Free workflow kit

AI deep research workflow: from question to verified findings

A practical 11-stage workflow for using AI to conduct structured, evidence-based research — from defining the question and finding sources to challenging conclusions and verifying the final findings.

Define
Plan
Discover
Evaluate
Extract
Compare
Analyze
Synthesize
Challenge
Verify
Deliver
Read the workflow first
  • Works with ChatGPT, Claude, Gemini and other capable AI tools
  • Built for deep research, reports, market research, strategy, content and client work
  • Covers source tracking, evidence mapping, contradiction analysis, findings and verification

The idea

What is an AI deep research workflow?

Deep research is not: ask AI a question, get a long answer, trust it. That produces something confident and unfalsifiable — you cannot tell which part came from a source, which part is inference, and which part is neither.

A stronger process separates the work into stages, so sources, evidence, assumptions, contradictions and conclusions can each be inspected on their own.

The chain that has to stay intact

Research question
Sources
Evidence
Contradictions
Findings
Verification
Deliverable

AI can accelerate searching, organising, extracting, comparing, analysing and synthesising information. What the process has to preserve is source traceability — and a clear line between AI-generated interpretation and the underlying evidence.

The workflow

The 11-stage AI deep research workflow.

Each stage has one job and one output. The output is what the next stage consumes — which is what keeps the chain from breaking.

  1. 01

    Define

    Clarify the research question, objective, audience, scope and constraints — and what the research actually has to establish before it can be called finished.

    Output: Defined research question and boundaries

    Recorded in: Research Planner

  2. 02

    Plan

    Break the main question into sub-questions and decide what kind of evidence each one requires. Deciding this before you search is what stops the answer being shaped by whatever turns up first.

    Output: Research plan and evidence requirements

    Recorded in: Research Planner

  3. 03

    Discover

    Search systematically for relevant sources, and deliberately look beyond the evidence that supports what you already expect to find.

    Output: Candidate source set

    Recorded in: Source Tracker

  4. 04

    Evaluate

    Assess each source for relevance, credibility, methodology, recency, stated limitations and potential conflicts of interest.

    Output: Evaluated source set

    Recorded in: Source Tracker

  5. 05

    Extract

    Capture the evidence that actually matters — statistics, findings, observations, quotations, the conditions they hold under, and the caveats attached to them.

    Output: Structured evidence records

    Recorded in: Evidence Matrix

  6. 06

    Compare

    Cross-check evidence across sources. Identify agreement, disagreement, and the cases where several sources ultimately trace back to the same underlying evidence.

    Output: Evidence comparison and identified contradictions

    Recorded in: Evidence Matrix + Contradiction Tracker

  7. 07

    Analyze

    Look for patterns, relationships, explanations, assumptions, anomalies, trade-offs and implications. Keep what the evidence says separate from your interpretation of it.

    Output: Analytical observations

    Recorded in: Findings

  8. 08

    Synthesize

    Combine individual pieces of evidence into higher-level findings, while preserving the uncertainty and limitations that came with them.

    Output: Evidence-backed findings

    Recorded in: Findings

  9. 09

    Challenge

    Actively test your emerging conclusions against counter-evidence, alternative explanations, anomalies and your own confirmation bias.

    Output: Stress-tested findings

    Recorded in: Contradiction Tracker + Findings

  10. 10

    Verify

    Check that important claims have sufficient evidence, credible sources, appropriate corroboration, accurate citations, current information, and wording that matches the strength of the evidence.

    Output: Verified or appropriately qualified findings

    Recorded in: Verification

  11. 11

    Deliver

    Organise verified findings around what the audience needs to understand, decide or do — not around the order you happened to research them in.

    Output: Research ready to become a report, article, presentation, strategy or recommendation

    Recorded in: Deliverable Planner

Depth

You don’t always need all 11 stages.

Rigour should match stakes. A quick background question does not deserve the same process as research behind an important business decision.

Quick research

Define → Discover → Evaluate → Synthesize → Verify

Background questions and low-stakes answers.

Deep research

All 11 stages

Reports, strategy and anything a decision rests on.

Verify existing research

Evaluate → Compare → Challenge → Verify

Pressure-testing work that already exists.

AI’s role

Where AI helps at each stage.

AI should help you operate the research process. It should not become the evidence itself.

DefineRefine the question and surface ambiguity in how it is framed
PlanGenerate sub-questions and evidence requirements
DiscoverBuild search strategies and locate candidate sources
EvaluateAssess methodology, relevance and stated limitations
ExtractStructure findings and evidence into consistent records
CompareCompare sources and diagnose where they disagree
AnalyzeSurface patterns, anomalies and unstated assumptions
SynthesizeCombine evidence into structured findings
ChallengeGenerate counterarguments and test conclusions
VerifyAudit claims, citations and evidence sufficiency
DeliverStructure the research for its intended audience

Ground rules

Four rules for better AI research.

01

AI ≠ evidence

AI can find, organise, analyse and synthesise information. Important factual claims should still trace back to an identifiable source, not to an AI response.

02

Preserve uncertainty

Don’t let synthesis quietly turn "may", "suggests" or "is associated with" into certainty or causation.

03

Challenge your conclusion

Go looking for credible evidence that could weaken or change what you are about to conclude, before you finalise it.

04

Match rigour to stakes

Deepen source evaluation, corroboration and verification as the consequences of being wrong get larger.

In practice

One question, all the way through.

Research question

How is generative AI affecting productivity in small marketing teams?

  1. DefinePin down what counts as a "small marketing team", the geography, how productivity is being measured, and over what timeframe.
  2. PlanIdentify the sub-questions: adoption rates, time savings, output quality, cost, and the limits of each.
  3. DiscoverFind surveys, academic studies, industry reports, company data and credible practitioner accounts.
  4. EvaluateAssess methodology, sample size, recency, who funded the work, and how well it maps to small teams specifically.
  5. ExtractRecord the useful statistics and findings together with the caveats attached to them.
  6. CompareInvestigate why different studies report very different productivity gains.
  7. AnalyzeLook for the conditions under which productivity improves — and the ones where it declines.
  8. SynthesizeBuild findings that hold across the evidence base rather than resting on the single most quotable study.
  9. ChallengeDeliberately hunt for evidence that AI creates rework or lowers quality.
  10. VerifyTrace the major claims back to their original sources and check the wording still matches the evidence.
  11. DeliverTurn the verified findings into the report or recommendation the audience actually asked for.

The kit

Put the workflow into practice.

Everything above is on this page for free. The kit solves the other problem: actually running a deep research project without spreading it across scattered chats, notes, bookmarks and documents.

Start HereThe workflow at a glance, the four rules and three depth paths
Research PlannerDefine the question, scope, objective and success criteria
Research WorkflowTrack all 11 stages, their outputs and where each one is recorded
Source TrackerEvaluate and organise sources, with credibility and recency
Evidence MatrixConnect each piece of evidence to a question, a claim and a source
Contradiction TrackerInvestigate conflicting evidence and diagnose why it conflicts
FindingsBuild evidence-backed conclusions with confidence levels and caveats
VerificationRun the final research QA before anything ships
Deliverable PlannerTurn verified findings into a structured output

ai-research-workflow-kit.xlsx

XLSX · 9 sheets · 11 stages

Enter your email and we’ll send the workbook straight to your inbox. Free, no account, no card.

We may email you occasionally for feedback, or to tell you about new Cognardo resources. Unsubscribe any time.

Better together

Pair it with the Research Prompt Pack.

Two different jobs. The Workflow Kit manages how you run and organise the research. The Prompt Pack gives you 66 ready-to-use prompts for the individual tasks inside it.

Prompt Pack

What to ask AI

66 prompts mapped to the same 11 stages — question framing, source screening, evidence extraction, contradiction diagnosis, the final trust audit.

66 AI research prompts

Workflow Kit

How to run the project

Which stage you’re in, which sources you have, what evidence supports which claim, what contradicts it, and what still needs verifying before it ships.

You’re on that page

The third layer

Run the stages inside your LLM.

The kit tells you which stage you're in and the Prompt Pack tells you what to ask. Workbench is where you run it — our browser extension keeps your prompt library organised, fills the variables in a form, and executes without leaving ChatGPT, Claude or Gemini.

Explore Workbench

FAQ

Questions, answered.

What is an AI deep research workflow?+

It is a staged process for doing research with AI rather than delegating research to AI. Instead of one long prompt and one long answer, the work is split into eleven stages — define, plan, discover, evaluate, extract, compare, analyze, synthesize, challenge, verify, deliver — so sources, evidence, contradictions and conclusions stay separately inspectable.

Can I use this workflow with ChatGPT?+

Yes. The workflow is model-agnostic; it describes what to do at each stage, not which tool to do it in. ChatGPT handles every stage well, particularly discovery, extraction and synthesis.

Does it work with Claude and Gemini?+

Yes. Claude tends to be stronger on analysis, challenge and verification work; Gemini is useful for discovery and search. The workflow is the same whichever you use, and nothing in the kit is tied to one model.

Do I need to complete all 11 stages?+

No. Rigour should match stakes. A quick background question might only need Define → Discover → Evaluate → Synthesize → Verify. Research supporting an important decision deserves all eleven. Checking someone else’s work is mostly Evaluate → Compare → Challenge → Verify.

How is the Workflow Kit different from the AI Research Prompt Pack?+

The Prompt Pack is what to ask AI — 66 prompts for individual research tasks. The Workflow Kit is how to run the project: which stage you are in, which sources you have, what evidence supports which claim, and what still needs verifying. They work well together and are useful separately.

Can I use this workflow for academic research?+

It works well for literature review, source evaluation and evidence synthesis, and the provenance and verification stages map closely to academic practice. It is not a substitute for your institution’s methodology requirements or its rules on AI use — check those first.

Can I use it for market, competitor or business research?+

That is where it earns its keep. Market and competitor research leans heavily on vendor reports, sponsored studies and practitioner claims, so source evaluation, contradiction tracking and verification matter more here than almost anywhere else.