My mother-in-law, Mother Nancy, uses a local pickup-and-delivery laundry service that, like many, runs everything through an app (For the record, she lives in a neighborhood that our stores don't service 😉). About three or four orders in, with weekly recurring service set up, the card on file rejects the purchase.
She tried to update it online. The app wouldn't let her. She tried the AI chat function. The chat connected her to the operator's support line, who routed the message to an answering service, who took down a callback note. When she called the store directly, someone picked up and immediately hung up. She tried again. Same result.
This went on for a couple days.
Her clothes were sitting at the laundromat. Clean. Folded. Bagged. Waiting for a delivery she couldn't unblock because no human being would pick up the phone or follow up on the message. By the third day, she described the situation in her own words. She said she was "caught up in the AI purgatory." She used the word "purgatory."
For the record, we use AI in this business. Our phone system uses it. This isn't a complaint about AI. It's a question about where, in our specific business model, AI's failure mode is recoverable, and what it costs when it isn't.
Two Dashboards, One Story
What Mother Nancy experienced and what the operator's dashboard showed were not the same story.
Her experience: three days, multiple channels, no resolution, and frustration that built to the point where she told me she was thinking about finding another service going forward. Her recurring weekly order was, at that moment, one short conversation away from becoming a recurring weekly order somewhere else.
The operator's dashboard, almost certainly, showed something different. AI handled an inquiry. Ticket closed. No complaint filed. From where the operator sat, this looked like a managed interaction.
Both views are accurate from where each party sits. Neither captures what's actually happening to the business.
The client is in crisis. The operator is reading the wrong metric. Only one of them will find out it wasn't the complete picture.
The Number Nobody Sells You
The customer service industry has two different metrics for AI performance. Most owner/operators have only been told about one.
The first is deflection rate. It measures whether the AI handled the conversation without escalating to a human. This is the number vendors put in their pitch decks. It's what shows up on the operator's dashboard.
The second is resolution rate. It measures whether the customer's problem was actually solved. This is the number nobody is selling you.
These two numbers can diverge by thirty to fifty percentage points.¹ According to McKinsey's 2026 service operations data, the industry average AI resolution rate sits at 44.8 percent. Action taking AI agents with proper handoff architecture reach 80 to 93 percent resolution. Legacy chatbots, the kind a small independent operator is most likely to deploy, top out at 10 to 30 percent.
When deflection runs high and re-contact runs high simultaneously, that combination has a name in customer experience analytics literature. It isn't called automation. It's called cost deferral.² The customer wasn't helped. The ticket was reclassified.
When a vendor tells you their AI handles 60 percent of customer service interactions, they may be telling you the truth about deflection. That number tells you nothing about resolution. In our business, those are the two numbers that decide whether the client comes back next week.
The Silent Twenty-Five
When clients don't complain, that isn't success. It's the most expensive sound a business can hear.
The foundational research goes back to TARP's work for the White House Office of Consumer Affairs in the 1970s, replicated repeatedly since. The headline finding, for every customer who complains, twenty-six others are unhappy and silent. Of those silent twenty-five, ninety-one percent never come back.³ Recent data from Netigate sharpens the point further, eighty five percent of customers who left a provider say they would have stayed if their problem had been addressed.⁴
The implication for an owner/operator running AI first customer service, if one complaint reaches your inbox, the right number to hold in your head is approximately twenty-five. Not one. Twenty-five other clients had a similar experience and said nothing.
This is where the deflection metric becomes actively dangerous, not just imperfect. The AI is optimized to produce silence. Silence registers in the dashboard as resolution. But silence is the danger signal in client behavior, not the win condition. The metric and the reality are inverted.
If I hadn't gotten involved in Mother Nancy's situation she would have become one of the silent twenty-five. The operator never would have known why and no dashboard would have explained it.
Why Our Business Carries More Risk
Every business with AI customer service faces the metric problem. PUD laundry faces it with