ISO 20252 AI Requirements: Five Points to Check in 2026

ISO 20252 AI requirements changed with the 2026 edition. What the public ISO text confirms, what only a consultancy summary says, and what a proposal can say...

The bid manager at a research agency opens the proposal template on a Monday and finds a new page. Somebody added it over the weekend, titled "AI use, per ISO 20252:2026 clause 4.1.5", with five bullet points copied from a post. The proposal is due Thursday. Nobody on the team has read clause 4.1.5, and it turns out nobody outside a certification process easily can. The ISO 20252 AI requirements are real and they did change, and that template page is still the wrong way to meet them.

Key Takeaways

  • ISO published ISO 20252:2026 as the fourth edition in September 2026, and the public preview confirms that AI drove changes to core requirements and annexes.
  • The requirement clauses are not in the public preview. The widely shared list of five AI points comes from a consultancy's paraphrase of clause 4.1.5, not from quoted standard text.
  • Each of those five points is already asked for, in other words, by the ICC and ESOMAR code, the AAPOR disclosure checklist, joint guidance from MRS, ESOMAR and the Insights Association, or the EU AI Act.
  • A proposal can describe the practice today and attribute each point to the source that states it, without citing a clause number nobody on the team has read.
  • Confirm the exact wording with your certification body before any document claims conformity with the 2026 edition.

What was actually published

ISO 20252 is the international service standard for market, opinion and social research, including insights and data analytics. It sets terms, definitions and service requirements, and agencies can be audited against it by a certification body.

The ISO catalogue entry for ISO 20252:2026 lists it as published, edition 4, with a publication date of 2026-09 and 77 pages. It was prepared by technical committee ISO/TC 225, and the 2019 edition now shows as withdrawn. The BSI Knowledge page for the 2019 edition marks it withdrawn on 3 September 2026.

Research Live reported the publication on 1 September 2026. According to that report, the committee added AI and automation policies with monitoring requirements, strengthened data quality for sampling and panels, and updated privacy and security. More than 400 research organizations hold third party certification to the standard. The report does not list specific requirements on client disclosure, consent, versions or participants.

None of this was a surprise. The Insights Association described the revision in December 2024 as a response to the impact of AI on research methods and to data fraud.

What the public text of the standard confirms about AI

ISO publishes the informative parts of its standards on its Online Browsing Platform preview of ISO 20252:2026. We read it on 15 September 2026. The page states plainly that only informative sections are publicly available, and that the full content requires buying the standard.

The foreword is explicit about AI. It says the document was reviewed to reflect the increasing use of automated techniques, AI and related technologies. It adds that requirements were amended for those techniques where appropriate. It also says that including AI significantly changed some core requirements and annexes, naming translation, coding and analysis as examples. The edition also absorbs ISO 19731:2017 and splits panel management into a new Annex G.

The introduction says AI has been reflected in a technology neutral manner, because the techniques change fast. That matters for anyone hoping for a list of approved tools. The standard does not name models.

The definitions are public too, and several are new ground for this standard:

An AI agent is an automated entity that uses AI to perform tasks, interact with its environment and act toward specific goals. That is definition 3.10, and it adds that agents often interact with users, data or other systems.

Automated coding is coding where computer methods replace manual coding, fully or in part. Definition 3.15 names machine learning and AI among those methods.

Synthetic data is information that computational or statistical models generate from real world data, usually historical, to approximate the patterns of the target group. That is definition 3.128.

A note to definition 3.92 says a questionnaire can be self completion, administered by a fieldworker, or administered by an AI agent. The standard now treats an AI interviewer as an ordinary case to be governed.

What the preview does not show is the requirement text. The table of contents lists clause 4.1, Core framework and legal requirements, without its subclauses. Clause 4.1.5 does not appear anywhere in the public preview.

What is confirmed about the AI clause, and by whom

Three stacked layers of evidence about the AI requirements in ISO 20252:2026, from the most direct to the least. Layer one, the standard itself, from the ISO public preview read on 15 September 2026: edition 4, published 2026-09; the foreword says requirements were amended for automated techniques and AI, naming translation, coding and analysis; new definitions cover AI agents, automated coding and synthetic data; the requirement clauses are not public. The owner line names the ISO catalogue entry and the Online Browsing Platform preview. Layer two, trade press, the Research Live report of the publication on 1 September 2026: AI and automation policies with monitoring, stronger data quality for sampling and panels, updated privacy and security, more than 400 research organizations holding third party certification, and no detail on disclosure, consent, versions or participants. Layer three, drawn dashed, a consultancy summary published on 2 September 2026 as a first analysis and paraphrase with no clause text quoted: clause 4.1.5 with five points plus monitoring, namely describing the AI and the stage in proposals, telling the client when AI generated insight, explicit client consent for data going through AI, documenting versions, and informing participants who interact with AI. The note says this is the only layer that names clause 4.1.5 and it is not the text of the standard. A closing line says to quote the practice, attribute the point and confirm the clause with the certification body. Every figure belongs to the source named under it, and none is a Cassi.ai reading of the clause.

The five points everyone is sharing sit in the bottom layer. The top layer confirms that AI changed the standard, and stops short of saying how.

Where the list of five points comes from

UNIQUE INSIGHT The five points travelling around the industry have one traceable origin, and it is a secondary source.

On 2 September 2026, a UK consultancy specializing in ISO 20252 published its first analysis of the new edition. It says clause 4.1.5 covers the development and use of AI and related technologies, and lists what it calls explicit requirements. Proposals describe the type of AI and the research stage where it was applied. Use of AI to generate research insight is recorded and communicated to the client. No project or client owned data passes through AI without the client's explicit consent. The versions of AI tools are documented. Participants who interact with AI are informed of its use. It adds a sixth point: monitoring procedures covering security, privacy and quality.

That list is useful, and it is plausible. It is also a paraphrase. The consultancy calls the piece its first step in analysing the changes, it quotes no clause text, and it sells certification support, which is a declared interest rather than a flaw. One detail is a reason to hold the summary at arm's length. It gives a page count for the new edition that differs from the one in the ISO catalogue.

So the honest status of the five points is this. They are attributed to clause 4.1.5 by one consultancy. Research Live independently reports AI and automation policies with monitoring. The public ISO text confirms that AI changed the requirements, and does not show which ones. We have not read clause 4.1.5, and this article does not quote it.

How agencies answer the AI question in proposals today

PERSONAL EXPERIENCE Before the new edition, the AI paragraph in most proposals I have seen was written the same way.

  1. The client's request for proposal asks, somewhere near the end, whether AI will be used.
  2. The person writing the proposal drafts a sentence that is true at the level of the whole company.
  3. The sentence says nothing about which stage, which tool or which data.
  4. The paragraph gets reused in the next proposal, for a different study with a different design.
  5. When the client later asks what exactly the AI did, the answer lives in someone's memory.

The risk now is a new version of step two. The paragraph gets replaced by a clause number and five bullets, which reads as more rigorous and is less accurate. The earlier piece on AI policy for research agencies covers the policy document behind the proposal. This one is about the page the client actually reads.

Point one: say what the AI does, and at which stage

The first point, as the consultancy words it, is that the proposal describes the type of AI and the research stage where it applies. You do not need the clause to justify doing that.

The AAPOR task force report on responsible AI in survey research, published in May 2026, sets required disclosures that it describes as part of the AAPOR Code. When AI acts as an interviewer, a respondent or an analyst, researchers disclose the task the AI performed and a plain language description of what it did. The report explains why a label alone is not enough. Saying that AI was used as an interviewer does not tell the reader the scope of its involvement.

A stage register is a list of the research stages in a study, with each one marked as manual, assisted or automated. It is the simplest way to write point one honestly, one line per stage, for this study rather than for the company.

Point two: tell the client when AI produced the insight

The second point is that use of AI to generate research insight is recorded and communicated to the client. The ethics code already asks for something close to it.

Research World, ESOMAR's publication, summarizes the 2025 ICC/ESOMAR Code this way. Article 7 asks researchers to disclose blind spots, gaps and the role of AI, along with the extent of human oversight. Article 9 asks that the public be informed when synthetic data or AI was used. ESOMAR's announcement of the revised code in June 2025 described new definitions and responsibilities around AI, synthetic data and synthetic personas.

The AAPOR report adds the part clients care about most. Researchers disclose whether AI output was validated by a person, and how. Without that, the report says, a reader cannot tell verified results from results that were trusted without review. The method we use to check agent output without a reference answer is in AI agent evaluation without an answer key.

Point three: explicit consent before client data goes through AI

The third point is the one with legal weight. No project or client owned data passes through AI without the client's explicit consent.

Joint guidance from MRS, ESOMAR and the Insights Association, reported in June 2026, points the same way from the contract side. It treats client ownership of research data as a sound starting point and says contracts should state it plainly. It notes that a de-identified dataset, an aggregated summary or a synthetic output can still carry the client's strategic thinking. It recommends approved tool lists, data classification, approval routes and retention rules.

Explicit consent, in a proposal, means a named data category, a named tool or deployment, and a signature from the client. A general clause saying the agency may use technology to process data is the opposite of explicit.

This is also where deployment choices stop being an engineering detail. Whether a model runs inside the client's environment or on a vendor's service changes what the consent has to cover. We laid out that decision in open weight models in market research.

Point four: write down the version

The fourth point is documenting the versions of the AI tools used. It sounds bureaucratic until a tracker changes.

The AAPOR report is the clearest source here. Its section on evaluation says not to assume that performance in one use carries over to another, and lists model version among the reasons. The next section says not to assume that past performance holds in the future, because providers update systems and integration layers change. The report's enhanced disclosures ask for the specific AI system and the dates it was used, noting that even minor version updates can change behaviour.

An older study shows why this is not a formality. How is ChatGPT's behavior changing over time? compared the March and June 2023 versions of the same commercial service across seven kinds of task. On one task, identifying prime numbers, the authors report GPT-4 accuracy falling from 84 percent to 51 percent between the two versions. Nothing about the product name told a user that anything had changed.

A model version is the specific release of a model that produced an output, as named by its provider, together with the date it was used. A model name alone, for example a product brand, is not a version.

For a proposal, the practical line is short: the model and version in use at the time of the study, and a commitment to record any change during fieldwork or analysis. For a continuous tracker it is not optional. A change of model between waves is a change of method, and the client will want to know which wave it hit.

Point five: tell participants when they are talking to AI

The fifth point, in the consultancy's words, is that participants who interact with AI are informed of its use.

The public ISO text supports the premise, even without the requirement. Its own definition of a questionnaire now includes administration by an AI agent. The AAPOR report treats AI as an interviewer as a distinct role with its own risks, including consent and auditability. It also draws a line clients should keep in mind. Responses generated by AI, such as synthetic responses or digital twins, are not research participants, and data created that way must be identified as such.

A participant, in the standard's definition 3.80, is a person or organization from whom or about whom data are collected for research. A person answering questions from an AI interviewer is still that. The disclosure belongs at the start of the interview, in plain words, before the first question.

For studies with participants in the European Union, there is a legal layer on top. The European Commission's page on the AI Act says that people using AI systems such as chatbots should be made aware that they are interacting with a machine. How that applies to a given study design is a question for the client's legal team rather than for a proposal template.

Five points, and who states each one

UNIQUE INSIGHT The table below is our mapping, and it keeps the attribution visible. The middle column is the only place where clause 4.1.5 appears, and it is always the consultancy speaking.

PointAttributed to clause 4.1.5 byAlso asked for byWhat a proposal can say today
Describe the AI and the stageThe consultancy summaryAAPOR required disclosures; ICC and ESOMAR Code article 7, as summarized by Research WorldOne line per research stage: manual, assisted or automated, and what the AI does there
Tell the client when AI generated insightThe consultancy summaryICC and ESOMAR Code article 7; AAPOR disclosure of human validationWhich findings were drafted with AI and who reviewed them
Explicit client consent for data through AIThe consultancy summaryGuidance from MRS, ESOMAR and the Insights AssociationNamed data, named tool or deployment, client signature
Document the AI versionThe consultancy summaryAAPOR enhanced disclosures and sections 4.1.2 and 4.1.3Model and version at study start, with changes logged
Inform participants who interact with AIThe consultancy summaryAAPOR on AI as interviewer; the EU AI Act, as the European Commission describes itA plain disclosure before the first question

Every row has at least one source outside the summary that asks for the same practice. Only the attribution to a clause number depends on the summary alone. Monitoring, the consultancy's sixth point, matches what Research Live reported and belongs in the policy rather than in each proposal.

Before a sentence about the standard goes into a proposal

Nothing above requires quoting a clause you have not read. It does require a small discipline about how the standard is mentioned.

  1. Buy the standard, or ask your certification body for the wording of the AI subclauses that apply to your scope.
  2. Until then, describe the practice in the proposal and attribute each point to the source that states it.
  3. Keep clause numbers out of client documents unless someone has checked them against the text.
  4. Do not write that the agency conforms to ISO 20252:2026 unless an audit against that edition has happened.
  5. Record the model, version and deployment for each study, whatever the clause turns out to say.

Before a sentence about the standard enters a proposal

A decision flow in three steps, each with an exit to the right. Step one asks whether someone has read the clause text, meaning the purchased standard or the wording supplied by the certification body. If the team only has a summary or a post, the proposal describes the practice and names the source that states it, with no clause number. Step two asks whether the point matches what the team does, with the stage, tool, model version and reviewer recorded for this study. If that is not recorded, the record is fixed first, because a claim without it fails the first client question. Step three asks whether an audit has covered the 2026 edition, meaning certification by an accredited body against this edition. If there is no audit yet, the proposal says what the agency does and does not state that it conforms. A closing band says that three yes answers allow citing the clause, in the words of the text that was read.

Every exit to the right still produces an honest proposal. The only thing it removes is a clause number standing in for a practice.

The consultancy summary itself notes that certification to the new edition depends on accreditation schemes being updated and auditors being trained. Ask your certification body for the transition dates that apply to you, and plan from its answer rather than from a post.

Where Cassi.ai comes in

@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights, founded by Cassiano Albuquerque. We are not a certification body, we do not issue compliance opinions, and nothing in this article is a claim of certification.

Deskor is a Cassi.ai product that connects company knowledge, AI agents and workflows, so the team builds proposals, decks, reports and dashboards from what the company already knows.

What we build is the record the five points depend on. Our Super Agents line runs inside the client's own environment, on premises, in a private cloud or air gapped. Every step it takes writes to a record outside its own session, by user, project and client, as we described in AI policy for research agencies. That record is where the stage, the model version and the reviewer come from when a proposal or a client asks. The rest sits in the cassiai.com portfolio.

FAQ

What changed in ISO 20252:2026?

The public foreword says the fourth edition was revised for the growing use of automated techniques and AI. It amends requirements for those techniques and names translation, coding and analysis among the areas that changed. It also aligns clause 4 with ISO/IEC 27001, merges ISO 19731:2017, and moves panel management into a new Annex G, per the ISO preview.

What is clause 4.1.5 of ISO 20252?

A consultancy summary describes clause 4.1.5 as covering the development and use of AI and related technologies, with five disclosure and consent points plus monitoring. The clause text is not in the public ISO preview, so confirm the wording with your certification body before quoting it.

Does ISO 20252 require disclosing AI use to clients?

Research Live reports that the 2026 edition adds AI and automation policies with monitoring. A consultancy summary says clause 4.1.5 requires communicating AI generated insight to the client. Independently of the standard, the ICC and ESOMAR Code asks researchers to disclose the role of AI and the extent of human oversight.

Do research participants have to be told they are talking to AI?

The consultancy summary says participants who interact with AI are informed of its use. The ISO text itself now defines a questionnaire as something an AI agent can administer. The AAPOR report treats AI as an interviewer as a distinct role with consent risks. Disclose it before the first question.

Can an agency say it complies with ISO 20252:2026 now?

Only after an audit against that edition by a certification body. A proposal can describe the agency's AI practice accurately, point by point, without claiming conformity to a clause the team has not read.

A clause number in a proposal is a claim. A clear description of what the AI did, at which stage, under whose review, is evidence, and it survives whatever the clause turns out to say.

If your template gained an AI page this month, check each bullet on it against the table above. Any line that names clause 4.1.5 and nothing else is the first one to rewrite.

About the author

Cassiano Albuquerque is the founder of @Cassi.ai. He has spent twenty years in market research operations, from fieldwork and tabulation to the reports that reach steering committees, and now builds AI systems for research teams. To discuss how your AI steps get recorded by stage, model and reviewer, contact him on LinkedIn or see the cassiai.com portfolio.

Published by CassiAI. Read the full article at https://www.cassiai.com/blog/iso-20252-2026-ai-requirements.