A customer service agent that hallucinates a refund policy or invents an ETA destroys CSAT in one conversation. Most "AI customer service" pitches optimise for resolution rate and quietly let the agent answer questions it should escalate. We don't. The agents we build resolve the easy 40-65% with cited sources, and escalate the rest with full conversational context. They sit on top of Zendesk, Intercom, Freshdesk, or Salesforce Service Cloud. Honest resolution-rate forecasts on the discovery call.
The pattern: agent handles common questions end to end, escalates the rest with full context, cites the source on every substantive answer. Trust earned because the trust is verifiable.
The agent reads your help centre, your release notes, your terms of service. It answers questions in your tone, citing the help article, the policy clause, the release note. Where the customer needs a human, it hands off the conversation with a written summary so the agent does not start over.
The agent on the other end of your support phone number. Sub-second latency, knows your product, follows your scripts, books call-backs when it cannot resolve. Built on the same stack as our Voice AI practice. TCPA scope built in for US engagements.
Inbound tickets categorised, prioritised, and routed in seconds. The agent reads the ticket, checks the customer's account, classifies the intent, applies the SLA, sends to the right queue with a summary. Cuts mean time to first response by 60-90% on routine tickets.
Inside your support team's tools, an agent that answers "how have we resolved this before, and what's the policy?" with cited internal context. Not customer-facing, agent-facing. Stops the policy-knowledge collapse that happens when a senior leaves and the new starter cannot find the doc.
Most "AI customer service" pitches are too optimistic. We will tell you the truth on the discovery call.
Resolution rate is not 100%, and it should not be. A typical mid-market customer service AI resolves 40-65% of incoming volume end to end on day one. Pushing higher than that means letting the agent answer questions it should escalate, which is where customer trust collapses. The right metric is "tickets the agent resolved correctly", not "tickets the agent touched."
Escalation quality matters more than resolution rate. The 35-60% the agent does not resolve must reach the human with full conversational context, customer history, and what the agent already tried. Without that, the customer repeats themselves, the human starts over, and the agent has made the experience worse, not better.
Hallucination is the existential risk. An agent that confidently invents a refund policy, an ETA, or a feature destroys CSAT in one conversation. We build with citation discipline (the same approach as our Data AI practice): if the agent cannot cite the source, it does not answer, it escalates.
Tone matters. Generic AI customer service sounds like generic AI customer service. We tune the agent to your voice (formal, warm, direct, technical, however your team writes) and validate against your existing top performers. If the agent does not pass an internal blind test, it does not deploy.
We have built voice AI in production at Bottie. We know what real users do when an AI agent fails them. The agent will fail less if you build it with the failure cases at the front of the design, not the back.
We do not replace your customer service platform. We add an AI layer that lives inside it.
Your support team is drowning in tier-one tickets and 60-70% are the same fifteen questions answered fifteen different ways.
You have a help centre that is up-to-date and detailed, but customers ask before reading it (because they always do).
You hired a customer service AI vendor and the demo was great but the production rollout hallucinated and you turned it off.
Mean time to first response has crept up and you are about to hire a third tier-one rep, quietly wondering if you should.
Your support phone line is staffed 9-5 UK time and US customers ring at 11pm UK time wanting a human.
Your senior CS staff are burning out doing the same triage their juniors should be doing, but the juniors do not know which is urgent.
We map your ticket volume, your top intents, your help-centre quality, your escalation rules. Pick the first agent shape (chat, voice, triage). You leave with a written specification, an honest resolution-rate estimate, and a build plan.
One agent built end to end (chat, voice, or triage) against a defined intent set. Goes live behind your existing helpdesk. Internal blind test against your top performers before deploy. We hand over runbooks and escalation rules.
We monitor resolution rate, CSAT, and hallucination patterns. Push fixes when output drifts. Retrain as your product evolves. Quarterly red-team review against new failure modes. Model and infra at cost.
The voice side of customer service AI. Live phone-line agents on the same stack as Bottie.
RAG over your help centre and internal docs. Citation discipline that customer service depends on.
Where customer service crosses into billing, refunds, and account-state queries. Wired to your finance stack.
We work with mid-market support and CX teams across the UK, US, and Australia. Cited answers, calibrated escalation, your tone of voice. The agent does the easy 60% so your team does the hard 40% with full context.
