Support automation usually gets a deflection target: reduce contacts by some percentage. The target is met, the contact rate falls, and satisfaction falls with it — because the fastest way to deflect a ticket is to make it hard to reach a person.
That is not a reduction in load. It is a transfer of cost to the customer, and it shows up later as churn.
Sort the volume before you touch it
Inbound contact is not one thing. It is roughly three:
Should never have happened. The customer is contacting you because something upstream is broken, unclear, or missing. A confusing invoice line. A status nobody communicated. A step in onboarding that fails silently.
Should be self-serve. A legitimate need with a deterministic answer, where the only reason a person is involved is that no interface exists for it.
Genuinely needs a person. Judgement, negotiation, an exception, or someone upset.
Only the second category is an automation problem. The first is a product problem, and the third should be protected, not deflected.
The first category is where the money is
Fixing an upstream cause removes the ticket permanently and improves the experience at the same time. Every other approach manages a symptom forever.
This is why the highest-value output of a support review is usually a list of things for other teams to fix, not a bot. Categorise a month of tickets by root cause rather than by topic, and the list is normally short and uncomfortable.
Automate the answer, not the deflection
Where self-serve is right, the goal is to give the answer in fewer steps than emailing would have taken. If the automated path is slower or less certain than typing a sentence to a human, people will type the sentence.
Concretely: answer in the channel they contacted you in, do not force a portal login for a status check, and make the escalation to a person obvious rather than hidden. Counterintuitively, a visible route to a human increases trust in the automated path, because using it stops feeling like a trap.
Watch resolution, not volume
Contact rate on its own can be gamed by anyone with a hard-to-find contact page. The pair worth tracking is contacts per active customer alongside first-contact resolution — falling volume with falling resolution means you hid the button.
The goal was never fewer conversations. It was fewer conversations that nobody should have had to start.