AI & automation
AI voice agents that know when to hand over
A voice agent is very good at the call that happens two hundred times a month and very bad at the one that does not. The design work is all in the boundary.
Context
What they are good at and what they are not
Voice agents handle structured, repetitive calls well: hours and directions, order status, booking and rescheduling, taking a repeat order, capturing details for a callback. They never get tired, never have a bad shift, and answer every simultaneous call at 6:30pm.
They are poor at emotional callers, unusual situations, heavy accents against background noise, and admitting uncertainty. Left unconstrained, a language model will produce a confident, fluent, wrong answer — and on a phone call there is no visible trace of it until a customer acts on it.
So the build is mostly boundary work: what the agent is allowed to answer, what it must escalate, and how it behaves when it is unsure. We publish a longer piece on the specific failure modes we design against, because a vendor who cannot name them has not deployed enough of these.
Scope
What a build includes
- Scoping to one clearly defined job with a testable right answer
- Knowledge grounded in your material, with refusal behaviour when it runs out
- Escalation design: the triggers, the handover, and the transcript that travels with it
- Adversarial testing before launch — noise, accents, confusion, and deliberate edge cases
- Supervised soft launch on a share of live volume
- Weekly transcript review and correction after go-live
Detail
Why the monthly fee exists
A voice agent that nobody reviews degrades. Your business changes, the questions change, and the gap between what the agent says and what is true widens quietly. Nobody notices until a customer is annoyed enough to tell you.
The monthly covers hosting, monitoring, and the transcript review that keeps it accurate. We will not build one without it, because an unmonitored agent eventually makes us look worse than doing nothing would have. It is the same discipline our own contact centre agents run on.
FAQ
Questions people ask
What does an AI voice agent cost?
It depends on scope, so it is published as three tiers — Simple, Standard, and Complex — from $2,500 to build plus $250 a month. What moves you up a tier is how many intents the agent handles, how many systems it touches, and whether it needs memory across calls. Every tier includes escalation design, adversarial testing, and a round of revisions after launch. The full table is on the pricing page.
Can it book appointments in our calendar?
Yes, where your booking system exposes an interface. Booking is one of the better fits for a voice agent because the task is structured and the outcome is verifiable — either the slot exists and is now taken, or it is not.
What if it gets something wrong?
Escalation triggers catch most of it before the caller does, and transcript review catches the rest within the week. The bigger protection is scope: an agent constrained to a narrow job with grounded answers has far less room to be wrong than one asked to handle everything.
Do we need a contact centre contract to have one built?
No. Voice agents are built as standalone projects, and plenty of clients run one with their own staff on the escalation path. Combining them makes sense when you want the escalation to reach someone who is already awake.
Related
Next step
Find out what this would look like for you
A 20-minute call. We ask what breaks, what it costs you, and who handles it now. If we are not the right fit we will say so.