An answering service or an AI: how to actually choose

The framing that makes this decision hard is treating an answering service and an AI agent as two suppliers of the same thing at different prices. They are two different products. An answering service sells coverage: somebody picks up, takes a message, gets it to you. An agent sells answers: the caller gets the information they rang for, and only the calls that genuinely need a person become a message.

That difference decides the purchase more than price does, because it maps onto a question about your own callers that only you can answer: when the phone rings, does the caller need something known, or something decided?

Spend a week writing down what people actually rang about. One line per call, no categories decided in advance. At the end of the week, sort the lines into three piles.

The first pile is questions with a knowable answer: hours, location, whether you have space on Saturday, whether you take a particular insurance, what a service involves. The second pile is transactions: booking, rescheduling, cancelling. The third pile is anything requiring judgement: an upset customer, an unusual situation, a decision about whether something is urgent.

If pile one is the biggest, an agent is likely the better purchase, because a message-taking service will simply write those questions down and hand them back to you. If pile three is the biggest, buy the people. If pile two is the biggest, ask both vendors very specifically what they do about it, because this is where both categories are weakest and where the marketing is loosest.

A person can tell that the caller is not fine. That sentence covers most of the real advantage. Tone, hesitation, the difference between a mild complaint and someone about to leave a review that costs you a month of marketing: a human hears it, and nothing available today reliably does.

A person also improvises. When the situation is not one anybody anticipated, a receptionist finds something reasonable to do. An agent follows its rules, and when the situation is off the edge of its rules the best case is that it says so and takes a message.

Then there is the caller who will simply not accept talking to a machine. That population is smaller than it was two years ago and it is not zero, and in some trades and some age brackets it is significant. You know your callers.

Concurrency, first. Ten calls at once are ten answered calls, and no staffing model does that at a small business price. If your problem is a spike rather than a gap, this is decisive on its own.

Depth of knowledge, second, and this one surprises people. A briefed agent can be given every service you offer, every policy, every price you are willing to state, and it will not have an off day or a first week. A service’s receptionist is working from a script that is necessarily short, because they are also answering for other businesses.

Cost shape, third. An agent’s cost does not rise per call in the way a per-call or per-minute human service does, so a business whose volume is growing is buying a flatter line. Ask any AI vendor to state the unit precisely, because the ones that bill per minute have reinvented the shape they claim to have escaped.

Ask these of a person on the phone, and ask for a demonstration rather than an answer.

A tempting answer is to buy both: an agent for the easy calls and a service for the rest. It can work, and it is more fragile than it sounds. Two systems mean two places a message can be sitting, two sets of hours to keep in sync, and a call-forwarding chain that will eventually route somebody into a loop at the worst possible moment.

If you do run both, keep the routing rule stupidly simple and write it on a card. Anything more complicated than one rule tends to survive about six weeks past the person who set it up.

If your calls are mostly questions and your problem is volume or hours, an agent. If your calls need judgement, or your callers will not tolerate a machine, people. If you cannot tell which, run a week of tallies rather than a week of thinking about it, and pick the vendor who will let you test on real calls before you commit.