Automate customer support with an AI agent

What can an AI support agent do, and what should it not do?

A well-built agent resolves the common, well-defined questions and escalates the rest with full context. The point is not to remove humans; it is to stop humans from re-typing the same answers all day. In practice the agent owns the first reply on every conversation, closes the ones that are genuinely routine, and hands the rest to a person with the history, the customer record and a suggested answer attached.

What it should not do: promise refunds or exceptions outside policy, argue with an angry customer, or answer a question it cannot ground in your data. Those cases are designed as escalations, not as things the model tries harder at.

How is the agent grounded in your information?

The agent answers from your actual policies, catalog and records using retrieval: before it writes a reply, it looks up the relevant passages from a knowledge base built from your documents and, where needed, the live record (order status, account balance, appointment). It writes from what it found, cites it internally, and where it is unsure it says so and hands off. That is what keeps it from inventing a return policy you do not have.

The knowledge base is the real deliverable. Building it means collecting the FAQs, the policies, the price list, the tone you want, and the exceptions your best support person carries in their head, then structuring them so the agent can find the right answer every time. When your policies change, the knowledge base changes and the agent follows; there is no retraining.

Which support workflows should you automate first?

Start where volume is highest and the answer is most deterministic: order status, business hours and locations, pricing and availability, how-to questions about your product, appointment changes. Those are usually half or more of inbound volume and the agent can close them on the first reply. Complaints, refunds outside policy and anything with legal or medical weight go to a person from day one, with the agent doing the intake and the routing.

If most of your inbound is new-customer inquiries rather than existing-customer support, the AI Intake Agent is the better-fitting build: it qualifies, answers and books rather than resolving tickets. The two share the same knowledge-base engine, so starting with one does not lock you out of the other.

What does it cost and how long does it take?

A fixed price agreed up front after a free 15-minute discovery call. Support automation on WhatsApp and email with a knowledge base and CRM sync typically takes three weeks: a week to collect and structure the knowledge, a week to build and connect the channels, a week to test against real past conversations and go live behind an approval queue. Where the scope matches the productized AI Intake Agent (published at USD 4,900 with an optional USD 1,200 a month operation), that list price applies; otherwise the build is quoted to scope. Infrastructure, messaging and AI usage are billed to you, in your name, with a cap declared in the proposal.

After launch, the 60-day operational guarantee applies to the metric agreed on the first call, for example first-response time under one minute on the channels in scope, or a share of routine conversations resolved without a person.

How does the handoff to a human work?

Every conversation the agent cannot close is routed to a person in the tool your team already uses (the shared inbox, the helpdesk, a WhatsApp group, the CRM), with the full transcript, the customer record and a suggested reply. The person answers from there, and the answer feeds back into the knowledge base if it reveals a gap. Nothing is lost between the agent and the human, and the customer never has to repeat themselves.