Six in ten heads of customer service abandoned their first AI platform within a year and a half. Not because AI doesn't work. Because they bought a promise and got a chart.
I say this as a vendor to whom this can also happen. Below is how to avoid it, including with us.
Gartner reports that 63 percent of customer service leaders dropped their first AI platform within 18 months. The two most common reasons: results stopped growing after the first weeks, and nobody understood what they were paying for. Zendesk looked at it from the other side: of the AI agents built in a given half-year, only about one in ten was still in production six months later.
Start with deflection, because most promises rest on it. Vendors write in their materials that the bot handles 30 to 60 percent of cases on its own. Independent surveys measure 10 to 25 percent. The 2026 median is 41 percent, the top quartile 59. The gap between claim and measurement reaches 30 to 40 percentage points, because the vendor shows its best deployment and you get an average one.
There is one more trap in that number. Deflection goes up when the customer doesn't write a second time. A customer who got a wrong answer doesn't write a second time either. If deflection is rising and so is the number of customers coming back through another channel, you are measuring an illusion.
Four things that separate a deployment that survives from one that gets thrown out after a year:
- A baseline before launch. How many cases a day, how many minutes per case, what cost. Without it, after six months nobody can prove anything improved, and then the budget beats the tool.
- A quality test on your questions, not on the vendor's demo. Before the agent speaks to a customer, it should pass an exam built from your documentation. If the vendor can't show the result of such a test, they can't show what breaks after a change in the knowledge base either.
- A hard time limit on the pilot. Eight to twelve weeks. If in that time the result doesn't beat the baseline, change vendor or close the project. Extending the pilot "by just one more month" is the most common road to those 63 percent.
- A price you understand. The cost of one conversation at your volume and what happens when volume grows. McKinsey reports that one in five companies already limits AI use because of operating costs. Usually those are the companies that learned about the costs from the invoice.
A limitation that applies to us too: no vendor will maintain quality without someone on your side who looks at the report once a week and approves the fixes. It can be done in fifteen minutes. It can't be skipped.
If you hear "our deflection is 70 percent", ask what share of customers wrote an e-mail or called after talking to the bot.
Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- Gartner, survey of customer service leaders 2025
- Zendesk ProductLab 2025; Zendesk CX Trends, Salesforce, Forrester 2025 (deflection benchmarks)
- McKinsey, The State of AI in 2026: On the road to ROI: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai