How we would help you put AI into customer service: what to build, what to buy, and the part nobody sells you.
Prepared following your introduction · highintentlabs.com
Sebastien de Bandt · Ali Goldsmith · Nicolas Meibohm
ElevenLabs provides a specific AI model. We are a consulting firm that advises on the implementation of voice agents. We are model agnostic.
Which means we have no reason to recommend one platform over anotherWe help you set up customer service agents on whichever platform fits, or build them in-house. The recommendation follows the use case, not our commercial interest.
We did substantial work setting up voice agents at Checkatrade: choosing the stack, building the knowledge base from scratch, and running the refinement loop that keeps quality from drifting after launch. That last part is where most projects quietly fail.
You add a phone channel you do not run today. Customers who want to speak to someone finally can, at a cost per call that would never work with people alone.
You take your existing assistant further: generative answers for the harder questions, and actions like cancellations and extensions rather than just information.
Both are good projects. They need different platforms, different budgets and a different order of work, so this is the first thing we would settle with you.
Your published contact routes are the in-app assistant, messaging and social. Phone numbers appear for recovery, not for general service.
You are not starting from zero. That changes the question from "should we use AI" to "what is the next step, and on which foundation".
Across public reviews, the recurring theme is how hard it is to reach a person. Price changes after a quote and unexplained declines follow close behind.
We would rather show you what we could see from the outside than pretend to know your numbers. Everything here is a hypothesis until you correct it.
A voice agent done badly makes that worse. Done well, it is the first time you can answer the phone at all.
Pick the platform, pick the calls worth automating, and get one agent genuinely working end to end.
2-3 weeksIt usually does not exist, because the knowledge lives in your agents' heads. Getting it out is the real work.
3-6 weeksReading transcripts, spotting errors, improving answers. Time consuming, and the single biggest predictor of whether this still works in month six.
OngoingEvery vendor assumes you have a knowledge base. Almost nobody does. What exists is a wiki that is two years out of date and a team that knows the real answers. Four methods, used together:
Several hundred real conversations across a full seasonal cycle, to find what customers actually ask rather than what we assume.
Days spent beside your most experienced people, asking "why did you answer it that way" at every edge case.
Structured sessions with your top performers, aimed at decision trees rather than anecdotes.
Wordings, scripts, FAQs. Where they contradict what the team actually does, that gap is itself a finding.
Not because the model degrades, but because policies change, edge cases accumulate and every fix quietly breaks something else. The loop that prevents it:
They give you the voice, not the agent. Cheap per minute, everything else is yours to build.
e.g. ElevenLabsFast to prototype, lowest running cost, but you own compliance, evaluation and maintenance. Check data residency carefully.
e.g. Vapi, RetellEnterprise compliance out of the box and far less operational risk, at six figures a year and real lock-in.
e.g. Decagon, Sierra, PolyAI, Parloa, CognigyIf you already run a contact centre platform, its own AI layer may be the cheapest path. Worth checking before anything else.
e.g. Genesys, Amazon ConnectOur first question is not which one you like. It is what you and Admiral already run, because that removes half the market immediately.
We would rather agree a conservative target now than explain a missed one later. And we build the business case from your numbers, because there is no credible industry benchmark to borrow.
Consumer Duty, vulnerable customer guidance, complaints handling and outsourcing rules apply to an AI conversation exactly as they do to a human one. Three hard limits we build in from day one:
The agent never refuses cover, rejects a claim or commits to a price on its own. Anything adverse goes to a person, which is also where the new automated-decision rules land.
Signs of confusion or distress route straight to a human. Under FG21/1 this is not a courtesy, it is the expectation.
Including the implicit ones. A complaint the agent files as a query is a regulatory problem, not a data quality one.
Plus the part that is easy to forget: using an external platform is outsourcing. Audit rights, monitoring and an exit route belong in the contract, and we would help you put them there.
seb@highintentlabs.com · ali@highintentlabs.com · nico@highintentlabs.com