AI customer service pays off when it changes how the work is organised. Bolt a model onto the existing workflow and you get a faster version of the old process, with the old costs mostly intact. The return comes from the redesign, and the redesign is the part most projects skip.
A familiar pattern: a company adds an AI agent to its support operation, the agent works, and a year later the return is disappointing. The review looks at the model and the vendor. The actual problem is that the operation around the model never changed.
This article is about the operating change that turns an AI deployment into a return.
Why adding AI to the old workflow does little
Kevin Davis, who writes KD Be Schemin, puts the diagnosis plainly: when the return on AI is missing, the redesign is what is missing, not the technology. A company that drops an AI agent into its existing process has automated one step of a workflow that was built for humans. The handoffs, the queues, the metrics, the team structure all still assume the old shape. The model is faster at its step, and the workflow around it absorbs the gain.
The common belief is that deploying capable AI is the project. Deploying it is the easy part. Re-organising the work around it is the project, and it is the part that produces the number.
The change: route decisions, not just tasks
The operating change that works is to use AI to route, not only to perform. In the old design, every contact enters one queue and is worked in order. In the redesigned operation, an AI layer reads each contact first and decides where it should go: resolve it now if it is simple and safe, send it straight to a specialist if it is complex, flag it for a human with full context if it is high-stakes.
That changes the economics. The simple contacts are resolved without a person. The hard contacts reach the right person first time, instead of bouncing through a general queue. The human team spends its hours on work that needs judgement, rather than on triage. The saving is not the model doing a step faster. It is the whole operation sorting its work better.
Set explicit escalation rules, not vague handoffs
A redesign like this depends on being precise about when the AI keeps a contact and when it passes one on. "Human in the loop" as a phrase is not precise enough. The operation needs defined rules: this contact type, above this value, with these risk signals, goes to a person, with this context attached.
Without explicit rules, the routing drifts back to the old default, where everything flows through one queue and the AI just answers first. The escalation rules are the redesign. They are also what makes the system measurable: you can check whether each tier of contact went where the rule said it should.
Where the return comes from
The return on AI customer service is not a property of the model. It is a property of the operation the model sits in. A capable model in an unchanged workflow produces a faster version of the old cost structure. The same model in a workflow redesigned around routing produces a different cost structure.
The decision worth getting right is not which model to buy. It is whether the project includes the work of changing how the operation routes its contacts, because that work is where the return is.