AI Customer Support Agent: Specification

What this agent is

A customer support agent that answers from your own documentation, cites what it used, and hands off to a person the moment it is out of its depth — on your website, and on the messaging channels your customers already use.

It is not a decision-tree bot with a language model bolted to the front. The distinguishing property is grounding: if the answer is not supported by retrieved content from your corpus, the agent does not answer. Refusal and handoff are the designed behaviour, not the error case.

The rule it is built around

An ungrounded model will state a refund policy you have never had, in your brand voice, to your customer, in writing. One screenshot of that costs more than the project. Every constraint in this specification exists to make that outcome structurally difficult rather than merely unlikely.

Oversight and failure modes

The corpus is the product. Answer quality tracks documentation quality and nothing else. The weekly list of questions the agent could not answer is the most valuable output the system produces, and it belongs to whoever owns the docs.

Silence during incidents. During an outage, ordinary retrieval is misleading — the docs say the feature works. The agent needs an incident mode that suppresses automated answers on affected areas and routes to a person with a status message.

Tone in the wrong moment. A cheerful, well-cited answer to an angry customer is worse than no answer. Sentiment routing to a human is a requirement, not a refinement.

Channel policy. Business messaging channels have template and session rules of their own. Breaking them risks the channel, not just the conversation, so the messaging integration is a compliance surface as much as a technical one.

Language. The underlying models are multilingual; your corpus probably is not. Answer quality follows the corpus, so a documentation set in one language answers best in that language, and the agent should say so rather than translate its way into a wrong answer.

What it might save

Dependent on documentation coverage and contact mix, so read it as an estimate: where the corpus covers the common question types, planning for roughly 15% to 40% of contacts resolved without a human in the first quarter is reasonable, alongside a first response measured in seconds. Thin documentation produces better triage and very little deflection — which is still worth having, but it is a different business case.