How to Plan a Source-Grounded AI Research Project

How to Plan a Source-Grounded AI Research Project

Plan an AI research project by defining the decision, audience, scope, evidence classes, exclusions, privacy boundary, deliverable, and stopping rule before the tool searches.

A long prompt is not necessarily a research plan. The plan should explain what evidence would make the work useful and what the system is not allowed to do.

Start with the decision

Write the decision the research will inform and who owns it.

“Research the electric-vehicle market” is a topic. “Help the product team decide whether to interview fleet operators in New York during the next discovery cycle” is a decision context.

The research may end with no recommendation. It should still know which uncertainty matters.

flowchart LR
    A["Decision and audience"] --> B["Claims and evidence needed"]
    B --> C["Scope, exclusions, and privacy"]
    C --> D["Research plan"]
    D --> E["Discovery and source ledger"]
    E --> F["Verification and stopping rule"]

Break the question into claims

List the material claims the final brief may need.

A market question may require market definition, customer segments, need, alternatives, regulation, economics, distribution, competitors, adoption, constraints, and evidence gaps. A product-history question needs dates, releases, official documentation, contemporary reporting, and later correction.

Do not ask the tool to discover every possible fact. Ask it to find evidence for defined claim classes and to surface missing or conflicting evidence.

Specify source classes

Name the preferred sources for each claim.

Product behavior should begin with current official documentation. Law should begin with primary legal authority and qualified review. Company claims should remain attributed to the company. Research claims should point to the paper or authoritative repository. Historical claims should use contemporary records where possible.

The Library of Congress source guide distinguishes first-hand records from later interpretation. Source type helps define what a document can establish, but it does not guarantee accuracy.

Require independent sources when a claim depends on a vendor, interested party, or disputed event.

Define time, place, and jurisdiction

State the evidence cutoff, search date, geography, jurisdiction, language, and historical window.

“Current” is not a date. A current product explainer should record the exact verification date and a refresh trigger.

For a global topic, specify which markets matter. For regulation, name the legal jurisdiction and effective date. For historical comparison, keep facts from different periods separate.

Write the exclusions

Exclusions prevent a broad agent from producing an impressive answer to the wrong question.

State topics, geographies, customer types, technologies, time periods, and recommendations that are outside scope. Also name prohibited sources such as unsourced aggregators or AI-generated pages when they cannot support the claim.

If a source is paywalled or inaccessible, require the report to say so rather than infer its contents from a citation.

Set the privacy and tool boundary

List the material the system may receive, the approved account, connected sources, storage, retention, logging, human-review conditions, export path, and who can access the result.

Google’s Gemini Apps Privacy Hub is the current starting point for Gemini app data handling. The Gemini API, Workspace, enterprise controls, and third-party tools require their own review.

Do not include customer, employee, guest, medical, legal, financial, security, or confidential business data without an approved basis and controls.

Define the deliverable

Specify the form the reviewer needs.

A useful research report can include an answer-first summary, claim-evidence table, source ledger, conflicting evidence, missing evidence, assumptions, dates, jurisdictions, glossary, and questions for domain review.

Ask for passages or page locations that help the reviewer open the source. Do not request long copyrighted quotations.

The final decision brief should remain separate. Research findings and recommendations are different artifacts.

Add the stopping rule

Research can expand without limit. Define when discovery is sufficient to enter verification.

A stopping rule may require a primary source for each material fact, two independent sources for a disputed claim, representation of major competing explanations, no unresolved critical citation, and explicit documentation of missing evidence.

The rule does not declare truth. It prevents the system from confusing more pages with better coverage.

Review the plan before execution

Google’s current Gemini Apps workflow and preview API agent both support planning concepts, though their controls differ.

Inspect the proposed plan for missing claim classes, weak source preferences, hidden assumptions, privacy problems, and scope drift. Revise it before the expensive search.

The approved commission should be preserved with the report so a reviewer can see what the system was asked to do.

About this guide

This guide was developed from Venture Step E049, current Google documentation, and source-evaluation guidance with AI assistance. It does not replace domain, legal, privacy, security, records, procurement, research-ethics, or professional review.