Let a person answer the phone
6 min read
I keep seeing ads for services where AI answers a company's phone. It puts me off. When I call, I usually have something a person needs to handle, and an agent A program built on an AI model that does tasks on its own, for example looks up information, answers questions or books appointments. on the line is just a barrier I have to get through. I think a company should automate everything the person on the other end can't see. That person should get to choose whether they talk to a human or an agent, and the company's website should be ready for them to send their own agent instead.
Why I call at all
When I want service A at time B, I book it through Booksy An app for booking appointments, for example at a hairdresser, beautician or physiotherapist. or a form on the website and I don't need a conversation for that. I call when I have questions.
Recently I wanted to make a large bank transfer, so I called my bank. I wanted to know whether the transfer would definitely go through, how much it costs on my type of account, and whether I could tell the security team in advance that it was coming so they wouldn't have to ask about it. An agent picked up first. I ended up talking to a consultant anyway.
It's similar when I book a doctor's appointment or call about anything else. My problems always seem to need a person, and before I reach one I have to answer the bot's questions and convince it that only a human can answer mine.
Why these agents can't help
I've wondered whether it's a weaker model or the people who set it up. My bet is the setup. A phone agent often has no access to what it needs to answer, like my account or the clinic's calendar. It may also run as a plain LLM A large language model, an AI program that understands and writes text, like the one in ChatGPT. in chat mode, without tool calling The ability of an AI model to use other programs, for example to check data in a system or add an appointment to a calendar.. Then all it can do is repeat what's in its instructions, and for any specific question it transfers you to a consultant anyway.
Giving it that access, on the other hand, is a serious security problem. Anyone who knows the helpline number can talk to an agent that sees bank accounts. Someone could pretend to be me or try to talk the bot into doing something it shouldn't. That's hard to secure properly, so I understand companies that prefer an agent without access. For the person calling, though, that agent is mostly a queue in front of the consultant.
Calls inviting you to product shows
My phone number is on a contract in my mum's name, and she's over fifty, so somewhere it's listed as belonging to a woman 50+. I get calls inviting me to cookware shows and "healing" bedding presentations, or offering English lessons. These calls are set up so that as much as possible happens automatically. I get the impression the lines are recorded in advance and just played back.
I suspect that because of calls like these, older people are more averse to a machine on the line. A bot answering at a clinic sounds to them a lot like the one inviting them to a bedding show.
Where AI makes sense for people outside the company
On the company's website. chatbot A window on a website where you can type with a program that answers questions. have been there for a long time: the model gets the site's context, instructions, and a base of questions or information it can share. That works for me. I want to check something quickly, I see what the bot can tell me, and I decide myself whether my problem needs a person. Next to the chat there could be a second window with a voice agent. I open these on my own, so they don't bother me. To me that's a much better setup than AI answering the phone.
The website should also be ready for agent traffic. More and more often it will be the other person's agent looking for information on the site: through an API An interface through which one program asks another for data or to do something., through a markdown A simple way of writing text with a few marks, such as # for a heading, easy to read for people and programs. version, or just by reading it. Anything a company wants to offer on its website, an agent should be able to do. I described how I did this on my own site in A website an AI agent can use.
MCP A standard that lets an AI assistant connect to apps and services, for example a calendar or an online shop. makes sense when someone uses a service regularly. For a one-time service it's overkill: nobody is going to install a connector in their app to ask a single question. That would scare people off sooner than AI on the phone.
How I would set up the phone
Ideally a person answers. If a bot is going to pick up, it should say up front that it's AI and offer: "I can connect you to a consultant right away or try to answer your question." Or make it a plain choice at the very start, with no conversation with a bot. What matters is that you don't have to convince the agent your question is for a human.
After hours, it's more honest to play a recorded message: we're closed today, there's an agent on our website you can ask, even through ChatGPT An AI assistant from OpenAI that you talk to by chat or by voice. (if the site is ready for it), and someone will call you back tomorrow. An agent could also make sense if it says it's an agent and that nobody can pick up right now. It will happily try to answer your questions, but someone will call you back regardless. That call can produce a transcript and a short brief for whoever calls back. It has to be clear that you're talking to AI, though, and that AI shouldn't settle matters on the company's behalf.
Automation where nobody sees it
Inside the company I would automate data processing with LLMs as much as possible and as far as the budget allows. We're practically forced to by now. Call recordings can go to a model afterwards that pulls out the important information and takes work off people's plates. A consultant can get suggested answers to the questions someone asked them. The person calling sees none of it and talks to a human.
I have my own email set up this way: an automation pulls important details and reminders out of my messages. Doing the same for calls wouldn't be a problem. For myself, though, I'd have to invent a use case, because I don't need it in my private life.
Agents will handle more and more things for people, and websites should be ready for them. When a person calls, though, it's usually to talk to a person.