8 min read

C29
Operation · Governance & Compliance

Claims About Your Agent

The riskiest sentence about your agent is not a lie — it is a number you published and cannot reproduce on demand.

Most AI-marketing enforcement is not about fabrication; it is about substantiation, and it is landing on business-to-business vendors. When the FTC finalised its orders against Cox Media Group and two suppliers on 27 August 2026 — $930,000 in total, $880,000 from CMG and $25,000 each from MindSift and 1010 Digital Works — it charged the two suppliers with providing the "means and instrumentalities" to deceive. The operational consequence is specific: every capability number on your site needs a dated, re-runnable eval behind it, and so does every number your customer repeats about you.

Step 1Read the docket before you read the think-pieces: the charge is usually "unsubstantiated", and the defendant is usually a vendor.

Two 2026 cases set the shape. In the CMG matter the FTC said the "Active Listening" product did not listen to consumers' conversations or use voice data at all; the sellers were reselling email lists bought from data brokers at a markup. That is a fabrication case. In Air AI, settled on 24 March 2026, the FTC's August 2025 complaint alleged false earnings claims and misrepresented refund guarantees under the Telemarketing Sales Rule and the Business Opportunity Rule — the stipulated order carries an $18 million judgment, largely suspended for inability to pay, with $50,000 actually payable, and bars earnings claims made without adequate substantiation.

The second shape is the one that should worry an engineering team, because "we had no evidence for it" is a complete cause of action on its own. You do not need to have known the claim was false. DLA Piper counted the CMG action as the thirteenth AI-washing case the FTC had filed since 2024, and noted that seven of the last eight involved claims made to other businesses rather than consumers — which is to say, claims that look exactly like the ones on an agent platform's pricing page.

Enforcement posture moves and the direction is not one-way: the Commission set aside its Rytr consent order in December 2025, a case that had turned on an AI tool's capacity for misuse rather than on a false factual claim. Read that as a narrowing of the theory, not a retreat — false and unsupported factual claims are exactly the part that survived.

Step 2"Means and instrumentalities" is why your customer's claim is your exposure too.

MindSift and 1010 Digital Works were not the companies with the deceptive marketing page. They supplied the capability that CMG then sold, and the FTC charged them for providing the means to deceive. If you sell an agent platform, that is the theory to internalise: the deck your customer builds on top of your datasheet is downstream of a document you wrote, and the number in it is traceable to you.

This flips the usual direction of a vendor datasheet. Teams write them to be as flattering as the lawyers will allow, on the assumption that the customer's own compliance function is the backstop. It is not — the customer's claim and yours are both yours. Three concrete changes follow:

  • Scope every number to the configuration it was measured on. "94% resolution" is not a property of your agent; it is a property of one model version, one tool set, one dataset and one grader. Ship the configuration with the number or do not ship the number.
  • Put the denominator in the sentence, not the footnote. "Resolves 94% of tickets" and "resolves 94% of the 61% of tickets it chooses to attempt" are different claims, and only one of them is yours to defend.
  • Give your customers a claims sheet they are allowed to copy. The cheapest control over what a reseller says about you is handing them pre-substantiated sentences, each with the eval and date attached. See third-party model and vendor risk for the same relationship viewed from the buyer's side.

Step 3Substantiation is an artefact, not an argument — keep a claim register that points at runs.

The practical test of substantiation is whether you can produce, on a regulator's or a customer's timeline, the evidence that existed when you published. Reconstructing it later from a model you have since upgraded is the failure mode. Treat each public claim as a tracked object with an owner and an expiry, stored next to the code that proves it:

# claims.yml — one entry per sentence that appears in public
- id: resolution-rate-2026q4
  # the exact published wording, copied, not paraphrased
  text: "Resolves 94% of tier-1 support tickets without escalation."
  surfaces: [website-pricing, sales-deck-p7, g2-profile]

  # what makes it true — must be re-runnable TODAY, not a screenshot
  evidence:
    eval: evals/support/tier1_resolution       # in-repo, versioned
    run: 2026-10-14T09:12Z                     # trace id, not a spreadsheet
    n: 1842                                    # sample size and how drawn
    sampling: "random 2026-09 tickets, 3 tenants"
    grader: human-adjudicated                  # LLM-judge? say so, and cite its calibration
    denominator: "all inbound tier-1, incl. ones the agent declined"

  # what invalidates it — the half teams never write down
  pinned_to:
    model: claude-opus-5-5@2026-09-02
    tools: catalog-v14
  expires: 2027-01-14                          # 90 days, or next model change
  owner: "support-platform-team"
  approved_by: "legal-review-2026-10-16"

The pinned_to and expires fields are what make the register more than a filing cabinet. A capability claim is only as current as the model under it, and a model migration silently invalidates every number measured on the old one. Wire the expiry into the same pipeline as your eval set maintenance: a claim whose eval has not been re-run inside its window is a defect, routed like any other.

Step 4Four claim shapes fail predictably. Three of them are about agents specifically.

  • Accuracy without a denominator. Agents get to decline, escalate and ask a clarifying question — all of which inflate a naive success rate by removing the hard cases from the bottom of the fraction. If your eval scores only attempted tasks, your published number is measuring your routing, not your resolution.
  • Autonomy claims. "Fully autonomous" is a factual claim about what happens with no human in the request path, and it is falsified by a single approval gate. If a human approves refunds over $50, you have a supervised agent; say so. Autonomy levels exist partly so that marketing has vocabulary that survives a deposition.
  • Comparative claims. "2× faster than <competitor>" invites the competitor's lawyers rather than the regulator's, and the evidentiary bar is the same. You need their configuration documented, which you almost never have.
  • Consent and data-handling claims. The CMG orders bar misrepresenting how voice data and consent were handled, and the FTC's position on the mechanism is worth quoting at your product review: clicking through mandatory terms of service when downloading an app does not constitute opt-in consent. If your agent ingests customer conversations for training or memory, "users consented" is a claim about a UI flow someone has to be able to show. See delegated access and consent records.

The quiet one in this list is the first. A team that adds a confidence threshold to improve quality has, in the same deploy, changed the denominator of its published accuracy number — upward, because the agent now declines more. Nobody notices, because the dashboard moved in the right direction. Put the attempt rate on the same chart as the success rate and the artefact becomes visible; see scope conformance evals.

Step 5In the EU the disclosure itself is now a claim, and silence is the violation.

The AI Act's Article 50 transparency duties have applied since 2 August 2026, enforced by national market surveillance authorities, with penalties up to €15 million or 3% of worldwide annual turnover, whichever is higher. The duties are narrow and mechanical: tell people when they are interacting with an AI system, mark synthetic content machine-readably, disclose deepfakes and AI-generated text published as news. The one transitional concession is on marking — generative systems already on the market have until 2 December 2026 — and content generated before 2 August 2026 but published after it still needs labelling.

For an agent product this changes the register of the compliance question. A US-style AI-washing exposure comes from over-claiming; an Article 50 exposure comes from under-disclosing — and the two pull in opposite directions on the same sentence. The marketing instinct to let a support agent read as human is now a documented violation with a number attached. Build the disclosure into the surface rather than the copy deck: disclosure and content provenance covers the marking machinery, disclosing the agent on a call covers the hardest surface to retrofit, and the EU AI Act for agents covers the rest of the obligation stack, whose high-risk tiers were pushed out to December 2027 and August 2028 while Article 50 was not.

Step 6Make the claim register the thing sales copies from, or it will be ignored.

A register nobody reads is worse than nothing, because it proves you had a process and bypassed it. The governance only holds if the path of least resistance runs through it, which means the register has to be the source for the artefacts people actually use — the one-pager, the RFP answer bank, the security questionnaire response. Four checks make that real:

  • A pre-publication gate owned by the team that can re-run the eval. Not legal alone; legal cannot tell whether the denominator is honest.
  • A diff on every surface. When the pricing page changes, the register entry's text field should fail to match. This is a string comparison, not a review meeting.
  • A kill path. When an eval re-run drops below the published number, you need an agreed answer to "what comes down first" and who may say so without a committee.
  • An answer bank for the questions you will be asked under contract. The RFP and questionnaire surface is where unsubstantiated claims are most often first written down, by someone with a deadline and no access to the eval.

Route the register's history into the same store as your audit trails, and keep the superseded entries: being able to show what you believed, when, and on what evidence is the whole asset. Related: impact assessments for the pre-deployment version of the same discipline, and insurance and liability for what an unsupported claim does to a policy.

Do this today: open your pricing page and your most-used sales deck, and write down every number and every absolute word ("fully", "automatically", "never", "all") about the agent. For each one, name the eval you would re-run this afternoon to defend it. The ones with no answer are your actual exposure, and the fix is almost never to find evidence — it is to rewrite the sentence so it says what you can show. Delete first, substantiate second.