What an AI agent inside a franchise system must never answer

Two releases and one order, in one week

On 5 October 2026 FranConnect published its October release notes with two AI agents for franchise operations: a franchisee support agent answering from brand-approved content with citations and permission awareness, and an agreement assistant that extracts key terms from franchise and lease agreements for an administrator to review and correct before the values are saved.

On the same day, the FTC announced a settlement with Premier Franchising Group, franchisor of Premier Martial Arts, and Franchise Fastlane, its former sales organisation: $3,875,424 against the franchisor, partially suspended on a $650,000 payment, plus $1.2 million against the sales organisation, for $1.85 million recovered. More than 200 people had paid initial fees of $49,500 or more. On 6 October the International Franchise Association said the Franchise Rule applies to everyone involved in selling a franchise, including sales organisations, and described the case as the first FTC action of its kind against a third-party seller.

A day later, a franchise performance coach published a piece arguing that AI in franchising should start with administration, surface insights rather than decisions, and keep humans in the loop where relationships matter.

These three are a single argument from three directions.

The architecture that survived scrutiny

Look at what the two shipped agents were not allowed to do. The support agent does not answer from general knowledge; it answers from the brand’s approved content and shows which document it drew on. It also does not answer beyond the asking user’s permissions, so a prospective franchisee and a ten-year operator do not get the same answer from the same system. The contract agent does not write to the record; it produces a draft that a person corrects first.

Both constraints look like product friction. They are actually liability design. In a franchise system, an answer about what the brand standard requires is an answer about a contract. A model that invents a plausible version of the operations manual has created a document the franchisor will be held to. Citation is what makes the answer auditable; permission scoping is what stops the system from leaking material that a given user has no contractual right to; human review is what stops a machine reading from quietly becoming the authoritative lease term.

Why the earnings question is different

Now put the FTC complaint against that architecture. Two of the alleged failures translate directly into machine behaviour.

The first is financial representations made outside the disclosure document. In an automated sales funnel, this is not an edge case. A chat assistant asked what a franchisee earns will answer if it can, because answering is what it is for. If the number it produces is not the number in the disclosure document, the franchisor has just made an unlawful representation at the speed of a web form, repeatedly, with a log.

The second is subtler and more interesting. The complaint describes a failure to disclose that existing franchisees had larger studios and martial arts experience while incoming ones did not. That is a comparability failure, and comparability failure is the default output of a statistical system. Give a model a set of unit results and ask what a new franchisee can expect, and it will give you a central tendency. The cohort behind that central tendency, how long those units have traded, how big they are, who ran them, is exactly the information that gets averaged away.

So the rule for an agent in a franchise system is narrower than a content policy. It is not do not lie about earnings. It is: do not generate an earnings or performance figure at all. Retrieve it, verbatim, with its cohort description, from the disclosure document, or decline and hand the person to someone who can answer. There is no version of generating this number that is safe, because the generation itself is the representation.

Five guardrails you can write into any deployment

Concrete enough to put in a deployment ticket:

• Hard refusal list. Earnings, profit, payback period, revenue per unit, cost to open and any projection are retrieval-only. The agent quotes the disclosure document and names it, or says it cannot answer and routes to a human.
• Cohort travels with the number. If the system can state a figure but cannot state how many units, over what period, and how comparable they are, it does not state the figure.
• Source on every claim. Any operational answer cites the approved document it came from, so a dispute can be resolved by reading, not by reconstructing a conversation.
• Permission scoping at the question, not the answer. A prospective franchisee and a current one are asking in different legal postures; the system should know which before it composes anything.
• Logged and reviewable. Every automated answer to a prospect is retained with its sources, because the claim you cannot produce later is the claim you cannot defend.

None of this is slow. The support agent in this week’s release already works this way. The point is that the architecture is not a nicety; it is the thing the order is about.

What we do not know

We do not know how accurate the agreement extraction actually is. The release notes describe capability and publish no accuracy rate, no deflection rate and no customer numbers, so nothing here should be read as evidence that it performs well.

We do not know whether generative tools played any part in the Premier Martial Arts case. The FTC complaint as reported describes human sales conduct; the connection to AI deployment in this analysis is ours, drawn from the structure of the violations, not from anything the Commission said about technology.

We do not know how often automated franchise sales channels currently emit earnings figures outside a disclosure document, because no one publishes that measurement.

And we do not know how regulators outside the United States will treat the same conduct. The Franchise Rule is a US instrument, and most Asian markets have disclosure regimes that differ in both scope and enforcement.

This is general regulatory and market information, not legal advice.

Get next week's brief by email. Free.