Key takeaways
- An AI phone provider hears your clients before anyone else does, so treat it as a processor of confidential data and test it like one.
- Ask in writing where calls are processed, who the sub-processors are and whether your callers' data trains anyone's models.
- The processor contract must cover the terms UK GDPR Article 28 requires, including deletion or return of data at the end[1].
- Check hand-offs and failover with your own scenarios; accountability for what the assistant says stays with the firm[2].
- Score suppliers against the same questions and walk away from vague answers on data and incidents.
Why a phone provider needs proper due diligence
An AI phone assistant is not a neutral pipe. It listens to new clients describing their problems, records what they say, turns it into text and passes it to your team. That makes the provider a processor of some of the most sensitive information the firm handles, often before any conflict check or client care letter. The Law Society's guidance on generative AI, aimed particularly at small and medium-sized firms, flags misuse or disclosure of confidential information as a key risk[3].
The regulatory position is equally clear. The SRA says firms remain responsible and accountable for outputs from AI they use, including errors in advice given by a third-party chatbot[2]. A supplier's assurances do not transfer that responsibility; they are evidence that the firm took reasonable steps.
Customer contact is where many businesses start. In DSIT's research, 46% of UK businesses using or planning AI use or plan it for customer service, against 16% for legal functions[4].
Source: Department for Science, Innovation and Technology (GOV.UK), 2026[4]
Show the data
| Item | Value |
|---|---|
| Customer service | 46% |
| Legal functions | 16% |
How to run the process
Shortlist on scope
Exclude providers that cannot support your basic requirements, such as UK English, call recording, escalation to several numbers and a way to keep your existing phone number.
Send the questions in writing
Use the same question set for every supplier and ask for written answers, so you can compare like with like and keep a record for your risk file.
Run your own scenarios
Ask for a demonstration using your scripts: an urgent family call, an advice request, an existing client chasing, a caller who goes silent.
Check references
Speak to at least one existing professional services customer about outages, support response and how the supplier handled a problem.
Review the contract
Have the data protection lead and a commercial solicitor review the processor terms, service levels and exit provisions.
Pilot before committing
Agree a short trial with clear success criteria and an easy exit, before any long-term commitment.
Questions on data location and sub-processors
- Where are calls, recordings and transcripts processed and stored? Name the countries and the hosting providers.
- Which sub-processors handle our data, for what purpose, and how will you tell us before adding or replacing one?
- If data leaves the UK, which transfer mechanism do you rely on, and can you show it?
- How long do you keep recordings, transcripts and call metadata by default, and can we set shorter periods?
- At the end of the contract, will you delete or return all our data at our choice, and confirm deletion in writing?
These questions map onto the law. UK GDPR Article 28 requires controllers to use only processors providing sufficient guarantees, requires the controller's written authorisation before a processor engages another processor, and requires the contract to cover processing only on documented instructions, confidentiality of staff, security, help with data subject rights, deletion or return of data at the end and information for audits[1].
Questions on training and reuse of your callers' data
- Is any of our audio, transcripts or call notes used to train or improve your models or anyone else's?
- If so, can we opt out completely, and is the opt-out the default?
- Which third-party AI services process our callers' words, and on what terms do those services hold the data?
- Can your staff listen to our calls, and under what controls and logging?
The Legal Ombudsman warns that not all free AI tools are secure or keep what users type confidential[5]. Paid business services can differ greatly from free consumer versions, which is exactly why the answer must be in writing and in the contract rather than in a sales conversation.
Questions on security and incidents
- Which security certifications do you hold, what is their scope, and when were they last audited?
- How is access to our account controlled: multi-factor authentication, role-based permissions, single sign-on?
- Is data encrypted in transit and at rest, and who holds the keys?
- How quickly will you tell us about a personal data breach affecting our data, and what will that notice contain?
- When did you last test your incident response, and what changed as a result?
Timing matters because the firm's own clock starts when it becomes aware of a breach. Under UK GDPR Article 33, a controller must notify the ICO without undue delay and, where feasible, within 72 hours[6]. A supplier that promises to tell you 'promptly' without a defined window leaves the firm exposed.
Questions on call handling, hand-offs and failover
- Can we set our own intake questions per practice area, and who can change them?
- How does escalation work: how many destinations, in what order, and what happens if nobody answers?
- If your service fails, where do our calls go? Show us the failover and how quickly it switches.
- How is the caller told the call is recorded and that they are speaking to an automated assistant?
- Can callers ask for a person at any point, and how is that honoured out of hours?
- How do you handle silence, speech differences, relay calls and interpreters?
Do not accept a slide or a recorded demo as the answer to these questions. Ask the supplier to set up a test line with your escalation numbers and ring it yourself: during office hours, after hours, with the first escalation number switched off, and while asking repeatedly for a person. Time how long each hand-off takes and note exactly what the caller hears while they wait. The failover test matters most, because an outage on a busy Monday morning is when the firm can least afford calls to disappear.
Questions on supervision, audit and exit
- Can we see every call's recording, transcript and summary, and export them?
- Is there a change log showing who changed wording or routing, and when?
- Can we keep our existing numbers, and port them away again if we leave?
- What notice period applies, and what help do you give with transition?
- What service levels do you commit to for uptime and support response, and what happens if you miss them?
Supervision questions show whether the firm can prove it oversaw the tool. If a complaint arrives months later, you will need the recording, the transcript, the wording that was live on that date and a record of who approved it. A supplier that cannot produce all four on request makes the firm's position harder to defend, however good the calls sound.
Red flags
| What you hear | Why it is a problem |
|---|---|
| 'We don't share that list of sub-processors' | You cannot authorise processors you do not know about |
| 'Your data helps improve the service for everyone' | Callers' confidential words may be reused beyond your instructions |
| 'We tell customers about breaches as soon as practical' | No defined window to support your own notification duty |
| 'Escalation is on the roadmap' | Urgent matters have nowhere reliable to go |
| 'Number porting is handled case by case' | Leaving may mean losing your published phone number |
| 'Our standard terms can't be changed' | The contract may not contain what Article 28 requires |
Scoring and deciding
Score each answer from zero to three: zero for no answer or a refusal, one for a vague answer, two for a clear answer with gaps, three for a clear answer backed by documents. Weight data and incident questions double. Any zero on data location, training, breach notification or failover should rule a supplier out regardless of the total.
Example: a fictional firm compared three providers. Supplier A scored highest on features but would not commit to a breach notification window. Supplier B had fewer features but answered every data question with documents and showed a working failover. Supplier C could not confirm whether recordings were used for training. The firm piloted Supplier B. For the trial itself, see running a 30-day AI reception pilot; for accessibility checks to add to your scenarios, see making AI voice assistants accessible; and for evening cover requirements, see AI for evening and weekend enquiries.
Frequently asked questions
Is an AI phone provider a data processor?
Usually yes. It processes callers' personal data on the firm's behalf, so the firm needs a processor contract with the terms UK GDPR Article 28 requires and must check the provider gives sufficient guarantees.
What is the most important question to ask an AI phone supplier?
Whether your callers' audio, transcripts or notes are used to train or improve any models, and whether you can opt out completely. The answer should be in writing and in the contract.
How quickly should a supplier report a data breach to us?
Within a defined window short enough for the firm to assess it and, where needed, notify the ICO within 72 hours of becoming aware. Vague promises are not enough.
Can we keep our phone number if we change provider?
Ask before signing. Check that the supplier supports porting your existing numbers in and out, and what notice and help they give if you leave.
References
legislation.gov.uk (The National Archives) (2016). UK GDPR Article 28: Processor (opens in a new tab) ↑
Solicitors Regulation Authority (2023). Risk Outlook report: The use of artificial intelligence in the legal market (opens in a new tab) ↑
The Law Society (2026). Generative AI - the essentials (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2026). Artificial Intelligence Adoption Research (opens in a new tab) ↑
Legal Ombudsman (2026). AI and Complaints: A Consumer Guide (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2016). UK GDPR Article 33: Notification of a personal data breach to the supervisory authority (opens in a new tab) ↑