The team page has fifteen names on it. Audrey, Cecily, Delia, Greta, Henry, Ivor, Margot, Owen, Porter, Quentin, Rhett, Sable, Silas, Tobias, Wren. None of them is a person.
The first sentence on that page says so. That was not a legal decision, it was the only version I could publish and still be able to answer the phone honestly.
Why name them at all
The alternative is describing them as a system, which is accurate and useless. "Automated proposal generation" tells you nothing about what happens on a Tuesday. "Quentin drafts the proposal, Cecily checks it for anything internal, and it does not leave the building until both have passed it" tells you exactly what happens, and the names are what make the sequence legible.
Naming them also makes the boundaries obvious. If Quentin drafts and Audrey audits and Audrey cannot change anything, that separation is easy to say and easy to check. It is a much harder sentence to write about "the system."
Why the first sentence has to say what they are
Because somebody will find out. A roster of invented colleagues with stock photographs is a lie with a discovery date, and the discovery happens in the worst available conversation: a client asking to speak to Margot.
There is a second reason that matters more. A one-person company with fifteen agents is genuinely unusual and genuinely interesting. Presenting them as staff converts the most distinctive true thing about the business into an ordinary lie. It is a bad trade even for someone with no scruples about it.
The rules, and where they live
Three of them, and all three are enforced in code rather than written in a policy document, because a policy is a request and a function is a constraint.
An agent signs as itself, never as me. Every client-facing message goes out under the agent's own name with a line saying it is an automated agent and that a person will read a reply. Enforced at the single point where anything is recorded as sent, so there is no path around it.
Internal reasoning never leaves the building. The proposal agent once drafted twenty documents that all carried internal notes, hedges and commentary meant for me. None went out, because a separate agent whose only job is compliance reads everything at the choke point.
The auditors cannot fix anything. The agents that inspect the live properties are read-only. Separating the finder from the fixer is the cheapest safety property available and it costs one architectural decision. It also creates a failure mode nobody warns you about.
What this costs
Honestly, some credibility. "One person and fifteen scripts" sounds smaller than "our team," and there are prospects who would have been more impressed by the second sentence.
I would rather have the ones who are impressed by the first. A client who discovers the truth later becomes a client who wonders what else was presented optimistically, and there is no version of that conversation that ends well.
If you are building something similar
Say what it is on the page where you say what it does. Put the disclosure inside the function that sends, not inside a document that describes what should happen. And if you find yourself writing a sentence about your company that you would be uncomfortable having a client read back to you, that sentence is the problem, not their reaction to it.
What the agents actually get wrong, at length, is here.
Common questions
Should a company disclose that its team members are AI agents?
Yes, and on the page where they are named rather than in a footnote. A roster of invented colleagues is a lie with a discovery date, and the discovery arrives in the worst available conversation, which is a client asking to speak to one of them. There is also a commercial argument: a one-person company running fifteen agents is genuinely unusual, and presenting them as staff converts the most distinctive true thing about the business into an ordinary lie.
Should AI agents sign emails with a human name?
No. An agent should sign as itself and state that it is automated, with a line telling the recipient that a person will read a reply. Enforcing that at the single point where a message is recorded as sent, rather than in a policy document, is the difference between a rule and a request.
Why separate the AI agents that find problems from the ones that fix them?
Because it is the cheapest safety property available and it costs one architectural decision. An agent that can only report cannot cause the failure it was inspecting for. It does introduce its own failure mode, which is findings that pile up unfixed, and that is worth knowing before adopting the pattern.