Every automated support deployment reports the same headline number. Contacts handled without a human went up, sharply, and the figure is real — the system genuinely answered those questions and the customer genuinely went away.
The number underneath it is the interesting one. Total human contacts fell by much less than deflection would suggest, and in a number of deployments did not fall at all. Both figures are accurate. What sits between them is the shape of the problem.
Deflection and resolution are different measurements
An automated system is extremely good at the top of the distribution: where is my order, reset my password, what are your hours, how do I return this. That is most of the volume and almost none of the difficulty, and removing it is worth doing.
What arrives at a human afterwards is everything else, and it has changed character. The easy contacts are gone, so the queue is now made entirely of the cases that are ambiguous, multi-part, angry, or wrong in a way the customer cannot articulate. Average handle time goes up, because the average is now taken over harder work. Agents who used to get a rhythm of simple tickets between difficult ones get no such relief, and attrition follows.
Then there is the contact the automation created. A customer who spends six minutes with a system that does not understand the question and then calls anyway has generated two contacts, not one, and arrives at the human already annoyed. Organisations that measure only deflection cannot see this at all — the first interaction counts as a success. The ones measuring resolution per issue rather than per contact see it immediately, and it is usually the single most expensive category in the mix.
The instinct is to widen what the automation attempts, which is where deployments tend to go wrong. Pushing it further up the difficulty curve raises deflection and raises the failed-then-escalated rate at the same time, and the second effect is invisible unless someone is explicitly looking for it. This is the general problem of not being able to tell whether the AI works showing up in a function where the metric is easy to produce and easy to misread.
The deployments that hold up have mostly done the unglamorous version: automate a narrow band, hand off early and cleanly with the full context attached, and let the agent start from what the system already established rather than from the beginning. That is close to how back-office agents earned their place — bounded scope, clean handoff, no pretence of handling the exception.
There is a staffing consequence that will take a few years to show. Support was one of the last large functions where a person without a specialised background could start, learn a business from the inside, and move into operations, product or account management. Removing the routine tier removes the training ground, which is the same trade being made across the entry level and booked, again, as a saving.
Topics aioperations
