← All posts
insight agenciesresearch operationsAI

Is it safe to put client research data into ChatGPT?

Sajin Kabeer · November 17, 2026

It can be, but safety depends on the plan, the contract and the data, not on the tool's name. Before client research goes into ChatGPT or any AI tool, check that your inputs are not used for training, that a data processing agreement covers the work, and that participants and the client permit the use. Removing names protects participants, not the client's findings.

Most agency risk sits in the gap between what a researcher can paste and what the agency has promised.

Disclosure: Meaningful publishes this guide and sells an AI platform for research agencies. The checks below apply to every tool, Meaningful included. This is practical guidance, not legal advice. Sources checked on 16 November 2026.

What actually makes it safe or unsafe?

Four things decide it: the plan or account the data goes into, the contract behind that plan, the data itself, and the permissions you hold. The same tool can be acceptable under one set and a breach under another.

A data processor is an organisation that handles personal data on someone else's instructions. A data controller decides why and how that data is used. In client research, the agency is usually the controller for the data it collects, and the AI provider is a processor.

What two kinds of data are at stake?

There are two layers, and they are protected by different means. The joint MRS and ESOMAR note, Client Data, AI, and the Boundaries of Consent, separates them clearly.

LayerWho it belongs toWhat protects itWhat anonymisation does
Personal data in the researchThe participantsData protection law and the consent they gaveAddresses it, if done properly, though confidentiality promises still apply
The client's intellectual propertyThe clientThe client contract and vendor termsNothing. The questions, hypotheses and findings remain the client's

The note puts it plainly: "anonymising participants does not anonymise the research." A de-identified dataset or an aggregated summary can still carry the client's strategy. If that feeds a model available to others, the client may feel it has paid for a competitor's advantage.

This is why "we removed the names" does not settle the question. It answers the participant duty and leaves the client duty open.

Does ChatGPT train on what you type?

It depends on the plan and its settings, so check the data controls for the exact account your team uses. Do not rely on what a colleague remembers from last year, because AI providers change plans and defaults.

Ask three questions of any AI assistant, ChatGPT included:

  1. Are inputs and files used to train or improve models on this plan, and who can change that setting? A setting a single user can switch back on is weaker than a workspace-level control.
  2. How long are conversations and uploaded files kept, and can the agency delete them?
  3. Is there a data processing agreement, and does it list the sub-processors?

Write the answers down with the date you checked them. The MRS and ESOMAR note warns that broad vendor wording such as "internal analyses", "benchmarking" or "aggregated and de-identified use" can cover model training. Read those clauses with your client contract next to you.

OpenAI's trust portal states that its products are covered by a SOC 2 Type 2 report for the period 1 July 2025 to 30 June 2026. A SOC 2 report is an independent audit of security controls. It does not answer whether your use is permitted by your client or your participants.

What does GDPR require before you upload respondent data?

If the data contains personal information about participants, you need a processor contract and, often, an impact assessment. These are the two GDPR articles agencies meet most often here.

A data protection impact assessment, or DPIA, is a written review of what could go wrong for the people in the data and how you will reduce it. AI analysis of identifiable interviews is the kind of processing many agencies will want to assess first. Your data protection officer or legal adviser should confirm what applies to you.

What do MRS and ESOMAR say?

They say the old assumption that client data belongs to the client still holds, but it can no longer be left unsaid. Their joint note is aimed at agency leaders and intermediaries and has six practical points.

  1. Restate the default in writing. Say in contracts that client data and derived insights belong to the client. "Silence is no longer safe."
  2. Know what you can consent to. Before accepting a platform's terms, check that participants, the client and upstream partners permit the use. "You cannot grant rights you do not hold."
  3. Separate what anonymisation does and does not do. It handles the participant layer, not the client layer.
  4. Look downstream as well as upstream. Your tools' terms must match your client promises, or the agency may be in breach even when the vendor is not.
  5. Make disclosure the norm. Tell clients which AI tools touched their project, at what stage and on what terms.
  6. Build a framework that matches actual work. An approved tools list, data classification, approval routes and retention rules, with training based on real research tasks.

The note also says the questions apply to older platforms, not only new AI services. A transcription or translation tool cleared years ago may now include AI features that use data differently. ESOMAR has also published 20 questions to help buyers of AI-based services, which work well as a vendor questionnaire.

What should you check before any upload?

Run the same seven checks for every tool and every project. If one fails, change the data, the tool or the permission before uploading.

CheckQuestion to answerWhere the answer lives
1. PlanIs this the agency's contracted workspace, not a personal account?Your IT or admin settings
2. TrainingAre inputs excluded from model training on this plan, at workspace level?The provider's data controls and terms
3. ContractIs a data processing agreement in place, with sub-processors listed?The provider's DPA and sub-processor page
4. LocationWhere is the data processed and stored, and does that suit the client?The DPA or data policy
5. ParticipantsDid the consent wording cover AI processing for this purpose?Your screener and privacy notice
6. ClientDoes the contract allow this tool, or require approval of processors?The client contract or framework agreement
7. Minimum dataHave you removed what the task does not need, such as names, employers and contact details?The file you are about to upload

Keep a one-line record per project: which tool, which data, which date, who approved. That record is also your client disclosure.

What can you safely use a general assistant for?

Tasks that do not need client or participant data are the easy wins. Drafting a discussion guide from a public brief, rewriting your own email, outlining a report structure or explaining a statistical method all carry little data risk.

The risk rises as material moves from your own words, to client documents, to identifiable participant data. Agencies that are introducing AI often start with the first group and add the others as contracts and checks catch up. Our first 30 days guide sets out that sequence, and our comparison of AI chatbots and research platforms covers when a general assistant fits the analysis itself.

How does Meaningful handle client data?

Meaningful publishes its data terms so an agency can run the seven checks before a trial. The main points, from its data policy, sub-processor list and customer GDPR guide:

None of this replaces your own checks. Participant consent and client permission are still the agency's to confirm, and Meaningful does not replace the agency's own methodological judgement.

Frequently asked questions

Is it safe to put client research data into ChatGPT?

It can be, but safety depends on the plan, the contract and the data rather than on the tool's name. Check that your inputs are not used for training, that a data processing agreement covers the work, and that both the participants and the client permit the use.

Does anonymising transcripts make it safe to use AI?

It protects participants, not the client. MRS and ESOMAR point out that the client's questions, hypotheses and findings are its intellectual property, and that anonymising participants does not anonymise the research. You still need the client's permission and suitable vendor terms.

Do I need a data processing agreement to use an AI tool with research data?

If the tool processes personal data on your behalf, GDPR Article 28 requires a binding contract with that processor. Check which plan and terms your team is actually using, and whether an agreement applies to that plan.

Do I need a data protection impact assessment before using AI on respondent data?

GDPR Article 35 requires one before processing that is likely to result in a high risk to people, in particular when new technologies are used. Many agencies will need one for AI analysis of identifiable interviews. Ask your data protection officer, as this article is not legal advice.

Should I tell clients which AI tools touched their project?

Yes. MRS and ESOMAR describe a short standing disclosure of which platforms are used, for what and on what terms as a baseline professional expectation. Some client contracts also require approval before you add a processor.

Does Meaningful train AI models on client data?

No. Meaningful's data policy states that customer data is not used to train or improve any AI models, and that customer data is processed and stored within the European Union. The agency remains the data controller and Meaningful acts as its processor.

What should an agency AI policy cover?

MRS and ESOMAR suggest an approved tools list, a data classification that says what is allowed, needs approval or is never allowed, an approval route for harder cases, rules on when clients are told, and retention and deletion requirements.

What should you do next?

This week, write a one-page approved tools list using the seven checks above, starting with the tools your team already uses. If Meaningful is a candidate, read its data policy with a client contract next to you, then test it on one permission-cleared project in the three-day free trial.