An agent that answers from your documentation, your ticket history and your live systems, escalates when it should, and can prove where every answer came from.
Not a chat widget with your FAQ pasted into a prompt. A retrieval system over your real content, wrapped in rules about what it may and may not do, connected to the systems that hold the answer, and measured against a test set built from your own tickets.
Marketing implies 30 to 50 percent. The average team sees 10 to 15 in year one. The variable that closes the gap is not the model, it is whether the system is anchored in real documentation. Grounded implementations reach 85 percent and above.
A confident wrong reply about a refund, a policy or a deadline costs more than a slow human one. Once a customer catches the bot inventing, they stop trusting the channel and every future deflection is lost too.
Where is my order, why was I charged this, did my payment go through. None of these are in a help doc. They are in the order system, the billing system and the product, so an agent that can only read documentation cannot resolve them.
Most deployments specify what the bot should answer and never specify what it must refuse. Refunds, legal commitments, medical or financial advice and anything touching a regulated obligation need an explicit boundary, not a hopeful prompt.
The parts that decide whether this survives contact with real data, real volume and real edge cases.
Content is chunked, embedded and retrieved per question, and the answer is constrained to the retrieved passages. If nothing relevant comes back, the system says so and escalates rather than filling the gap.
Retrieval augmented generationRefusal rules, topic boundaries and action permissions are enforced outside the prompt, so a persuasive customer cannot argue the agent into issuing a refund or making a commitment it should not make.
Guardrails and policy enforcementWe build a test set from your resolved conversations and score every change against it. That is how you get a defensible deflection number rather than a vendor claim, and how a prompt change stops silently regressing quality.
Evaluation harnessLow retrieval confidence, negative sentiment, high account value or a restricted topic all trigger handoff, and the human receives the conversation, the sources consulted and the reason for escalation.
Human in the loop escalationCaching, routing simple questions to smaller models, trimming retrieved context and, where the volume justifies it, fine tuning a smaller model to do the repetitive work. Token spend is tracked per conversation so cost per resolution is a number you can see.
Inference cost optimisationEvery conversation is logged with the sources retrieved, the decision taken and the confidence. When something goes wrong you can replay it rather than guess.
Observability and tracingA 30 minute audit, no sales pitch. We map where this fits in your stack, what it would take to build, and whether it is worth doing at all.