An insight agency should introduce AI in one month of deliberate steps: approve a short tool list and decide what client data may enter those tools; pilot one repetitive, checkable workflow on a finished project; measure time saved after correction; and keep a named researcher accountable for the result.
The first month is a data and accountability decision before a tooling decision. This guide covers days 1–30; Meaningful's AI-native playbook covers the longer path.
Disclosure: Meaningful publishes this guide. Its product is mentioned as one possible pilot tool; the policy and evaluation questions apply to any vendor.
What should an agency decide before anyone touches a new tool?
In week one, find out what the team already uses, then approve a small set of tools and a clear boundary for client material. Do not start by buying licences for everyone: tool access without a shared workflow leaves quality and data handling to individual habit.
Ask which tools researchers use, on what material, and which outputs they distrust. Treat the inventory as learning, not a hunt for rule-breakers. Check personal accounts and files copied between projects.
The failure mode is sidecar AI: individuals become faster using private prompts and unshared processes while the agency learns nothing it can inspect, teach or govern. The AI-native playbook names this pattern. A shared workflow needs a home for source material, an approved tool for the job, an output someone else can review and a named owner.
By week's end, publish an approved-tool list, choose the pilot environment and name an exception approver. Record data destinations and retention; if unknown, use only public or cleared test material. See the stack nobody designed.
What should a research agency's AI policy cover?
Keep the first policy to one page answering five operational questions. It should say what a researcher can do on Monday morning, what requires approval, and who handles an error; a broad promise to “use AI responsibly” is not a control.
| Policy question | Decision to write down | Research-specific check |
|---|---|---|
| What may staff use? | Approved tools, workspaces, roles and an exception route | Distinguish a personal account from a contracted agency environment |
| What data may go in? | Public, internal, client-confidential and respondent data rules | Check project contract, participant consent, vendor training terms, residency and retention |
| What must be checked? | Source and method QA before any client use | Verify bases, quotes, coding, omissions and contrary responses |
| When do we tell clients? | Contract and readout language naming material AI use | State the tools or processing steps and the extent of researcher oversight |
| Who is accountable? | Named project reviewer and incident escalation owner | Stop delivery, correct the evidence and notify the right people if a boundary fails |
The note Client data, AI and the boundaries of consent came from the Market Research Society (MRS) and ESOMAR's Professional Standards Committee, following an Insights Association statement. It urges clarity on client ownership, consent for AI training and vendor terms. Anonymising a transcript does not automatically release client intellectual-property rights.
Translate it into approval, classification, retention and deletion rules. Meaningful's customer GDPR guide is a provider-specific example, not certification against the note. Have legal and data leads check the policy against client contracts.
Which workflow should you pilot first?
In week two, choose a recurring task that can be checked against its source. Use a completed project only if contract, consent and retention allow reuse; otherwise create a cleared test pack.
| Candidate | Potential value | How to check it | First-pilot caution |
|---|---|---|---|
| Interview transcript summaries | Cuts first-pass reading time | Compare each theme and quote with transcripts | Do not mistake a tidy summary for complete coverage |
| Open-end coding | Speeds a repetitive quant/qual hand-off | Audit code assignments, exceptions and achieved bases | Agree the codebook before measuring accuracy |
| Public desk scan | Helps frame a brief | Open sources, dates and original claims | Fastest to approve, but less representative of client-data work |
| Proposal first draft | Reduces blank-page time | Check scope, method and commercial assumptions | Do not input an unapproved confidential brief |
| Multi-source synthesis | Tests a costly agency bottleneck | Reconcile survey, interviews and desk evidence | Higher value, but harder to review as a first experiment |
A transcript summary or open-end coding pass is a practical start. Once QA is stable, Meaningful is one option for multi-source synthesis. The AI tools guide helps build a shortlist.
Write down inputs, output, reviewer and stopping rule first. Use the same study as the manual baseline and keep its delivered result hidden during blind review.
How do you know the pilot worked?
In weeks three and four, measure net time saved and research quality, not the speed of the first AI draft. A workflow that saves four hours of coding but adds five hours of correction has not improved delivery; a fast, untraceable finding should fail even if the prose reads well.
| Scorecard item | Record for the manual and AI-assisted routes | Pass question |
|---|---|---|
| Net time | Preparation + run + human correction + review | Did total researcher time fall at comparable quality? |
| Factual and method errors | Wrong bases, fabricated or altered quotes, miscodes, missed caveats | Were material errors caught before client use? |
| Coverage and traceability | Included records, source links and contradictory cases | Can a reviewer reproduce the main finding? |
| Blind senior review | Usefulness, clarity and confidence without knowing which route produced it | Would the director send this after normal sign-off? |
| Reuse | Inputs, instructions and QA steps needed for another team member | Can the agency repeat the workflow without its inventor? |
Quirks gives an illustrative 20-interview calculation: 10 hours manually versus 20 minutes of AI work plus two hours refining, a net seven hours and 40 minutes before other review. It is not a benchmark.
Set acceptance criteria before seeing output. Log AI errors separately from poor inputs or instructions. If a conclusion cannot be traced to a respondent or table, revise the process or stop.
How do you bring the team with you?
Ask researchers what work they want help with and what they fear losing before announcing a rollout. A mandate framed only as efficiency invites quiet workarounds; a credible plan names which decisions remain with researchers and shows the pilot's errors as openly as its time savings.
Invite qual, quant and delivery volunteers to challenge the scorecard. Give analysts and directors separate review roles. Save inputs, corrections and sign-off steps in a shared library.
Do not promise job security you cannot authorise. Explain the tested task, the intended use of freed time and who makes staffing decisions. The next hire should inherit the approved route.
Should you tell clients you're using AI?
Yes: explain material AI use before it surprises the client, especially when their data or respondents are involved. Agree the platform, permitted inputs, human oversight and deliverable language in the contract or project method, then describe what the researcher checked in the readout.
The 2025 ICC/ESOMAR Code sets expectations for transparency and professional responsibility; the MRS and ESOMAR note presses for clearer client arrangements around platforms and data. Neither a vendor's “no training” setting nor a de-identified upload grants permission under a client contract. Check the exact obligations with your legal and research leads.
Decide billing before live use. On an hourly contract, bill for work actually done under the agreement. On a fixed fee, price scope, evidence quality and senior judgement, with the AI-assisted method clear to the client.
What comes after the first month?
At day 30, keep only a pilot with a repeatable input pack, a measurable net gain and a review standard the team trusts. A failed pilot can still be useful if it shows that permissions, source packaging or QA need work before wider use.
The next steps are to standardise one lane, connect the evidence layer and build institutional judgement, not to roll every task into the first tool that worked. Meaningful's full AI-native playbook develops those later stages and the longer rollout. This post is the starting checklist, not a replacement for that plan.
Frequently asked questions
The first month should make AI use visible, permitted and reviewable. These answers are starting rules; client contracts and the study's method set the final boundary.
Where should an insight agency start with AI?
Start with an approved-tool list and a rule for client data, then pilot one repetitive task against a finished, permission-cleared project. Name the reviewer and record the manual baseline before comparing output.
What should a research agency's AI policy include?
Cover approved tools, permitted data, checks, client disclosure and a named sign-off owner. Add classification, vendor terms, retention, deletion and an escalation route for a mistaken upload or unsupported finding.
Can we put client transcripts into ChatGPT?
Only if the specific workspace and workflow are approved and the client contract and participant consent permit it. Official OpenAI documentation says its business products do not use business data for model training by default, but you still need to check access, retention and connected services; that product setting does not grant client permission.
Which research task is best for a first AI pilot?
Choose a frequent, slow, checkable task such as transcript summary or open-end coding. Use cleared source material so a researcher can compare every substantive claim with the original record.
How long should an AI pilot run?
Use four weeks to inventory, approve, test and review one workflow, not to declare the agency transformed. Extend only if the scorecard lacks enough comparable work or a data permission remains unresolved.
Do we have to tell clients we use AI?
Tell them about material AI use and human oversight in the contract or method, subject to the project's obligations. Be specific about what data the tool processes and what a named researcher checks before delivery.
Will AI replace researchers at our agency?
It can reduce repeatable processing, but it does not take responsibility for sample, interpretation or client advice. Tell the team which tasks are changing, what remains human-led and how you will measure the effect before making broader promises.
Next step
Use the first month to make one AI-assisted research workflow safe, checkable and shared. Then read the full AI-native playbook for the four stages from first pilot to institutional judgement. Keep the scorecard and the data rules beside the enthusiasm.