Before the Bot Picks Up: Seven Procurement Questions for AI Voice Agents

AI receptionists and voice agents are now a familiar item in the supplier conversations of UK small and medium-sized businesses. The pitch is straightforward: fewer missed calls, cover outside office hours, and a lower cost than a full-time receptionist. The procurement question is less straightforward. An AI agent that answers the phone is handling personal data, representing the business and creating records that may later matter in a dispute.

This is a buyer's checklist, not a forecast. It assumes you may pilot or buy an AI voice service in 2026 and want to ask the questions that sit behind the demo.

1. What is the lawful basis for recording and processing calls?

Ask how the service handles call recordings, transcripts and AI processing under UK GDPR and the Data Protection Act 2018. Who is the controller and who is the processor? What is the lawful basis for each processing activity? How are callers told that an AI or recording is in use, and at what point in the call? What retention period applies, and how are deletion requests handled?

Do not accept a statement that the vendor is GDPR compliant as a complete answer. Ask for the data processing agreement, the list of subprocessors, storage locations and any international transfers. If the service monitors or records calls for quality or training, check that the notice and consent model matches your organisation's policies. A short data protection impact assessment may be appropriate before a pilot.

2. Which calls must never reach the AI?

A voice agent needs a triage rule, not just a script. Ask which call types are routed to a human immediately: emergencies, safeguarding concerns, vulnerable customers, complaints, legal threats, press enquiries and existing customers with complex accounts. Define the handoff method and the maximum wait before escalation.

Test the failure path. What happens if the AI cannot understand the caller, the CRM is unavailable or the human queue is closed? The answer should be a named fallback, such as a voicemail box that is monitored, an on-call number or a call-back promise with an owner. Saying that the AI will handle it is not a fallback.

3. How does the agent know what to say, and who approves changes?

Ask to see the knowledge base, call flows and escalation scripts. Who writes them, who approves them and how often are they reviewed? Can the supplier change the model or the prompts without your sign-off? How are errors logged and corrected?

A demo often uses a narrow set of questions. Your callers will not. Ask for a test against real anonymised scenarios from your business, including awkward ones: a caller who gives a partial postcode, a supplier chasing an invoice, or someone who asks for a named colleague. Check how the agent handles uncertainty. A confident wrong answer is worse than a clear transfer to a human.

4. What does the integration actually touch?

If the AI is to book appointments, check availability or log tickets, ask exactly which systems it reads and writes. Does it connect to your CRM, calendar, helpdesk or telephony platform? What permissions does it use? Can it create, amend or delete records? How are duplicate contacts and conflicting appointments handled?

Integration is where procurement quietly becomes an IT project. Ask for a data map: what fields are read, what is written back, where the record of truth sits and how the integration is monitored. Confirm whether the supplier needs persistent access to your systems and how that access is revoked at the end of the contract.

5. Who owns the number, the voice and the data?

Number ownership matters. If the AI service provides a new number, can you port it away if you leave? If it uses your existing number, who controls the routing and the porting authorisation? Ask for the position in writing before signing.

Voice is another asset. If a synthetic voice is cloned from a real person, obtain their written consent and agree what happens when they leave or withdraw consent. Ask whether your call data or transcripts can be used to train models, and how to opt out. Confirm the export format for transcripts, call logs and contact data at termination.

6. What happens when the AI gets it wrong?

Ask for the service levels that apply to accuracy, availability and call handling, not just uptime. How quickly are transcripts and recordings available for review? What is the complaint process when a caller says the AI misled them? Who is liable for a missed appointment, a wrong quotation or a disclosure to the wrong person?

Check the audit trail. Can you see what the agent said, what it heard and why it transferred or ended a call? Review the supplier's insurance and indemnity position, and make sure your own professional indemnity cover is not silently undermined by an automated process you cannot evidence.

7. What are the commercial and exit terms?

Finally, price the whole service, not the headline rate. Ask about per-minute or per-call charges, minimum monthly spend, overage rates, setup fees, integration fees and charges for changes to scripts or call flows. Check the contract term, notice period and termination assistance.

Exit deserves its own question. What happens to your number, recordings and integrations when the contract ends? How long does data remain available for export? What deletion certificate will you receive? A voice agent that cannot be exited cleanly is not a low-risk purchase, however good the demo sounds.

Pilot before you sign

A short pilot with agreed success measures is usually more informative than a longer demo. Define the call types, volumes, escalation rules and review process in advance. Listen to recordings, check the transcripts and ask staff who handle the consequences what they think. The right AI voice agent should make the phone easier to manage, not add a new layer of uncertainty to your customer service.