: reviewer effort, factual corrections, source reopening and accessibility, for example. These are proposed measures, not claims that the new approach improves them. Participants, permitted data and review capacity need agreement before launch.

Keep high-impact and regulated actions under the organization's appropriate authority. Research can identify options, but cannot grant access to confidential records, determine legal compliance or substitute for medical, legal, financial, scientific or security expertise. Make the authorization boundary explicit in the handoff.

12. Reviewer and revision history

Record who prepared the report, what AI actually did, who checked sources, who reviewed domain conclusions and what has been approved. If a reviewer has not yet reviewed it, say pending rather than attaching an invented name, credential or approval date. Maintain the exact report version that was reviewed.

When a source, definition, calculation or conclusion changes, identify affected claim IDs and downstream summaries. Record the previous and replacement wording, reason, date and approval state. Notify people who received a materially misleading earlier version. A source update should not silently regenerate a report that retains an old approval label.

Preserve previous versions according to the approved retention policy, not indefinitely by default. The record should enable correction and accountability without becoming an unnecessary store of sensitive material. The current version must be distinguishable from a draft, a withdrawn report and an archived historical copy.

Run a six-step research and review workflow

1. Approve the brief. Set the question, decision owner, source scope, permitted tools and stopping conditions. Agree what would count as enough evidence for the intended decision.
2. Collect and register evidence. Open sources, record dates and locators, screen consistently and preserve permitted context. Mark inaccessible or incomplete records accurately.
3. Build and test the claim ledger. Extract claims before composing the executive narrative. Recalculate figures, identify common origins and attach contradictions or missing evidence.
4. Draft a bounded report. Use the accepted claim set and carry its qualifications into the summary. Label implications and recommendations; do not allow polished language to broaden findings.
5. Review claims and proposed use. A source reviewer checks support; an appropriate domain reviewer checks interpretation. The business owner separately decides whether and how to act.
6. Release a version and maintain it. Save the report, ledger, approved supporting records and review status together. Define update triggers and how corrections reach previous recipients.

Test the process before using it for a consequential report. Include an invented reference, an accessible but irrelevant document, an unsupported percentage, an old version, a contradictory footnote, a duplicated press release and an inaccessible underlying study. A document containing hostile instructions should remain evidence to inspect, not permission to change the research task.

Measure what was checked. The share of material factual claims with verified support needs an explicitly defined claim set and a review method. A report with many citations can still omit a decisive issue. Coverage of the necessary questions requires a separately defined reference list of those questions; link validity alone cannot establish completeness.

Worked example: correct a report that changes its denominator

This is a fictional training exercise, not a real employee survey or a measured AI result. A forty-person team is considering a pilot of an inline weekly briefing instead of an attachment. All forty eligible people receive one preference question; twelve valid, unique responses are recorded: eight prefer the inline format and four prefer the attachment. Twenty-eight do not respond.

The designed evidence packet contains four records. S1 is the aggregate response register. S2 is an invented vendor announcement claiming a 20% review-time reduction in its own setting without accessible methods in the packet. S3 republishes S2, so it is not independent corroboration. S4 is the referenced methods document, which is unavailable in this exercise. No real vendor or URL is implied by these identifiers.

The response rate is 12 ÷ 40 × 100 = 30%. Among respondents, the inline preference is 8 ÷ 12 × 100 = 66.7%, rounded to one decimal. Among all invitees, the observed favorable responses represent 8 ÷ 40 × 100 = 20%. The final number is not an estimate that only 20% actually prefer the new format: the other twenty-eight preferences are unknown.

If all nonrespondents preferred the attachment, the inline preference across the group would be 20%. If all preferred inline, it would be 36 ÷ 40 × 100 = 90%. These are logical extremes given the fictional counts, not a confidence interval, a forecast or a recommended statistical estimator. The example shows why a population-majority conclusion is not established by the response subset.

There are five proposed factual claims, C1–C5. Three are unsupported at their stated scope, so the unsupported share in this deliberately flawed draft is 3 ÷ 5 × 100 = 60%. C6 is assessed as a recommendation and is not added to that factual-claim denominator. This is an exercise result, not an accuracy score for AI tools or OpenMax.

A corrected executive answer could read: “Eight of twelve respondents preferred the inline format; twelve of forty invitees replied. The preferences of the remaining twenty-eight and any effect on review time are unknown. Consider an owner-approved, reversible pilot with pre-agreed quality and effort measures before deciding on a wider change.” The vendor's claimed percentage is not needed to justify that limited next step.

The filled reasoning also tells the reviewer what to reject. Do not turn nonresponse into opposition, count a republication as a second study, cite S4 as read, or label a proposed pilot as approved. If a later source corrects a result, update the claim ledger and every summary that depends on it before releasing a new version.

Choose a document, reference-manager or agent-assisted approach

Manual documents and spreadsheets suit a narrow question with a manageable evidence set. Keep the twelve-field template, source IDs and calculations together. The main discipline is maintaining links between claim wording and evidence as edits accumulate. A short report does not justify skipping that record.

Reference-manager features help organize source material and notes. Zotero supports notes attached to individual items as well as standalone notes, with tagging and relationships. Those features can support record organization; they do not decide whether a source entails a claim. Zotero notes documentation.

Deterministic automation can check file existence, document IDs, duplicate identifiers, missing fields and arithmetic. It can also flag a broken URL, but a successful request is not proof that the source supports the sentence. A renamed table column or altered unit needs semantic review, not just a green validation result.

Agent-assisted research is worth evaluating when the scope is approved and repeated extraction, synthesis or reviewer coordination is the bottleneck. Bound the sources and tools, require traceable draft claims, and retain an explicit insufficient-evidence outcome. A model that generates a convincing report from unread documents is not an improvement over a slower but inspectable process.

As volume grows, separate ownership of collection health, source verification, domain interpretation and final decisions. The market research agent use case provides broader workflow context; the competitor monitoring guide covers recurring change detection. This page's deliverable remains the research report and its supporting record.

Evaluate OpenMax with one approved evidence packet

OpenMax describes its product as a human–agent collaboration platform with agent integration and team workflows. That positioning does not verify a specific research connector, an automatic citation-verification system or a complete research archive. OpenMax product overview.

Use the template to define a proposed evaluation rather than asserting those functions already exist. Start with a small approved packet containing a supported claim, a misleading denominator, a duplicate source and an inaccessible record. Ask the assistant to prepare a draft ledger and report, then have the reviewer inspect the source links and withheld claims.

Verify actual import and retrieval behavior, permissions, data retention, downloadable artifacts, source locators, logs and reviewer handoff in the intended environment. Keep missing capabilities visible. A product demonstration that writes a summary is not evidence that it preserves versions or prevents unsupported statements across all tasks.

If a document and reference manager already meet the need, keep them. If coordination is the problem, bring the approved packet and the decision brief to a bounded OpenMax workflow discussion. The next step is an inspectable report with a clear refusal to invent missing evidence, not an automatic claim of research authority.

Privacy, security and professional-review boundaries

Use only sources and data the organization is authorized to access and process for the purpose. Public availability does not settle copying, quotation, redistribution, personal-data use or retention rights. Qualified legal, privacy and security owners need to assess the actual context. This guide does not provide jurisdiction-specific legal clearance.

The National Institute of Standards and Technology describes its AI Risk Management Framework 1.0 as voluntary and intended to help organizations manage AI risks across contexts. It is not a certification of this template or OpenMax. NIST AI RMF 1.0.

Treat retrieved pages and files as untrusted content. OWASP describes indirect prompt injection through external material and recommends measures such as least privilege and human control of high-risk actions. These controls need implementation and testing; a written instruction alone does not establish immunity. OWASP prompt-injection guidance.

Do not upload confidential source packets to an unapproved service or send internal strategy to a website to obtain an answer. Limit access to the research records and their exported copies. Separate publication permission from permission to read. Do not bypass a login or paywall to make a report appear complete.

For scientific, medical, legal, financial, safety-critical or similarly consequential conclusions, assign appropriately qualified review and use methods suited to the field. A template can reveal missing review but cannot supply the missing qualification. Mark pending approval honestly and keep the report out of uses that require approval until that review is complete.

Frequently asked questions

How many sources should an AI research report contain?

There is no universal count. Use enough relevant, independent and accessible evidence to support the important claims and address meaningful disagreement. Several copies of one announcement are not several independent studies, and extra links cannot compensate for missing evidence on the decisive question.

Can a real citation still be wrong?

Yes. The source may exist but concern a different population, date, product version or outcome. Verify the cited location and exact wording, not only the URL. A reference that supports association does not automatically support causation or a prediction for another team.

Is this a template for scientific papers about AI?

No. It is for business research reports that may use AI assistance. Scientific manuscripts and systematic reviews require the relevant discipline's study design, reporting standards and review process. Do not relabel a business desk report as a systematic review because it has a method section.

What should we write when sources disagree or cannot be accessed?

Record the disagreement or access limit, compare scope and methods where possible, and narrow or withhold the conclusion. Identify what evidence would resolve the issue. Never claim to have read an unavailable source or conceal a credible counterpoint to make the recommendation sound stronger.

Can OpenMax provide the final expert approval?

Do not assume so. The organization must assign qualified people and decision authority. Any proposed OpenMax role in drafting or routing needs verification in the actual environment; it does not replace professional judgment, source checks or authorization for the report's intended use.

Sources, authorship and revision scope

Prepared by the OpenMax content team for OpenMax's own website. The product discussion is commercially connected to the publisher. Official references were checked on September 4, 2026; they support the particular research, documentation or risk statements next to their links, not a claim that their publishers reviewed this page or endorsed OpenMax.

The twelve fields, worksheet and fictional survey exercise are an editorial synthesis. No real employee survey, first-hand product-performance experiment or named professional review is claimed. Arithmetic in a designed example is not customer evidence. The page intentionally distinguishes citation support from source reliability and proposed next steps from approved action.

This revision expands the earlier outline with source and claim records, field-specific instructions, a completed denominator example, downloadable materials and explicit product-verification limits. It retains the original September 2, 2026 publication date with a September 4, 2026 content revision. Send corrections with the page URL and non-sensitive supporting material to contact@openmax.com; do not email confidential research records.