Deflection is the one that leaders reach for because it visibly shows the AI handling volume. But it’ s measuring the wrong thing.
THE DEFLECTION TRAP
Deflection counts interactions that are handled by automation. But it says nothing about whether the customer ' s issue got solved, like in the all-too-common scenario I outlined.
Within that gap lies the damage. Gartner research( eesel AI) puts AI deflection above 45 % of incoming queries, while only about 14 % of deflected queries reach full self-service resolution.
That is a huge 31-point gap: interactions that were touched by automation, marked as handled, and left unresolved. Every one of them is a customer who is about to come back. But almost nobody is tracking that rate of return, and it’ s expensive.
WHAT STARTING OVER COSTS
Our analysis of 33.4 million customer conversations across more than 50 organizations, published this year, put a number on it.
In a comparison between customers whose issues were handled in a single contact, and customers who had to open new interactions about the same problems, the ones forced to start over scored 11.2 % lower on satisfaction.
The penalty held across every channel. And it got worse depending on where the customer had to re-engage.
Email was the worst on both counts. One in five email conversations produced a repeat contact, and those repeats lost 17.7 % in satisfaction, the steepest drop of any channel.
In email, a repeat is not more backand-forth on an open thread. It means the prior thread failed completely, so the customer had to reopen a new one, retype the issue, and re-attach the documentation.
When the first email did not resolve the issue, customers came back at 2.3 times the rate. Email also happens to be the channel absorbing the most AI-generated responses right now. If automation is closing tickets without confirming resolution, this is where it costs you first.
THE REPEAT CONTACT STILL HAPPENS; IT JUST LANDS IN ANOTHER CHANNEL ' S NUMBERS OR IT DISAPPEARS ENTIRELY BECAUSE THEY QUIT.
THE FAILURES THAT LOOK FINE
Here is the finding that should change how you run quality assurance( QA): the biggest satisfaction hit didn ' t come from tense, obviously bad conversations; instead, it came from the ones that looked fine.
We found that in interactions with normal sentiment, the repeat-contact penalty was roughly seven times larger than in conversations that were already negative.
It comes down to the expectations:
• A customer who was already frustrated has( subconsciously) priced in the friction, so a callback just confirms what they expected.
• A customer whose first interaction left them feeling like the issue had been addressed had no reason to expect a second round.
So, when the problem resurfaces and they have to start from scratch, the gap between what they expected and what they got is at its widest.
METRICS
Think about what that means for a program built on sampled reviews. A closed ticket with a professional tone, but with an unresolved outcome, sails through QA. So, while it may look like a good interaction, it ends up being your most expensive one.
Sampled QA is structurally blind to this because the failure is invisible inside any single transcript. You only see it when you connect the first contact to the second.
WHY YOUR CLEANEST CHANNELS LIE
Chat posted the lowest repeat rate in the study, roughly half of email ' s, despite having the lowest first contact resolution( FCR) rate of any channel measured.
Logic would say that a channel that resolves the fewest issues on the first try should generate the most follow-up, but chat generated the fewest.
THE METRIC THAT MATTERS MOST
The industry is spending more on AI-driven deflection than at any point in its history, and the number most vendors lead with keeps climbing.
But the most important data tells you how many customers had to come back and say it twice. Fix resolution and context together and that number falls.
If you celebrate deflection without measuring repeat contact, you are optimizing for a metric that hides your most expensive failures.
The likeliest explanation is channel exit. A customer who does not get resolved in chat rarely comes back through it. They call, email, or give up. The repeat contact still happens; it just lands in another channel ' s numbers or it disappears entirely because they quit. Either way, it never shows up in chat ' s stats.
The channel that looks cleanest can be the one pushing its failures somewhere else. So, if you evaluate channels in isolation, you will reward the one that is best at hiding the problem.
CONTINUED ON PAGE 26
NOVEMBER 2026 23