Short answer: treat the launch as a data-protection job first and a software choice second. The five checks below are our own practice advice, built on ICO guidance. They are not a legal formula, and they apply whichever product you pick.
What an AI receptionist actually does, and what it should not do
It answers the phone, asks who is calling and why, takes a message, and may book a slot or pass the call on. For a small business where one busy person does everything, that can be useful.
What it should not do is make judgement calls: whether a complaint is serious, whether a caller gets a refund, whether someone’s symptoms sound urgent. Our rule of thumb is that if the answer could change what happens to the caller, a person decides.
Most search results on this topic are vendor pages and ranked lists. They compare features and skip the data questions, so this post names no products and quotes no prices. For the wider admin picture, the office managers hub covers the drafting jobs AI suits, and AI for small business: start with the back office is a gentler first step than anything customer-facing.
Why the phone is a data-protection question before it is a software question
Every call carries personal data: a name, a number, a reason for calling, and sometimes things the caller did not need to say. You decide why that data is collected and what happens to it, so in most set-ups you are the controller under UK GDPR. Those duties stay with you whichever product you buy.
The ICO’s guidance on AI and data protection is organised around accountability and governance, transparency, lawfulness, accuracy and fairness. On a phone line, that means telling callers what is happening, being able to show your reasoning, and checking that what the system writes down is right. The ICO also notes that much of its guidance is under review following the Data (Use and Access) Act, so check the current version of each page before you rely on it.
Check 1: Who is the processor, and what does the contract say?
In most set-ups, the company running the AI receptionist handles call data on your behalf, which makes it your processor. The ICO’s guidance on contracts between controllers and processors says that every time a controller uses a processor there must be a written contract (Article 28). When that processor uses another processor, it needs a written contract with them too.
Read the contract and its data processing terms before you sign. Questions we would ask:
- Who else handles the audio and transcripts (the sub-processors)?
- Is call data used for anything beyond running your service?
- Where is it stored, and what happens to it when you leave?
If nobody can answer in plain English, treat that as your answer.
Check 2: What do callers hear at the start of the call?
The ICO’s page on the right to be informed says you must give people privacy information at the time you collect their data. That includes your purposes, your retention periods and who you share it with, in clear and plain language. The ICO also says layering often works well: a short message first, with fuller detail one step away.
On a phone line, the first layer could be a few seconds that:
- Say the caller is speaking to an automated assistant.
- Say what it will note down and why.
- Say whether the call is recorded.
- Tell them how to reach a person and where to read the full privacy notice.
Write it, read it aloud, time it and cut it. Then make sure your website privacy notice covers the phone line, because the ICO says you must tell people about any new use of their data before you start.
Check 3: Recordings, transcripts and how long you keep them
Many systems keep everything unless you change the setting. The ICO’s storage limitation page says you must not keep personal data for longer than you need it, you must be able to justify how long you keep it, and you need a policy with standard retention periods wherever possible.
Set one period for recordings and a separate one for transcripts and message logs. You may need one for much longer than the other, or you may not need recordings at all. Write the periods down, then check that the system actually deletes on schedule.
Whether to record at all is the earlier decision. We could not find a current, dedicated ICO page on recording business calls to cite here, so check the ICO’s current guidance before you rely on any specific rule. Either way, the duties to tell callers and to justify how long you keep data still apply.
Check 4: Where a human takes over (complaints, health, money, urgent calls)
Decide in advance which calls go straight to a person. Our own starting list: complaints, anything about health or money, anything urgent or distressed, and anyone who asks for a human.
Health needs particular care because callers often mention it without being asked. The ICO’s special category data guidance says this data needs more protection because it is sensitive, and health is on the list. To process it lawfully you need a lawful basis and a separate Article 9 condition, decided and documented before the processing begins. A receptionist that transcribes everything it hears can end up holding health details you never planned for. Route those calls to a person, and decide what happens to anything already captured.
Then there are decisions. The ICO’s page on automated decision-making explains that Article 22 adds rules where a decision is made solely by automated means and has legal or similarly significant effects on someone. Keep the receptionist well clear of that: it takes messages and books slots, and a person makes any decision about the caller.
Finally, test the handover out of hours. Where does a call go at 5.30pm on a Friday when nobody is in?
Check 5: Who checks the messages and bookings it writes down
The ICO’s AI guidance has a chapter on accuracy, and on the phone the errors are mundane: a misheard name, two digits swapped, a booking on the wrong day. Somebody has to own the checking.
Name one person and give them a set time each day to review what came in, correct it against your diary or booking system, and chase anything odd. “Whoever is free” usually means nobody. Treat the booking system as the record and the transcript as a draft that a person checks.
Route rights requests the same way. If a caller asks what you hold about them, or asks you to delete it, that message needs to reach whoever handles data requests promptly. The principle that runs through the HR hub applies here too: AI can draft and note things down, and decisions about people stay with people.
Five test calls to make before you go live
Ring it yourself, then ask a colleague who has not seen the setup. Use made-up names and details for every test.
- A plain booking. Does it capture the details correctly and read them back?
- A muddled caller. Change your mind mid-sentence and talk over it. Does it cope, or loop?
- A complaint. Does it hand over to a person, or try to resolve it itself?
- A fictional health detail dropped in casually. Where does it end up, who can see it, and can you delete it?
- “Can I speak to a person?” Does it put you through the first time you ask?
Then check the records. What was stored, where, and under which retention period? If anything surprises you, fix it before launch.
When to leave the phones with a person
If most of your calls are sensitive, emotional or high-stakes, an AI at the front of the queue is probably the wrong fit. If you cannot get clear answers on the contract, retention or handover, wait until you can.
Think about a data protection impact assessment too. The ICO’s DPIA guidance says one is required where processing is likely to result in high risk, and it recommends doing one if you are in any doubt. A line that may hear health details is a good reason to do one.
A receptionist that is switched off is annoying. One that quietly collects data you cannot account for is a much bigger problem. If you are still weighing up tools in general, choose by task and data sets out the approach, and the guides index lists every task manual on the site.