Voice agents
How we build and test a voice agent
How a build runs, from the first contact to the review after go-live. You can ask us about any step.
Why testing matters more than the voice
A voice agent that tells a caller their request is booked when it is not does more harm than one that says nothing. That is the kind of failure we design against. It is why most of the work goes into controls and testing, not into how the agent sounds.
1. Discovery
We start with your call types, your systems and your rules. Who rings, what they ask, what your team does next, and which calls must reach a person. We also agree who on your side signs off the wording.
2. Your systems and your wording
The agent is built around your systems, data and processes. It reads from and writes to the tools you already use, follows your hand-over rules and uses your wording. We agree what it reads, what it writes and where, before we build it.
3. Controls
What it may say
Wording and topics you approve. Anything outside them is set to be handed over.
What it may do
Only the actions you allow, such as taking details or logging a request.
Hand-over rules
You decide which calls go to a person and who rings. Notes go with the call.
It says it is an AI
It is set to say it is an AI at the start of the call.
Built to reduce hallucination
Hallucination is an AI making up an answer it does not have. The agent keeps to your scripts and records, has a narrow job and a hand-over rule for calls outside your scripts.
4. Testing before go-live
Calls written in advance
We write test calls from your own call types before the agent takes a real one.
Difficult callers included
Distressed callers, bad lines, people who interrupt, people who give details in pieces and people who push for advice or try to trick the agent.
Each failure becomes a test
When a test call goes wrong, we fix the agent and keep that call. It reruns after every change, so we can see if an old problem returns.
5. Run alongside your team first
The agent starts next to your people, on one call type. Your team sees what it does and what it hands over. The next call type is added once the first has settled.
6. After go-live: call analysis and review
Calls are recorded and logged where you choose. Call analysis shows which calls come in, what callers struggle to say and which calls get handed over. We review the handed-over calls with your team. When they show a gap, we change the agent and rerun the tests before the change goes live.
What we do not claim
No voice agent is error-free, and we can't promise a particular result. What we can show you is how we work, and you're welcome to question any step of it.
Voice agents by sector
The sector pages show how this works for a particular kind of business.
Questions
Does testing mean the agent cannot make a mistake?
No. Testing looks for problems before go-live, and each one found becomes a test that reruns after every change. After go-live, call analysis and the review of handed-over calls look for what testing missed.
What counts as a difficult caller?
Someone distressed, someone on a bad line, someone who interrupts or gives details in pieces, someone who asks for advice the agent must not give, and someone who tries to make it say something it should not.
Does the agent say it is an AI?
Yes. It is set to introduce itself as an AI assistant at the start of the call.
Who reviews the handed-over calls?
We review them with your team. When a call shows a gap, we change the agent and rerun the tests before the change goes live.
Where are calls recorded?
Calls are recorded and logged where you choose.
Tell us which calls to start with
Tell us about the process you want to improve and we'll say whether we can help.