AI phone agents and AI receptionists: what to decide before one talks to your customers
Three different products are sold as AI phone agents. This guide separates them, says when an off-the-shelf AI receptionist is the right purchase, and lists what to decide before a phone agent reads or writes your CRM. It ends with how to test one before a customer hears it.
Three things sold under one name
"AI phone agent" and "AI receptionist" cover three different purchases. They need different amounts of integration work and carry different risks. Work out which one you mean before you compare vendors or talk to a builder.
| What it is | What it does | What to settle first |
|---|---|---|
| AI receptionist | Answers inbound calls, gives hours and basic answers, takes messages, books appointments | Whether it supports your calendar or booking tool, and how it hands a caller to a person |
| Inbound agent connected to your systems | Identifies the caller, looks up their account, answers from live data, routes by account owner, writes the call back to the record | What it may look up, what it may change, and who reviews what it did |
| Calling help for your own team | A person makes or takes the call. Software records it, transcribes it and writes a summary and next steps onto the record | Recording consent, where transcripts are stored, and who reads the summaries |
The first is a product you subscribe to. The second is an integration project. The third can be either, depending on how much of it lives inside your CRM.
When an off-the-shelf AI receptionist is enough
Buy one if most of your calls are about hours, directions, booking and message taking, and your booking tool is on the vendor's supported list. Many good products do this well, and a custom build would cost you more for the same result.
Before you sign a long contract, run it on your own calls. Listen to a sample of recordings every week. Check each call type it handles. Check what happens when a caller asks for a person.
When the phone has to read and write your records
The picture changes when the answer to a call depends on data in your own systems. These are the usual signs:
- The caller should be recognized by their account.
- The answer depends on an order status, a balance, a delivery date or a contract term held in your CRM or ERP.
- The call should create or update a record, such as a lead, a case or a follow-up task.
- Calls should reach the person who owns the account.
- Someone needs a reviewable record of what was said and what was promised.
At that point you are connecting a voice model to your systems of record. That is the same work as any AI integration: scoped access, logged actions and a person reviewing exceptions.
Decide these before anything answers a call
Write down an answer to each of these before anything is built.
- Which calls it takes, and which it never takes. Start with one call type. Complaints, disputes and anything involving a person's health or money usually go straight to a person.
- What it may look up. Name the systems and fields. A phone agent that can read every field in the CRM can say every field out loud.
- What it may change. Creating a task or a note is low risk. Changing an order, a price or an address is not. Start with notes and tasks.
- How it hands off. Decide how a caller reaches a person, how quickly, and what the person sees when they pick up. After-hours calls need their own rule.
- Disclosure. Decide how the agent says it is automated. Some jurisdictions have rules about this, so ask your counsel.
- Recording and consent. US states differ, and some require the consent of everyone on the call. Settle the notice and the storage location with counsel before launch.
- What goes back on the record. Decide which of these land on the customer record: the transcript, a summary, next steps, promised follow-ups.
- Who reviews. Name the person who reads a sample of calls every week, and the screen where they do it.
- The off switch. Decide who can send every call back to people, and how fast.
What we have built: calling inside the beLoved CRM
beLoved Health is a distribution company David runs. He built the CRM it runs on, and calling is part of it.
Reps place calls from the lead, contact or account record in the browser. Each call is recorded and transcribed as it happens. When the call ends, a model reads the transcript and writes a summary onto the record. It also writes action items, split between the rep and the customer, a suggested next step, and the days and times the contact prefers to be called. The rep reads it and decides what happens next.
This is the third type in the table above: people make the calls, and AI does the work around them. We have not run an inbound AI agent that answers customer calls at beLoved. If that is what you need, the honest first project is a pilot on one call type, with everything on the list above decided in writing.
How to test a phone agent before a customer hears it
- Replay past calls. Take transcripts of real calls your team handled and check what the agent would have said and written at each step.
- Script the hard calls. Make test calls for each call type, plus the difficult ones: a caller who is angry, one who has the wrong number, one who speaks another language, one who asks for a person in the first sentence.
- Read the record after every test call. Open the CRM and read what the agent wrote. That is where quiet mistakes show up.
- Agree the pass mark in advance. Our pilot acceptance checklist covers owner, baseline, acceptance criteria, rollback and the production decision.
- Start at the edges. Send it after-hours or overflow calls first, while your people take the rest.
If the phone agent is one of several agents you are weighing, our worksheet on enterprise AI agents helps you choose which workflow goes first.
Using this with us, or without us
If an off-the-shelf receptionist covers your calls, buy one and use the test list above. If your calls need to read and write your own records, write down your answers to the list in this guide and send them to us. We will tell you what we would look at first.