# Claims Checker: An AI Agent That Reviews Marketing Claims

An AI marketing claims checker compares draft copy with an approved claims register and required disclosures, then flags what needs attention. The Claims Checker serves lifecycle marketers and compliance reviewers at L2, the propose level, alongside any customer data platform (CDP). It reads copy rather than customer data. It never approves a regulated claim, publishes copy, invents a substantiation or personalises from raw personal data.

The profile sits within the vendor-neutral [AI agent rails](https://privycdp.org/ai/rails) method.

## Rail at a glance

The Claims Checker's rail limits what it may cite as well as what it may do. A retrieval allow-list holds the approved documents, and a validator rejects any register ID that does not exist.

| Field | Value |
|---|---|
| Job | Check marketing copy against an approved claims register and required disclosures. |
| Trigger | Copy submitted for review. |
| Data scope | Draft copy, claims register, required-disclosure list and audience description. No customer data. |
| Tool scope | Read: draft copy, claims register, disclosure rules. Propose: annotated findings with suggested wording. Execute: none. |
| Output contract | Flagged claims, the rule each one relates to, suggested wording, and items needing legal review. |
| Gate | A compliance reviewer signs off regulated claims, and the copy owner resolves the rest. |
| Limits | Budget per run, and approved reference documents only. |
| Ledger | Copy version hash, rule set version, findings, reviewer decision, model version. |
| Evaluation | Seeded claims per regulated topic, both compliant and non-compliant. |
| Owner and kill switch | Head of lifecycle marketing, with compliance. A pause from the [Orchestrator](https://privycdp.org/ai/rails/orchestrator-agent) returns review to compliance. |

## What it does and who uses it

The Claims Checker gives copy a first read before a reviewer sees it, looking for claims writers did not realise they were making. Rules confirm that a required sentence is present, and the model looks at what a reader would take away.

An illustrative submission from a vitamin brand reads: "Say goodbye to winter colds with our new immune blend, now in eco-friendly pouches." The first phrase implies a health benefit, so the agent offers the nearest wording from the register. "Eco-friendly" has no register entry, so it goes to legal review with no suggestion. The reviewer signs off the amended copy against its hash.

A reader's impression is what regulators test. Staff at the US Federal Trade Commission assess the "net impression" of an advertisement, including implied claims, and expect evidence before the ad runs [3](#source-3).

A register needs different logic for each regime; these starting points were checked on 1 November 2026.

| Regime | What it asks | Agent can flag | Stays with a person |
|---|---|---|---|
| EU food health claims, Regulation (EC) No 1924/2006 | A register of authorised claims with conditions, and refused claims with reasons [1](#source-1) | Wording that departs from an entry | Whether the product meets the conditions |
| UK advertising, CAP Code rule 3.7 | Evidence before publication for claims consumers read as objective [2](#source-2) | An objective claim with no evidence link | Whether the evidence is adequate |
| UK financial promotions, Financial Conduct Authority (FCA) | Fair, clear and not misleading, with capital risk made clear [4](#source-4)[5](#source-5) | A benefit with no nearby risk | Sign-off under the firm's process |
| EU green claims, Directive (EU) 2024/825 | Generic environmental claims banned from 27 September 2026 without recognised excellent performance [7](#source-7) | Terms such as "green" or "eco" | Whether the evidence meets the bar |

String matching falls short here. The FCA prescribes no phrase for mainstream investments and says consumers often do not read "capital at risk" statements [5](#source-5), so a rule looking only for those words misses the purpose.

## Why this level

The Claims Checker stays at L2 because sign-off on a regulated claim is accountable work. Copy reaches many people at once, and a published claim cannot be pulled back from an inbox.

Meta's Agents Rule of Two counts a single property in this session [8](#source-8). Copy is untrusted input, since agencies, feeds and merge fields supply it. A hostile line in a template could still push the model towards a false all-clear, so findings stay advisory.

No promotion path exists here. Speed comes from routing instead: copy with no regulated flag goes straight to its owner.

## Three ways it fails, and the control for each

The Claims Checker's failures matter most in copy that looks harmless, so each control works whether or not the model noticed anything.

| Failure | What it looks like | Control |
|---|---|---|
| Implied claim missed | A benefit is suggested through tone or a product name, and no listed word appears. | Seeded implied claims, with catch rate reported by topic. Regulated categories reach a reviewer whatever the agent finds. |
| Invented support | The model cites a study or suggests wording that is not in the register. | Each reference must resolve to a register ID, and a validator blocks anything else. |
| Stale or wrong-market rule | A claim cleared last year, or for another country, passes where the rule differs. | Register rows carry market, expiry and owner. A new rule set version queues affected copy for another check. |

## How to evaluate it

The Claims Checker needs an evaluation set with compliant claims as well as bad ones, because a checker that flags everything scores perfectly on catch rate and helps nobody. Compliance reviewers should build it from past campaigns.

Measure these, split by regulated topic:

- **Catch rate on seeded non-compliant claims.** Look at implied claims separately from express ones.
- **False-flag rate on compliant claims.** Too many false flags teach reviewers to skim the queue.
- **Injection cases.** Instructions hidden in copy must not change the findings.

Report results by topic and rail, not by named reviewer. The EU AI Act's Annex III, point 4(b), covers systems that evaluate workers, so take advice before comparing individuals [6](#source-6).

## Where it runs

The Claims Checker reads copy and reference documents, so it can use an external model where the contract permits. A private model suits regulated teams, as the [boundary architecture patterns](https://privycdp.org/ai/ai-agents-on-customer-data-inside-your-boundary) explain. An immutable audit ledger in PrivyCDP names the actor that authorised each AI-assisted action, while a claims rail would be your own design.

## Related

- [Marketing claims compliance check on rails](https://privycdp.org/ai/rails/use-cases/copy-claims-compliance-check): the use case in full.
- [AI agents for lifecycle and CRM marketing](https://privycdp.org/ai/rails/functions/lifecycle-marketing): the playbook for this team.
- [Rail card template](https://privycdp.org/tools/agent-rail-card): a card for your rule set and reviewers.


Browse related reviewers at [AI agent rails](https://privycdp.org/ai/rails). Scope your own claims review with the [rail card template](https://privycdp.org/tools/agent-rail-card).

Rail level: L2 · Risk: medium · Function: lifecycle marketing

By Art Kukharenko
Published: 2026-11-02
Updated: 2026-11-02

## FAQ

### Can AI approve marketing claims?

No. It can find likely problems and suggest wording from your register. Approving a regulated claim stays with a named compliance reviewer, because a model cannot judge whether evidence is adequate.

### Why not just use a banned-word list?

Word lists catch exact phrases only, so a claim implied by tone or a product name can pass. Use lists for required disclosures, and the model as a second reader for meaning.

### What should a claims register contain?

Each entry needs the approved wording, product, market, evidence reference, expiry date and owner. Keep refused claims too, with the reason, as the EU food register does [1](#source-1).

## Sources

1. European Commission, EU Register of nutrition and health claims: https://food.ec.europa.eu/food-safety/labelling-and-nutrition/nutrition-and-health-claims/eu-register-health-claims_en
2. Advertising Standards Authority and CAP, Substantiation: https://www.asa.org.uk/advice-online/substantiation.html
3. US Federal Trade Commission, Health Products Compliance Guidance: https://www.ftc.gov/business-guidance/resources/health-products-compliance-guidance
4. FCA, Regulating financial promotions and adverts: https://www.fca.org.uk/firms/financial-promotions-adverts
5. FCA, Risk warnings for mainstream investments: https://www.fca.org.uk/firms/risk-warnings-mainstream-investments
6. EU AI Act, Annex III: https://artificialintelligenceact.eu/annex/3/
7. Mayer Brown, EU green transition directive (27 October 2026): https://www.mayerbrown.com/en/insights/publications/2026/10/the-eu-directive-on-empowering-consumers-for-the-green-transition-two-years-on-what-it-requires-and-where-the-member-states-stand
8. Meta, Agents Rule of Two (31 October 2025): https://ai.meta.com/blog/practical-ai-agent-security/

This article is general information, not legal advice.

## Further reading

- [Marketing Claims Compliance Check: Review Copy With an AI Agent on Rails](https://privycdp.org/ai/rails/use-cases/copy-claims-compliance-check)
- [AI Agents for Lifecycle and CRM Marketing: The Playbook on Rails](https://privycdp.org/ai/rails/functions/lifecycle-marketing)
- [Orchestrator agent](https://privycdp.org/learn/glossary/orchestrator-agent)
- [AI Agent Rails: An Operating Model for Running Agents on Customer Data](https://privycdp.org/ai/rails)
- [AI Agent Rail Card Template: YAML, Spreadsheet and JSON Schema](https://privycdp.org/tools/agent-rail-card)