NewLondon office open. British English agents now live on UK numbers.See our offices
VoiFlow
All resources

Buying guides 8 min read

How to choose an AI voice agent platform: 12 questions to ask

Twelve questions that separate a convincing demo from a platform that works on real calls, with what a good answer sounds like.

Strategy team
VoiFlow

Two colleagues review printed papers at a meeting table with a speakerphone in a bright city-view room

Most AI voice agent demos look impressive. The agent answers quickly, understands a sentence about a booking and sounds natural. The question that matters is whether the same system will still work on a busy afternoon, with a caller who has a second question, a noisy line and a backend that is slow to respond. Choosing the best AI voice agent platform means testing for that, not for the demo.

Choose the platform that works on your real calls, not the one with the best demo. Ask twelve questions covering latency, languages, handoff to staff, integrations, memory, channels, guardrails, data hosting, quality review, telephony fit, pricing clarity and rollout support. Ask for evidence on each one, and test the answers on a live call before you sign.

How to use this list

Use the list if you are comparing two or three vendors. Give each one the same twelve questions, ask for the same evidence, and record the answers side by side. The platform overview explains the parts most vendors will describe, so you can check their answers against it.

Some answers will be strong and some will be vague. A vague answer is not a no, but it is a reason to ask again in writing.

The 12 questions

How quickly does the agent respond on a live call?

Latency is the gap between the caller finishing a sentence and hearing the reply. Long gaps make callers repeat themselves or talk over the agent, and that is the fastest way to lose trust on the phone. Ask for response times measured on live calls rather than in a test lab, and ask to hear a recording of a real call made at a busy time of day.

A good answer sounds like: we measure the time from the end of the caller's speech to the first word of the reply, on live calls, and we can share the figures for your region.

Which languages and dialects does it handle, and how well?

Check the list the vendor gives you, then test it. A language on a list can still fail on local words, mixed speech or a regional accent. If your callers switch languages within one sentence, ask how the platform handles that and test it with your own staff.

A good answer sounds like: here are the languages we support, here is how the agent follows a switch in the middle of a sentence, and here is a recording from a live call in the dialect you need.

How does a handoff to a person work?

A handoff is where many deployments fail. Ask what the staff member sees first, how the call is passed on, whether the caller has to repeat anything, and what happens when nobody is free to take it.

A good answer sounds like: the person receives the caller's details, the reason for the call and a short summary before they answer. Supervisors can listen in, and staff can take over a live call. If no one is free, the agent takes a message and books a callback.

Which systems can it connect to, and who builds the connection?

Integrations decide whether the agent can book, update or look up anything. Ask for the list of supported systems, then check the ones you need on the first day. Ask who builds each connection, how long it takes, and who maintains it when your system changes.

A good answer sounds like: here are the calendars and customer systems we connect to today. For yours, here is the method, who builds it, how long it takes, and who looks after it when either side changes.

What does the agent remember between calls and messages?

Memory is useful when it is right and harmful when it is wrong. Ask what the agent keeps about a customer, how long it keeps it, how a customer or staff member can correct it, and whether the memory is shared across channels. Good memory is visible to you, not hidden inside the model.

A good answer sounds like: the agent keeps the customer's open matters, promises and preferences across calls and messages. You can see what it holds, correct it and delete it under your retention rules.

Which channels does it cover, and is it one conversation?

A call is only one channel. Customers also message on WhatsApp, SMS and email. Ask whether a call and a follow-up message share one customer record, or whether each channel starts from scratch.

A good answer sounds like: a call, a WhatsApp thread and an email share one customer record. If the agent promises a reply, that promise appears in the same record and stays open until it is done.

Which guardrails are enforced outside the model?

A language model can be talked into the wrong action in the middle of a conversation. Permissions, spending limits, consent checks, redaction of sensitive data and an audit trail should sit in the platform itself, not only in the instructions written for the model. Ask which rules the model cannot override, and how each one is logged.

A good answer sounds like: permissions and limits are checked by the platform before any action runs. Consent and do-not-call rules are enforced on outbound calls. Sensitive data is redacted, and every action is written to an audit trail you can review.

Where is our data hosted, and who can read it?

Ask where calls, recordings, transcripts and customer records are stored and processed, and whether that location is fixed in your contract. Ask which other companies act as sub-processors, meaning the firms that handle your data on the vendor's behalf, and who at the vendor can access your data. The answers should be in writing, not only on a sales call. For the full set of questions to put to a vendor on data and security, use the security checklist.

A good answer sounds like: data is hosted in the region named in the contract. Here is the current list of sub-processors, with notice of any change, and here is how staff access to customer data is restricted and logged.

How are calls reviewed, scored and searched?

Ask whether every call is recorded, transcribed and scored, whether you can replay a call, and how supervisors find the calls that matter. A quality review should show the calls where the agent went wrong, not only an average score.

A good answer sounds like: every call is recorded, transcribed and scored. Supervisors can filter for escalations, failed bookings and complaints, and replay any call alongside its transcript.

How does it fit our phone system?

SIP is the standard way phone systems connect over the internet, and a PBX is the company phone system that routes calls inside a business. Ask whether the platform connects to your current PBX through SIP, or whether you need new numbers, and what changes on your side. The answer decides how much work the rollout takes.

A good answer sounds like: we connect to your current PBX over SIP, or we provide new local numbers. Here is what changes on your side, who makes the change, and how calls are moved across step by step.

Is the pricing clear enough to forecast?

Ask for each cost on its own line: the licence, included minutes and messages, phone numbers, carrier charges, setup, integration and overage. Ask for the total at your own call volume. The cost guide explains how to put two quotes on the same basis.

A good answer sounds like: here is the monthly licence, what it includes and the rate above the allowance, and here is the total for the call volume you described, with every line shown.

What support do you give during rollout, and who writes the knowledge?

Ask who writes the first version of the agent's knowledge and call flows, how many test calls happen before go-live, who is on call during the first weeks, and what support you get after launch. Ask for the rollout plan in writing, with dates and an owner on each side.

A good answer sounds like: we write the first version with you in a set working session. Here is the test plan, the go-live checklist and the named contact for the first 30 days.

Red flags in a demo

Some problems only show up after the contract is signed. These are the signs to look for during the demo itself:

  • The demo works on scripted phrases, and a surprise question breaks it.
  • The handoff is described but never shown on screen.
  • Response speed is described in general terms and never measured.
  • Data is called secure, but nobody can say where it is stored.
  • The price depends on questions you did not ask.
  • Staff cannot see transcripts or change the knowledge themselves.
  • The vendor says the agent never makes mistakes.

A checklist to share with your team

Use this table to score each vendor on the same page. A pass needs evidence, not a verbal yes.

#AreaEvidence to ask forRed flag
1Response speedMeasured times on live callsFigures from a test lab only
2LanguagesRecordings in your dialects and mixed speechA list with no examples
3HandoffA live demonstration with the summary shownHandoff described, not shown
4IntegrationsThe list, plus the method and owner for yoursConnections promised for later
5MemoryWhat is stored, how long, and how to correct itMemory that cannot be seen
6ChannelsProof that calls and messages share one recordSeparate threads per channel
7GuardrailsThe rules enforced outside the model, and the logSafety described only in the prompt
8Data hostingHosting region, sub-processors and access controlsNo stated region
9Quality reviewScored calls, replay and supervisor filtersAverage scores with no call review
10TelephonyConnection method for your PBX or number planNo plan for your phone system
11PricingEach cost line and the total for your volumeA single headline price
12RolloutA written plan with dates and named ownersGo-live left to chance

How VoiFlow handles this

VoiFlow's AI voice agents are hosted in-region in the UAE and GCC, Malaysia and Southeast Asia, the UK and the US. They work in more than 30 languages and regional dialects, including Gulf Arabic, Bahasa Malaysia and Manglish. A warm handoff gives staff the full summary, and every call is recorded, transcribed and replayable. VoiFlow's agents are never trained on your data, and the Trust Center lists the documents you can review.

Frequently asked questions

How many vendors should I shortlist?

Three or four is usually enough. A longer list adds comparison work without adding much insight, and a fair comparison needs each vendor to answer the same questions.

Should I run a pilot before I sign?

Yes, when the call type matters to your business. A short pilot on one call type shows how the platform handles real callers. The 30-day pilot guide sets out how to run one.

What if a vendor will not answer a question in writing?

Treat that as a finding. Written answers are what you rely on later in the contract, and a vendor that will not put an answer on paper may not stand behind it.

Use the twelve questions to narrow your list, then test the top two or three on a live call. Try a live agent to hear how the answers sound in practice, or contact our team to talk through your own call types.

Call it now · United States line+1 (831) 282-9226

Hear it run one of your calls.

Call the agent yourself, start a free trial, or bring us your hardest workflow. Most teams are live in days.