Every business that adds an AI customer agent hopes for the same thing: fewer tickets reaching a human. The first generation of chatbots delivered that number by deflection, answering with a help article, a link, or a rephrased version of the question, and customers learned to skip straight past them. The deflection was real; the resolution was not.
An agent that actually resolves needs two things the old bots did not have: access to the customer's real situation, and the ability to act on it.
Knowing the customer
Resolution starts with context. If a customer asks where their order is, an agent that can look up that specific order, see it is stuck at a depot, and say so is useful. An agent that responds with the general shipping policy is not. Connecting the agent, read-only is fine to start, to the order system, the CRM, and the booking calendar is what moves it from FAQ to support.
Taking action
The next step is letting the agent do the thing the customer wanted: reschedule the appointment, resend the invoice, apply the credit, update the delivery address, cancel the order within the allowed window. Each of these is a defined action with clear rules, and each one is a ticket that never reaches a human. This is where tool-using models earn their place, the conversation and the action happen in the same exchange.
The guardrails that make it safe
Giving an agent the ability to act means giving it the ability to act wrongly. The pattern that works: the agent can do low-stakes things freely (send a document, check status), needs to confirm medium-stakes things with the customer explicitly (I'm going to cancel order 4471, yes?), and simply cannot do high-stakes things (refunds above a threshold, account changes) without a human. The rules live in the system, not in the prompt, because a prompt can be argued with.
Measuring the right thing
Deflection rate flatters everyone. The number that matters is resolution rate, conversations that ended with the customer's problem actually solved and no follow-up ticket, plus how often customers re-contact within a few days. If that second number climbs, the agent is deflecting again, just more politely.
Where to begin
Start with read access and a handful of safe actions on your highest-volume, most repetitive request type. Prove the resolution rate on that one thing before widening scope.