AI market research workflow with citations

Lab

AI Research and Customer Intelligence

Build a traceable evidence pipeline that separates source material, customer language, analyst inference, and unanswered questions.

DIFFICULTY
Beginner
ESTIMATED TIME
85 min
UPDATED
2026-08-20
COPY REVIEW
blader/humanizer
2 passes
On this page
  1. 01Working definition
  2. 02Field situation
  3. 03Worked example
  4. 04Build it, with checkpoints
  5. 05Hands-on lab
  6. 06Failure clinic
  7. 07Production boundary
  8. 08Sources and claim limits

Learning objectives

  • Frame research around a decision rather than a broad topic
  • Apply an explicit source policy and evidence schema
  • Preserve claim-level provenance through extraction and synthesis
  • Report disagreement, missing evidence, and confidence without false precision

Before you start

  • • A current product, positioning, or campaign decision
  • • Access to approved public and first-party research material

Working definition

Research and intelligence

An AI-assisted research workflow collects approved material, extracts evidence into a consistent ledger, and produces an analyst-reviewed synthesis whose important claims remain traceable. Retrieval and summarization are separate operations. The model may organize evidence, but it cannot turn repetition, search rank, or a plausible sentence into proof.

A long market summary is difficult to reuse if readers cannot inspect where a claim came from, when the source was accessed, or whether the sentence is observation or inference.

Customer language deserves special handling. A direct quote, a researcher paraphrase, and a model-generated theme are different evidence types and should not collapse into one field.

A decision-led brief narrows collection. It records what evidence could change the recommendation and prevents the research agent from searching indefinitely.

A durable source ledger can support content, sales, product, and leadership without forcing every team to rediscover the same material.

Field situation

Meridian Robotics APAC buyer research

Named synthetic scenario. Meridian Robotics and all research outputs are fictional teaching material, not client evidence.

Owner
You are the regional product marketer deciding whether to lead with maintenance cost or worker safety in a manufacturing campaign.
Decision
Recommend the lead message for one campaign, state the counterargument, and identify the next research action that would most reduce uncertainty.
Starting state
The team has six anonymized interview transcripts, two win-loss notes, product documentation, competitor pages, and public regulator guidance. Previous briefs mixed direct customer words with analyst assumptions and contained links that no longer supported the claims.
Expected outcome
A two-page decision brief backed by a normalized evidence ledger and a reviewer-readable claim map.

Constraints

  • • Interview identities and company details must remain redacted
  • • Only approved domains and uploaded first-party files may be collected
  • • The recommendation must distinguish regional evidence from global evidence
  • • Every pivotal claim needs a source pointer or an explicit evidence gap

Worked example

Comparing safety and maintenance narratives without inventing consensus

Evidence status: Named synthetic scenario

Meridian's six synthetic interviews contain frequent maintenance complaints, while two procurement notes emphasize safety documentation. Public regulatory guidance establishes obligations but does not show which message causes buyers to respond. A naive summary would count mentions and declare maintenance the winner.

The analyst extracts each passage with source ID, evidence type, speaker role, market, date, verbatim status, claim, and confidence. Themes are generated only after extraction. The brief recommends maintenance as the testable lead for operations managers, retains safety proof as a required supporting block, and proposes three additional procurement interviews before extending the choice across APAC.

The example produces a bounded message-test decision with a visible countercase. It does not claim that the recommended message will improve pipeline. The useful artifact is the chain from source passage to interpretation, recommendation, and next test.

Limits

All company material and conclusions are synthetic. Six interviews are not a representative market sample, mention count is not preference strength, and regulator guidance cannot validate campaign performance. Translation, interviewer bias, missing losses, and regional differences may change the interpretation.

Method

Build it, with checkpoints

Evidence lineage diagram linking approved sources to extracted passages, reviewed themes, disputed claims, and a decision brief

Field situation

Recommend the lead message for one campaign, state the counterargument, and identify the next research action that would most reduce uncertainty.

  1. 01Frame the decision and reversal condition
  2. 02Write and apply the source policy
  3. 03Extract evidence before themes

Acceptance checks

A brief that remains useful after the meeting because its evidence can be inspected, refreshed, and reused. Readers can see which statements came from sources, which came from analysis, where the material disagrees, and how uncertainty affects the recommended action.

Why this visualA provenance diagram can show the separation between source, extracted passage, theme, claim, and decision. That lineage is the core skill and benefits from spatial inspection.
  1. 01

    Frame the decision and reversal condition

    State who will act on the brief, the options under consideration, the deadline, and what finding would change the choice. Narrow geography, segment, and time horizon.

    CHECKPOINT · The question can be answered with a recommendation or an explicit insufficient-evidence result, not a general market overview.

  2. 02

    Write and apply the source policy

    List allowed source types, preferred primary material, recency rules, exclusions, consent constraints, and the treatment of paywalled or changing pages. Snapshot material when permitted and record access dates.

    CHECKPOINT · Every collected item has a source ID, type, date, region, locator, and review status before synthesis begins.

  3. 03

    Extract evidence before themes

    Capture the smallest passage that supports a claim. Mark whether it is verbatim, record the speaker or publisher, and add limitations. Redact personal data before any external model call.

    CHECKPOINT · A reviewer can open the cited location and determine whether each pivotal ledger row is faithful to the source.

  4. 04

    Generate a contested synthesis

    Ask for supporting and conflicting rows, segment differences, missing evidence, and plausible alternative interpretations. Keep generated themes separate from extracted passages.

    CHECKPOINT · Every recommendation cites reviewed source IDs and includes at least one counterargument or states that none was found in the bounded set.

  5. 05

    Run analyst and stakeholder review

    Verify pivotal claims against source snapshots, test whether the recommendation exceeds the sample, and have the decision owner record accept, revise, or reject with a reason.

    CHECKPOINT · The final brief includes reviewer names or roles, review date, unresolved gaps, and the next evidence-gathering action.

Hands-on lab

Build a cited customer-intelligence brief

Use the Meridian decision or a live question with approved evidence. Finish with a recommendation that a stakeholder can challenge claim by claim.

Prepare

  • • Write the exact decision, audience, deadline, and reversible next action
  • • Obtain permission for any first-party transcript or CRM material
  • • Create a folder or table where source snapshots and access dates can be retained

Deliverable

A reviewed source ledger and a two-page brief containing the decision, evidence, counter-evidence, recommendation, confidence note, open questions, and next research action.

Starter kit: Evidence ledger and synthesis prompt

Copyable CSV header plus prompt
source_id,source_type,publisher_or_role,title_or_interview,date,accessed_at,region,url_or_file_pointer,passage,verbatim,claim,confidence,limitations,review_status

SYNTHESIS INSTRUCTION
Decision: [insert]
Use only reviewed ledger rows. Separate observation, direct quotation, analyst inference, and open question. For every pivotal sentence, return supporting source_ids and conflicting source_ids. Say "insufficient evidence" when support is absent. Do not estimate prevalence from this convenience sample. End with the smallest reversible action and the evidence that would change it.

Expected result

A brief that remains useful after the meeting because its evidence can be inspected, refreshed, and reused. Readers can see which statements came from sources, which came from analysis, where the material disagrees, and how uncertainty affects the recommended action.

Carry forward

Keep the ledger as the evidence layer for the content operations and SEO/GEO modules. Add new rows rather than copying claims into an untraceable prompt.

Acceptance checks

  1. 01Every pivotal claim resolves to one or more reviewed ledger rows
  2. 02Direct quotations, paraphrases, and generated themes occupy distinguishable fields
  3. 03The brief contains counter-evidence, an alternative interpretation, and material limitations
  4. 04Restricted identifiers are absent from model inputs and the final artifact
  5. 05The recommendation names a reversible next action and evidence that would change it

What breaks

Failure clinic

F1The brief contains polished claims whose supporting link is irrelevant or missing.
Inspect
Sample pivotal sentences and follow their source IDs to the exact passage and snapshot.
Likely cause
Collection, extraction, and synthesis were combined in one unconstrained prompt.
Repair
Rebuild the ledger first, remove unsupported sentences, and regenerate from reviewed rows only.
Prevent next time
Block synthesis until required provenance fields pass validation.
F2Customer quotations appear stronger or cleaner than the transcript.
Inspect
Compare punctuation and wording with the timestamped passage; check the verbatim field.
Likely cause
A model paraphrase was formatted as direct speech.
Repair
Restore exact wording where permitted or label the sentence as an analyst paraphrase.
Prevent next time
Store quotation status as structured data and prohibit generated text in the passage column.
F3The recommendation claims a market-wide pattern from a few convenient sources.
Inspect
Review sampling method, role mix, geography, loss coverage, and source independence.
Likely cause
Frequency in the retrieved set was treated as population evidence.
Repair
Narrow the claim to the observed set and define research needed for broader inference.
Prevent next time
Include sample limitations and a claim-scope field in the acceptance rubric.
F4A refreshed brief silently changes position after sources update.
Inspect
Compare source snapshots, access dates, extraction versions, and changed ledger rows.
Likely cause
The workflow stored live URLs but no material snapshot or change record.
Repair
Re-verify affected claims and publish a concise change note with the new evidence.
Prevent next time
Retain permitted snapshots, content hashes, and a scheduled review date for pivotal sources.

Beyond the demo

Production boundary

  1. 01The research decision, segment, geography, time horizon, and decision owner are explicit
  2. 02The source policy covers primary preference, recency, consent, storage, and exclusions
  3. 03Source IDs, dates, access dates, evidence types, locators, and review status are required fields
  4. 04Personal and confidential data is redacted or kept inside an approved processing boundary
  5. 05Extraction is completed and reviewed before synthesis
  6. 06Pivotal claims expose support, counter-evidence, limitations, and analyst inference
  7. 07A human decision owner records acceptance or rejection with a reason
  8. 08Refresh triggers cover source change, new customer evidence, and scheduled expiry
Turn the ledger into a content pipeline →Build retrieval systems in the agent track →

Evidence status

Sources and claim limits

Sources support the named claims; they do not guarantee the same result in another system.

  1. [1]
    Structured Outputs guide

    OpenAI · Official documentation · 2026-08-20

    schema contracts · output validation · refusal handling
  2. [2]
    How we built our multi-agent research system

    Anthropic · Published research · 2026-08-20

    research decomposition · citation quality · effort controls
  3. [3]
    Introducing Contextual Retrieval

    Anthropic · Published research · 2026-08-20

    retrieval evaluation · semantic and lexical search · reranking
  4. [4]
    Web search tool guide

    OpenAI · Official documentation · 2026-08-20

    source-aware research · citations · tool configuration
  5. [5]
    Creating helpful, reliable, people-first content

    Google Search Central · Official documentation · 2026-08-20

    content quality · authorship · automation disclosure
  6. [6]risk ownership · measurement · human oversight
  7. [7]
    Workflow vs. Agent Architecture

    Tenten AI · Tenten field method · 2026-08-20

    research workflow framing · agentic retrieval · human validation

Related Tenten resources

Field guide

Workflow vs. Agent Architecture

A practical guide to choosing fixed automation or variable agent planning.

Open resource

Apply the track

Start with one constrained workflow.

Tenten can work with your marketing, data, and technical owners to validate the workflow boundary, build the production controls, operate the first release, and transfer ownership against visible evidence.

Discuss a workflowRead the FDM playbook