Search for an AI receptionist and you will get a dozen lists of the ten best AI receptionists. Every one of them is published by a company that sells an AI receptionist, and every one of them ranks its own product somewhere near the top. That is not a scandal, it is just what the search results are made of, and it means the genuinely useful question goes unanswered.
The useful question is not which one is best. It is: which parts of intake does this technology actually hold up on, and which parts will quietly fail in a way you will not notice for a month.
We run these systems for clients. Here is where they break.
1. The handoff, not the conversation
The conversation is the part everyone demos, and it is the part that works. Modern voice agents answer immediately, sound fine, and handle a straightforward "what are your hours, can you come Tuesday" without difficulty.
The failure is at the seam. The call ends, and something has to happen: a record written, a job created, a human notified, a calendar hold placed. That handoff is ordinary integration work, it is where the value actually is, and it is the part the demo skips.
What we see when we inherit a setup: calls handled beautifully and appointments that exist only in a transcript nobody reads. The AI did its job. Nothing downstream was connected, so the job was pointless.
2. Confident wrong answers about your business
A voice agent will not say "I don't know" unless it has been built to. Asked something outside what it was given — do you service this zip code, do you take this insurance, can you fit a unit this size — it will produce something plausible.
Plausible and wrong is worse than nothing here, because the customer acts on it. They take the day off for an appointment in an area you do not cover. They expect a price you do not charge. Someone in your office then has to make an apologetic phone call, which costs more goodwill than the missed answer would have.
This is fixable, but the fix is unglamorous: enumerate the questions the agent is allowed to answer authoritatively, give it the real answers, and make everything else an explicit escalation. That means somebody has to sit down and write out your service area, your exclusions, your actual pricing rules. Most of the work in a good deployment is this, and it is why deployments that skip it disappoint.
3. Qualification that optimises for the wrong thing
Ask an AI to qualify leads and, unless told otherwise, it will optimise for completing the form. It will get an answer to every field, because that is what looks like success.
But qualification is a judgement about whether this is a customer worth your time, and the signal for that is often in what the caller says rather than in which boxes got ticked. Hesitation about budget. A landlord asking on behalf of a tenant. Someone comparing you against two other quotes with a decision already made. A job just outside what you actually want to do.
A scored intake form full of complete answers can be entirely accurate and still send your best technician to the wrong job. The fix is partly better instructions and partly accepting that some qualification is a human read, and building the escalation rather than pretending the model will get there.
4. Compliance, at the moment the call connects
This one is not a quality issue, it is a legal one, and it is the one most likely to be missed entirely because it is not a marketing question.
Most AI voice systems transcribe. Transcription is recording. A number of states — Maryland among them, which is where we are — require all parties to a call to consent to it being recorded, not just one. That obligation attaches the instant the call connects, before any of the clever part happens.
If you are running a voice agent in a two-party-consent state, or taking calls from one, the disclosure has to be built into the opening of every call. That is a design constraint on the greeting, not a checkbox in a settings page, and it is the reason we treat "which states do your callers come from" as an early question rather than a detail.
What this actually means for deploying it
None of the above is an argument against AI in intake. We use it and we sell it. It is an argument against the framing in the roundups, which is that you pick a tool and switch it on.
What works, in our experience:
Start where the alternative is nothing. After-hours and overflow calls are currently going to voicemail, and voicemail converts terribly. An AI that captures the detail and books a callback is competing against silence, and it wins easily. That is a much better first deployment than replacing a receptionist who is good at their job.
Give it a narrow, well-specified job. Capture the situation, confirm the address, offer real appointment slots, escalate anything unusual. Systems that try to handle everything fail in the ways above; systems with an explicit escalation path degrade gracefully.
Instrument the escalations, not just the successes. The number that tells you whether it is working is how often it handed off and whether those handoffs were caught. A system reporting a high resolution rate may be resolving calls it should have escalated.
Keep a person in the loop where judgement lives. Qualification, pricing exceptions, anything where an unhappy customer is the failure mode. The point is not that AI cannot do these — it is that the cost of it being wrong is paid by your reputation, and the cost of a human doing it is an hour a week.
The pattern generalises past intake. The parts of an operation that AI takes over cleanly are the ones with a clear input, a clear output and a low cost of being wrong. The parts that need a person are the ones where being confidently wrong is expensive. Most disappointing AI projects we are asked to rescue put the technology on the wrong side of that line, then blamed the model.
Is an AI receptionist better than an answering service?
They fail differently. An answering service gives you a human who can exercise judgement but may know little about your business. An AI knows exactly what you told it and never gets tired, but cannot improvise. For after-hours capture, the AI is usually better value; for complex or sensitive calls, the human is.
How do we know if it is working?
Track booked jobs from calls it handled, not calls handled. A system can answer everything and book nothing. Also track how often it escalated and whether anyone picked those up.
Will customers know they are talking to AI?
Usually, and increasingly they do not mind if it is fast and useful. What they object to is being trapped — an agent with no path to a person, on a call that needed one.
What should we not point it at?
Anything where a confidently wrong answer costs real money or goodwill: firm pricing on non-standard jobs, eligibility questions, complaints, and anything a lawyer would want to see the wording of.
More in AI operations. Related: the five ways missed-call text-back breaks, which is the same lesson at lower cost.