AI Voice Agent Red Flags: 6 Signs the Phone Line Isn't Yours

Six AI voice agent red flags from a vendor shutdown and live builds: numbers, wallets, exports, silent failures, test calls and disclosure, and what to ask.

By Carlos Familia, CRM & AI automation expert. Updated October 15, 2026. 6 min.

On October 5, 2026, an agency owner posted a warning on r/gohighlevel. A voice AI vendor had shut down around September 25, taken its site offline and let its staff go. One of the poster's clients only found out because his texts started failing.

The advice for anyone affected came as a short list. Copy the agent prompts, mid-call actions and webhook URLs while the vendor's dashboard still loads. Find out where your number lives, because one the vendor bought can disappear with its account. And check the HighLevel wallet and workflows, since theirs sat under the vendor's agency and premium actions failed when that wallet ran dry.

A red flag, then, is anything that leaves your phone line depending on an account you do not control, or that lets a failure go unreported. These six cover most of what I have seen, each with the question that tests for it. If you are still choosing a builder, pair this with the AI voice agent developer interview questions.

Red flag 1: The number was bought inside the vendor's account

What it looked like: the poster's second point. A number the vendor bought can vanish with the vendor's account. If your main line only forwards to it, you need a new number to forward to.

Ask this instead: "Whose name is on the phone number, and can I see its billing page?" Retell's custom telephony documentation describes setting up a SIP trunk with your own telephony provider and importing your own number into Retell, which is one way to keep the number apart from whoever builds the agent.

Red flag 2: There is no copy of the agent outside the platform

What it looked like: the replies to that thread were mostly about what people wished they had saved. One said to grab call transcripts and recordings while the dashboard still loaded, then replay the nastiest calls (emergencies, price hagglers, mumbled addresses) against the replacement before switching the forwarding. The same reply said to keep prompts and business facts in your own documents, so the next shutdown is "a rewire, not a rebuild."

Ask this instead: "At handover, which exports do I get: prompt, settings, tool list with endpoints, knowledge base, recordings and transcripts? Where do they live today?" If the answer is "in the platform," the copy does not exist yet.

Red flag 3: One wallet pays for every client's premium steps

What it looked like: in the thread, the agency's workflows and wallet sat under the vendor's agency, and its premium actions began failing when that wallet emptied. I know how quietly that happens. In GoHighLevel, a premium action such as a custom webhook is skipped without an error when the agency wallet is empty, and the reason shows only inside that step's details. On a roofing build in Kentucky, email stopped for the same kind of reason: the email wallet had no balance.

Ask this instead: "Which wallet or prepaid balance pays for calls, texts and premium steps? Is it separate per business, and who gets the low-balance alert?" A builder who runs every client from one balance has chosen to give them all the same outage.

Red flag 4: "Success" in the log is the only signal

What it looked like: on a speed-to-lead dialer I built in Pennsylvania, the prepaid voice credit hit zero and the runs kept finishing green. Behind them were 79 dials that failed and 11 Facebook leads who never heard from anyone.

ElevenLabs' post-call webhook documentation describes another quiet path. After ten straight failed deliveries, the platform turns a webhook off, provided its last good delivery is over a week old or it never had one. Different cause, same silence.

Ask this instead: "What alerts me when no calls arrive, when dials fail and when a webhook goes quiet? Show me an alert that fired, and the evidence it carried." The watchdog I added to that dialer alerts only with evidence, and it now catches this exact case.

Red flag 5: It has only met polite test callers

What it looked like: a reply in the same thread put it plainly: "real calls break things scripted tests never do," such as bad lines garbling words. My own builds agree. On a tattoo studio's phone agent, fast transcription heard "a tattoo" as "account two," and the agent once invented a location in another country. Accurate transcription mode, boosted phrases and a keypad fallback fixed the hearing, and stating the studio's real details fixed the rest.

Ask this instead: "Show me a test where it failed and how you fixed it. Has anyone replayed messy recorded calls against it, and was the transfer tried on a real phone?"

Red flag 6: Nobody decided what the caller is told

What it looked like: a human-sounding name, no disclosure, and nobody able to say whether calls are recorded.

California's Penal Code section 632 punishes intentionally recording a confidential communication without the consent of all parties, with a fine of up to $2,500 per violation, jail, or both. On California calls I work to two-party consent, so the agent tells callers it is an AI and that the call is recorded. Other states set different rules, and the wording needs your counsel's approval.

Ask this instead: "How does the caller learn it is an AI and that the call may be recorded, and who approved that wording?"

What if you are already stuck with one?

The thread doubles as a recovery plan, and the order matters. While the dashboard still loads:

1. Copy the prompts, mid-call actions, knowledge base and webhook URLs into your own documents.
2. Download the recordings and transcripts. They are your best test material for whatever replaces the agent.
3. Find out who owns the phone number. If the vendor bought it, get a number you hold and forward your main line there.
4. Check which wallet pays for texts and premium steps, and which agency account your workflows live under.
5. Ask whether your sub-account can move to an agency account you control. One commenter noted that, depending on setup, a transfer may keep contacts, workflows and funnels.
6. Replay the worst recorded calls against the replacement before you switch the forwarding.

Flags 1 to 3 cost minutes to check today and days to repair later. Once a vendor passes them, the next questions are cost, in what an AI receptionist costs, and upkeep, in AI agent maintenance. Marketplace buyers should also read my red flags for an AI automation freelancer, and the wider hiring guide is AI automation consultant. To check a vendor you already use, bring its answers to a call on the calendar below.

Six lines to check before you sign

• The number, the voice account, the wallet and the workflow platform are in my name.
• I get exports of the prompts, tools, knowledge base and recordings.
• Each business has its own balance, with a low-balance alert that reaches me.
• An alert fires on silence and failed dials, not only on errors.
• The agent was tested on messy calls, and the transfer was tried on a real phone.
• The AI disclosure and the recording notice are written down and approved.

About the author

Carlos Familia, CRM & AI automation expert. GoHighLevel, n8n, AI agents, servers. Full-stack developer.

I build CRM systems in GoHighLevel, n8n automations on my own servers and AI voice and chat agents for service businesses. Full-stack developer: I write the workflow, the API integration, the database and the server it runs on, and I test every build with real journeys before handoff.