What should you ask an AI voice vendor before you buy?
The nine questions below will tell you more about an AI voice vendor than any demo will. They are the ones that surface how a platform behaves in production rather than in a controlled script, and most of them have a wrong answer you can recognise without being technical.
There is a reason the demo is a poor test. It runs on a clean line, a rehearsed script and a call type the vendor chose. Your contact centre runs on a network that occasionally wobbles, callers who interrupt and switch language halfway through a sentence, and a long tail of questions nobody wrote a flow for. The gap between those two environments is where AI voice projects fail, and no demo will show it to you. Before you can run a trial, the question set is the only instrument you have.
1. Does it run on our existing telephony and CRM, or does something have to be replaced?
This is the first question because it sets the size of the project.
An AI voice agent that integrates at the telephony and CRM layer is a four to six week engagement. One that requires a platform migration first is an eighteen month programme with an AI feature at the end of it. Both get sold using the same three words, and the difference between them is most of your budget and most of your risk.
The right answer names your actual stack, whether that is Avaya, Genesys, Cisco or a local PBX, and describes the integration point rather than a replacement. A vendor who can tell you what stays exactly where it is, your routing, your records, your reporting, has done this before. A vendor who opens with a discovery phase to assess platform suitability is telling you that a migration is coming.
2. Where is our data processed and stored?
For a South African business this is a POPIA question before it is a technical one.
The Protection of Personal Information Act places conditions on transferring personal information outside the country, and a call recording containing an ID number is personal information. Depending on the sector it may be special personal information as well. So ask where the audio goes, where the transcript goes, how long each is retained, and who else can see it.
Then ask the version most buyers forget: does any of that change when the vendor uses a third party model provider, and would you be told if it did. A vendor who can draw the data flow on a whiteboard in the meeting has thought about it. One who says hosted securely in the cloud has not.
3. What happens when the underlying model changes?
This is the question almost nobody asks and it is the one that matters most over a three year contract.
The model behind an AI voice agent gets updated. When it does, behaviour you tested and approved six months ago can shift without anyone touching your configuration. That is not a hypothetical risk, it is how these systems work.
So ask how changes are controlled. Are you notified before or after. Can a version be pinned. What regression testing runs against your specific call flows before a change reaches your callers. Who signs it off, on your side and theirs. If the answer is that they are always on the latest model, presented as a benefit with nothing attached to it, you have found a gap.
4. Who is accountable when it gets an answer wrong?
In South African law the accountability sits with the business that deployed the system, not the vendor that supplied it. That will not change when the national AI framework arrives.
So the useful version of this question is not who carries the liability. It is what the vendor does to reduce your exposure. Look for logging you can audit, a defined escalation path with a person at the end of it, control over what changes and when, and an independent standard behind the governance rather than a policy document written in house.
A vendor whose answer is a contract clause moving the liability to you, and nothing else, has answered a legal question and dodged an operational one.
5. How does it hand over to an agent, and what does the agent see?
A handover that drops context is worse than no automation.
The customer explains the problem twice. The agent starts cold. The call is now longer than it would have been with no automation at all, and you have paid for a platform that made your operation slower.
Done properly, the system passes the caller identity, the account, the intent it detected and the transcript of what has already been said, and the agent picks up with all of it on screen before they speak. So ask to see the agent screen at the exact moment of handover. Not the flow diagram. The screen. If a vendor cannot show it, they have not built it.
6. Which languages are live today, and which are on a roadmap?
South Africa has eleven official languages and no vendor supports all of them at production quality today. That is fine and it is not the problem.
The problem is a roadmap presented as a capability. Ask which languages are live right now in a paying customer environment, and get the rest in writing as a roadmap with dates against it. Ask specifically about code switching, because South African callers move between languages inside a single sentence and a system that handles two languages separately may not handle both at once.
7. What is the resolution rate, not the containment rate?
Containment counts every call that did not reach an agent. That includes the customer whose problem was solved and the customer who gave up and phoned back the next morning. Same number, opposite outcomes.
Resolution counts the calls where the customer got what they phoned for. Ask for both, then ask for repeat contact rate inside seven days, which is the number that catches deferred calls masquerading as contained ones. Then ask for customer satisfaction measured on the automated path on its own, never blended into the overall contact centre score.
A vendor who leads with containment and cannot produce a repeat contact figure is showing you the flattering number.
8. Can we test it on one call type before committing?
If the answer involves a long contract before anything runs, the vendor is asking you to carry the risk of their implementation.
A proof of concept on a single high volume call reason, scoped to four to six weeks with the measures agreed in writing before it starts, gives you a decision based on your own numbers rather than someone else’s reference. It should run on your existing infrastructure, which means there is nothing to unwind if the answer is no.
A platform confident in its own product has no reason to refuse this.
9. What independent certification do you hold for AI governance?
ISO/IEC 42001 is the international standard for AI management systems. It covers how an AI system is designed, deployed, monitored and corrected when its behaviour drifts, and it is independently audited rather than self declared.
That distinction is the whole point. South Africa has no AI law and the national framework is not expected before 2027, which makes independent certification the only external benchmark a compliance function has to work with today.
Watch the wording carefully. A company is certified to the standard. Compliant with, aligned to and built around are not certification. Ask any vendor to evidence the certification rather than assert it, and ask what the certified scope actually covers, because a scope that excludes the product you are buying tells you very little.
1Stream is certified to ISO/IEC 42001, and a client deploying our AI voice agent inherits that governance position rather than having to build one.
The short version
A vendor who answers questions 1, 5, 7 and 8 well is probably a working platform.
A vendor who answers 2, 3, 4 and 9 well can survive a procurement process and a compliance review.
You want both, and they are not the same vendor as often as you would hope.
Frequently asked questions
What is the most important question to ask an AI voice vendor?
Whether the platform runs on your existing telephony and CRM. The answer determines whether this is a six week integration or a multi year replacement programme, and it is the single biggest driver of cost and risk in the decision.
Is containment rate a good measure of AI voice performance?
No. Containment counts any call that did not reach a human agent, including callers who abandoned. Resolution rate paired with repeat contact rate inside seven days is the honest measure.
Does POPIA allow an AI voice agent to handle customer calls in South Africa?
Yes, with conditions. Under the Protection of Personal Information Act, personal information must be processed lawfully and securely, moving it outside South Africa is restricted, and there are limits on decisions taken solely by automated means where those decisions affect a person materially. Service and lookup calls sit comfortably inside those limits. Decisions that affect a customer materially do not. This is general information rather than legal advice.
What is ISO/IEC 42001 and why does it matter when buying AI?
It is the international standard for AI management systems, covering how AI is governed across its lifecycle. It matters because it is independently audited rather than self declared, which gives a compliance officer something to verify at a point where South African law offers no benchmark.
How long should an AI voice proof of concept take?
Four to six weeks on a single defined call type, with the measures agreed before it starts. Anything shorter will not produce a defensible number. Anything longer usually means the scope was never tight enough.
Where to go from here
If you want the nine questions as a working document, with the answers to expect, the answers that should worry you and a scoring sheet you can take into a vendor meeting, download the AI Voice Procurement Pack. It is vendor neutral, which means you can point it at us as well.
If you would rather have the conversation, book a CX AI Readiness Assessment. Forty five minutes, ending in a written recommendation on which of your call reasons automate first and which should be left alone.
