Quick answerIf a fraudster clones your brand voice or your AI agent's persona to run scam conversations with your customers, the first move is not a support ticket, it is a fraud incident: verify the impersonation is real, lock down every channel a customer could mistake for your legitimate agent, and push an authenticated verification signal (a real order number lookup, a signed session token, a published channel list) that a clone cannot fake. This is a different problem from your own agent making a real mistake or drifting off brand tone. Here, the agent making the mistake is not yours at all.
Why this is a fraud response, not a support or QA response
Two existing failure modes get confused with this one, and the confusion slows down the response. A viral PR crisis is your own agent saying something genuinely wrong in public, which you own and can fix by correcting the agent. Off-brand tone drift is your own agent's real answers sounding inconsistent, which you fix through internal quality control. Impersonation fraud is neither. It is a third party standing up a fake chatbot, a cloned voice line, or a spoofed social account that looks and sounds like your real agent, and using it to extract payment details, account credentials, or gift card codes from people who reasonably believe they are talking to you.
Because the attacker is external, your agent's own logs and transcripts are useless for detecting it. You will not find the incident by reviewing your agent's conversations, since your agent never had them. You find it from customer reports, brand monitoring, or a support ticket that starts with "I already paid the fee you asked for."
Step 1: confirm it is actually a clone, not a legitimate but confusing channel
Before treating this as fraud, rule out an internal explanation: a reseller or partner running their own branded bot, a regional team that stood up a channel you were not looped in on, or an old, deprecated bot that was never decommissioned. Check your own channel inventory first. If none of those account for it, treat it as external impersonation.
Step 2: verify with evidence a clone cannot reproduce
A cloned voice or a copy-pasted chat script can mimic tone and wording almost perfectly, but it cannot originate a real, live lookup against your backend. Build (or confirm you already have) a lightweight verification signal your legitimate agent exposes and a fraudster's clone cannot: a real-time order status pull tied to an account, a one-time code sent through a channel you control, or a published, dated list of the only domains and phone numbers your agent will ever use. Put that list somewhere customers can check independently of the channel they are currently on, since if the fraudulent channel is the only place they'd see it, it is useless.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Step 3: lock down and notify in parallel, not sequentially
Once confirmed, three things happen at once, not one after another: report the impersonating domain, number, or social account to the platform hosting it (most have expedited fraud takedown paths), push a notice through your own verified channels warning customers about the specific impersonation pattern (what it asks for, what it looks like), and loop in whoever owns How to Respond When a Competitor Runs a Disinformation or Fake-Review Campaign Against Your AI Agent, since the takedown and public-communication playbook for a bad-faith external actor overlaps heavily even though the underlying tactic here is impersonation rather than disinformation.
Step 4: give customers a standing way to verify they are talking to the real thing
The single most durable fix is not reactive, it is a permanent verification affordance baked into your legitimate agent's UI: a visible badge, a session fingerprint customers can check, or clear sourcing of every claim the way described in How to Design a Trustworthy AI Agent UI: Sources, Confidence, and Labeling. A clone can copy your agent's words. It is much harder for a clone to fake a live, verifiable trust signal that is specific to a real, authenticated session.
What this is not
It is not the moment to relitigate your agent's tone guidelines, and it is not a moment to treat a real customer-facing mistake your own agent made the same way you would handle a viral internal error, covered separately in How to Handle a Viral Social Media Moment When Your AI Agent Makes a Public Mistake. Conflating the two slows the fraud response down, because a fraud takedown and a legal notice move on a different track than an internal correction and apology do.
FAQ
How is this different from my AI agent just giving a bad answer? A bad answer comes from your own system and you can fix it by correcting the agent. An impersonation clone is not your system at all, so there is nothing in your own logs to correct, only an external actor to report and a verification gap to close.
Can we legally force a takedown of a cloned voice or chatbot? In most jurisdictions, trademark and brand impersonation used to solicit payment gives you a fast path through the hosting platform's fraud team and, if needed, a cease-and-desist, but the platform-level report is almost always faster than a legal filing and should go out first.
Should we warn all customers or only the ones targeted? Warn broadly through your verified channels once the impersonation is confirmed active, since you usually cannot tell in advance who has been targeted, and a general notice with a clear "how to verify us" instruction protects customers who have not been contacted yet.