What it does reliably
Answering, first ring, at any hour and any number of calls at once. This is the part that is genuinely solved. There is no queue, no bad Monday, and no difference between the first call of the day and the four hundredth.
Answering questions from material it has been given. Hours, location, parking, services, policies, what to bring, whether you are open on a public holiday. This works well and is the bulk of the value, and it works in proportion to how good the material is, which is a point the next guide is entirely about.
Capturing a caller properly. Name, number, what they wanted, spelled back and confirmed. A written record of that is worth more than a voicemail, because somebody can act on it in fifteen seconds without listening to anything.
Holding a consistent line. It says the same thing to the caller at eight in the evening as at eight in the morning, and it does not get worn down by the fourth person asking the same question.
What it does with caveats
Reading a live schedule. An agent connected to a real scheduling system can tell a caller whether the Saturday class has space. Whether yours can depends entirely on whether that system has an interface anyone can connect to, and a large share of small business software does not. Ask the specific question about your specific system rather than the general one about integrations.
Handling an accent, a bad line, or a caller in a car. Speech recognition today is good and not perfect, and the failure is usually not a wrong answer but a repeated question, which callers tolerate about twice.
Long or wandering calls. An agent is at its best in a two minute exchange with a clear purpose. A caller who tells a story for four minutes before getting to the question is a harder problem, and the honest outcome is often a handoff.
- Scheduling: ask whether it reads your calendar, writes to it, or neither. These are three different products described with the same word.
- Languages: assume one unless the vendor names the others, and ask whether a switch is automatic or a separate line.
- Numbers and spellings: ask how it handles a surname it has never heard. Confirmation back to the caller is the tell.
What it does not do
It does not exercise judgement. It follows the rules it was given, which is exactly what you want on a Tuesday and exactly what fails on the call that does not fit any rule. There is no version of this that reads a room.
It does not know anything you did not tell it. This is the most commonly misunderstood point in the category. An agent does not know your prices, your policies or your availability by being intelligent; it knows them because somebody wrote them down. Every gap in that material is a question it has to refuse.
It does not take responsibility. When a caller asks whether something is an emergency, a well built agent for a clinical business refuses and routes them, because the alternative is a machine making a medical judgement. Any vendor whose demo answers a clinical or legal question directly is showing you their biggest liability as though it were a feature.
It does not replace the person at the desk. It replaces the ringing, the voicemail, and the calls that arrive while that person is already busy. A business that fires its front desk on the strength of one of these is going to have a bad quarter.
The failure modes to ask about by name
These are the four ways this technology goes wrong in production. A vendor who has run one of these in the wild will have an answer to each; a vendor who has not will be surprised by the question.
- Confident wrong answers. Ask what happens when the question is close to something in the material but not the same. The safe behaviour is a refusal, and refusals are what you should ask to see demonstrated.
- The endless loop. A caller who is misheard twice needs an exit. Ask what triggers a handoff to a message or a person, and after how many turns.
- The runaway bill. Anything billed per minute has an incentive problem and a spend risk. Ask about a hard cap per call and per day, and whether it exists in the product or in a promise.
- The stale answer. Your hours change and the agent keeps quoting the old ones for a month. Ask how the material gets updated and who does it, and how you would find out that it had gone stale.
How to test one in ten minutes
Do not ask it the easy questions. Every demo survives the easy questions. Ask the three you would be most embarrassed to have answered wrong: the pricing question you never quote over the phone, the question about a policy that has an exception, and the one where a caller asks for advice you are not licensed to give over the phone.
A good result is not three correct answers. A good result is a clean refusal on the ones that should be refused, and a clean capture of the caller so a person can follow up. If it refuses well, the rest is a matter of writing better material. If it guesses well, you have learned nothing about the day it guesses badly.
The short version
The gap between a useful agent and an embarrassing one is almost entirely in what happens when it does not know. Test that, not the greeting.