The AI forgiveness gap
Your customers forgive human mistakes. They don't extend the same grace to AI. Here's why that changes everything about how you deploy automation.
There is a stat buried in Medallia’s 2026 State of Customer Experience report that I keep coming back to. When human agents make mistakes, 42% of customers say they find those errors more forgivable than AI errors. Not forgivable in general. More forgivable than the equivalent thing going wrong in a bot.
That asymmetry matters more than most CX leaders have accounted for.
We have spent two years having a conversation about AI efficiency: containment rates, deflection, cost-per-contact. The assumption baked into most of those conversations is that if AI handles the same tasks humans did, customers will judge the outcomes roughly the same way. Faster, maybe. Cheaper, definitely. But similar.
They don’t.
Customers hold AI to a higher standard. And when AI falls short, they are significantly less inclined to let it go.
Why the forgiveness gap exists
Think about how you feel when a human does something wrong in a service interaction. You might be annoyed. But somewhere in the background, there is a quiet understanding that humans are fallible. They have bad days. They misread situations. They sometimes get it wrong despite genuinely trying to help.
AI does not get that credit.
When a bot fails to resolve your issue, the interpretation shifts. There is no bad day to blame. No nerves. No distraction. The technology was either built properly or it wasn’t. The company either equipped it correctly or they didn’t. The failure feels designed-in rather than incidental, and that changes what the customer does with their frustration.
Qualtrics surveyed more than 20,000 consumers across 14 countries in late 2025 and found that nearly one in five people who used AI for customer service saw no benefit from the experience at all. That failure rate is almost four times higher than for AI use in general. And the thing those customers saw as worst: AI is ranked by consumers as among the least convenient, least useful, and least time-saving of any AI application they use. Only “building an AI assistant” scores lower.
This is not a minor gap in satisfaction scores. It is customers saying, clearly, that when AI does not work in service contexts, the experience is actively negative.
The compounding problem: error rates are higher than most leaders think
Close to half of customer service representatives in a recent 360 Magazine analysis said they are regularly fixing mistakes made by AI tools. One in ten admitted they do not even know something went wrong until a customer tells them.
So the situation is this: AI is failing more often than companies realise, and when it fails, customers are less forgiving than they would be if a human made the same mistake. Those two things together create a compounding problem that dashboards tend to obscure.
Containment goes up. CSAT holds steady (the customers who gave up quietly did not fill in your survey). The vendor deck shows green. Meanwhile, a steady trickle of customers who had one bad AI interaction have quietly decided not to come back.
Forrester put a number to this in their 2026 predictions: one in three brands will damage customer trust through premature AI deployment in customer service this year. That is not a fringe outcome. It is the modal one.
Why disclosure makes it worse before it makes it better
Here is where it gets genuinely complicated. The natural response to customers feeling deceived or let down by AI is to be more transparent: tell people when they are talking to a bot. Add disclosures. Build honesty into the interaction design.
That is the right instinct, but it does not fully solve the problem.
Research published in the International Journal of Advertising in 2025 found something counterintuitive. When AI sounds human during a service recovery and then discloses its AI status, customers experience what researchers called an “identity-contingent trust violation.” The empathy and apology they received was reinterpreted as performative. The interaction that felt meaningful became hollow.
In other words, transparency helps when it comes at the start. It creates a different problem when it comes after the customer has already formed a connection.
Most AI deployments do not think carefully about this sequencing. They bolt a disclosure on somewhere because it feels responsible, without considering what it does to the interaction arc.
Service recovery is where AI falls furthest short
Zendesk’s 2026 CX Trends data shows 80% of CX leaders agree that transparency about AI will be non-negotiable this year. Only 37% currently offer customers any reasoning behind AI’s decisions.
That gap matters most in the specific moments where customers most need to understand what is happening: complaints, billing disputes, policy exceptions, service failures. These are exactly the situations where a customer needs to feel that someone, or something, genuinely understands their frustration and has the authority to fix it.
AI gets these technically right and tonally wrong, with some regularity. The resolution might be correct. The process might be compliant. But the interaction leaves customers feeling processed rather than heard, and that is what they remember when they decide whether to stay.
A survey of US customers from Netfor shows 68% still prefer a human agent when their problem is complex or emotional. That is not a preference the industry is successfully moving. It is a structural reality that service design needs to accommodate.
Bruce Temkin, one of the founders of the CX discipline, gave a talk on exactly this territory in 2025 that is worth your time. He argues that AI-driven service without genuine human judgment at the critical moments does not just create bad experiences; it erodes the relational foundation that makes loyalty possible.
Watch it before your next AI deployment review. It is a more useful 40 minutes than most vendor briefings.
What to do with this
None of this is an argument against AI in service. It is an argument for being honest about where it fails and designing accordingly.
Three things that actually help:
First, map your forgiveness risk. Not all service interactions carry the same stakes. A billing enquiry answered incorrectly at 9pm feels very different to a claim declined without explanation. Identify the moments where customers have the least tolerance for error and make an active decision about whether AI should own those or hand off to a human.
Second, set AI up to fail well, not just to succeed. Your deflection rate tells you about volume. Your error recovery rate tells you about resilience. Measure what happens in the 5% of interactions where AI does not resolve the issue. Is the handoff seamless? Does the human pick up with context? Or does the customer have to start again from scratch?
Third, treat disclosure as interaction design, not a legal checkbox. When customers know from the start that they are talking to AI, their expectations shift. They are more forgiving of limitations, more likely to accept a structured resolution, and less likely to feel deceived when the bot cannot handle nuance. The problem is almost never disclosure itself. It is disclosure at the wrong moment, in the wrong tone, with no thought given to what comes after it.
The companies getting this right in 2026 are not the ones with the highest containment rates. They are the ones who have been honest with themselves about what AI cannot do, and have built their service model around that honesty rather than around the vendor’s most optimistic slides.
Worth reading
How to get your customers to trust AI — Harvard Business Review’s January 2026 piece argues that transparency alone is not enough; it needs to be sequenced into the interaction design from the start, not added as an afterthought.
AI-powered customer service fails at four times the rate of other tasks — The Qualtrics 2026 Consumer Experience Trends Report is worth reading in full. The failure rate data is based on 20,000+ consumers across 14 countries and is one of the most credible datasets available on actual AI performance versus expectations.
One-third of brands will damage trust with AI self-service in 2026 — Forrester’s prediction for 2026, grounded in their research on premature AI deployment. The argument is precise: the cost pressure to reduce headcount is driving organisations to deploy AI in contexts where the failure rate is predictably high.
When customers know it’s AI: Comparing human and LLM communication in service recovery — An academic study from 2025 on identity-contingent trust violations. Not light reading, but the forgiveness gap data is more granular here than you will find anywhere else.
If you have been forwarding the AI strategy slides and quietly wondering whether the numbers stack up, this newsletter is for you.
CX Decoded is a weekly read for CX professionals and business leaders who want the practical version of what is happening, not the vendor version. No case studies paid for by the vendor. No recycled frameworks. Just thinking from someone who has done this work.
If this was useful, forward it to one person who should be thinking about this.