AI Customer Service · Resolution not deflection
AI Customer Service Satisfaction: Why Companies Say 86% and Customers Say 51%
86% of companies believe their customers are satisfied with their AI customer service; 51% of consumers agree (Twilio, 9 November 2026, n=7,652 consumers and 660 business leaders, 18 markets). That 35-point gap is the clearest measure yet of AI customer service satisfaction as customers actually feel it.
8 min read · Last updated 13 November 2026 · Checked 13 Nov 2026
Compiled by Rob Jiang, Chief AI Engineer. First published 24 August 2026.
Twilio's 35-point gap is not proof that AI fails at service; it is proof that companies grade AI on containment and on surveys of customers who stayed, while customers grade it on whether the problem got solved.
- →Twilio 2026 Customer Insights Series: 7,652 consumers who bought online in the prior six months and 660 director-level-or-above leaders at companies with 500+ employees, 18 markets, fieldwork 9 April–1 May 2026, released 9 November 2026 (press release).
- →Corroborating study: Pegasystems/YouGov, 4,700+ consumers in North America and the UK, February 2026 (CX Dive).
- →Aissist figures: 83% average resolution and 4.8/5.0 CSAT, as published on the AI customer service benchmark. Nothing on this page is derived from them. The two illustrative charts are not measured datasets.
- →Checked 13 November 2026. Disclosure: Aissist sells AI customer service agents; Twilio sells customer-engagement software. Both have an interest. Figures are quoted as each source published them.

The Data
Are customers happy with AI customer service?
About half are: 51% of consumers say AI-powered service is satisfactory, against the 86% of companies who believe it is, according to Twilio's 2026 study. The gap is not confined to satisfaction. On four questions Twilio asked both sides, companies overestimate what customers get or underweight what they want.
| Question (Twilio, Nov 2026) | What companies say | What consumers say | Gap |
|---|---|---|---|
| Are customers satisfied with AI customer service? | 86% believe yes | 51% agree | 35 pts |
| Should AI identify itself? | 32% actively prioritize transparency | 76% want the AI to say so up front | 44 pts |
| Do customers want more human-like AI? | 73% think so | 54% do | 19 pts |
| Is there an easy route to a person? | 47% offer easy opt-outs | 63% want one at any time | 16 pts |
Rows two to four pair related but differently worded questions; the 47% comes from Twilio's automation report.
The behaviour matches the sentiment. Twilio reports that 78% of consumers have tried to bypass an AI agent to reach a human, and that 91% of brands have seen satisfaction drop because of AI errors such as lost context. Pegasystems' YouGov survey, reported by CX Dive, found 46% of consumers say AI service rarely or never leads to a successful outcome.
None of this is a slow-adoption story. Twilio says AI agents already handle half of service interactions and expects 63% by 2027. The gap is growing at the same rate as the volume.
“Right now, too many organizations are rating their AI on a scale that their customers aren't actually using.”
The Shape
What is the resolution–CSAT tradeoff?
The resolution–CSAT tradeoff is the drop in customer satisfaction that follows when an AI's resolution rate is raised by blocking or delaying human escalation.
The tradeoff comes from escalation design, not from automation itself. There are two ways to push a resolution number up. One is to solve more problems. The other is to make the alternative harder: hide the “talk to a person” option, loop the customer through clarifying questions, close the chat when they go quiet. Both raise the rate. Only the first raises satisfaction.
We call the difference the deflection penalty: the satisfaction lost on conversations the dashboard counts as resolved but the customer counts as abandoned. It is invisible from the inside, because the deflected customer has stopped talking to you. They tell Twilio instead.

That is why the 35-point gap does not mean AI is bad at service. It means the scoreboard and the game are different sports. A company reading its own containment rate sees success; the customer who gave up after the third loop sees a closed door. Deflection and resolution look identical in a report and feel opposite in a conversation.
The practical consequence: past a certain point, chasing a higher resolution number with friction lowers the number that predicts whether customers come back. Earned resolution has no such ceiling.
The Measurement
Why do company and customer satisfaction numbers disagree?
Company metrics are built from the customers who finished the conversation; consumer surveys include the ones who didn't. Each standard metric hides a specific customer experience:
| Company-side metric | What it counts | The customer experience it hides |
|---|---|---|
| Containment rate | Conversations that never reached a human | Customers who gave up, called instead, or churned |
| Deflection rate | Contacts diverted to self-service | Whether the article or bot answer actually worked |
| CSAT on resolved tickets only | Ratings from conversations marked solved | Every escalated, abandoned or re-opened conversation |
| Post-chat survey response | Customers who stayed and chose to answer | Frustrated customers, who rarely stay for the survey |
Twilio's own content team names the mechanism plainly:
“A post-interaction CSAT survey only reaches people who stayed to the end.”
A worked example, with illustrative round numbers rather than measured data: an AI handles 1,000 chats and contains 700. Of those, 450 are genuinely solved and 250 end with the customer leaving. The survey fires at the end of completed chats, so nearly all responses come from the 450. CSAT reads high and containment reads 70%, with the 250 people who had the worst experience counted as contained successes. They show up only when someone like Twilio asks consumers directly.

Wasted time compounds it. Twilio found an ineffective AI agent spends 81 seconds per interaction on failed attempts before handing off. That is a long time to be stuck with something that won't let you leave.
The Fix
How do you measure AI customer service satisfaction honestly?
Measure satisfaction across every AI-handled conversation, count only problems that stay solved, and treat a fast human handoff as a success. Three rules close most of the gap:
Report abandonment beside containment
Chats that end without a confirmed answer tell you what containment alone conceals.
A re-opened ticket wasn't resolved
If the customer comes back within seven days on the same issue, take it out of the resolved count.
Hand off the moment they ask
Pass full context so nobody repeats themselves. Per Twilio, 63% of consumers want that route available at any time.
The full definitions are in AI resolution rate. Then widen the survey. Sample customers who left mid-conversation, not only the ones who stayed, and track a relationship metric alongside transactional CSAT — the NPS vs CSAT breakdown covers when each fits. The broader set of customer service KPIs worth tracking is longer than most dashboards.
“AI should enhance service quality on the journey towards operational efficiency, not compromise it.”
What this looks like in practice: in its published customer story, eSIM provider Holafly reports a 75% resolution rate in customer service alongside 4.8 out of 5.0 CSAT. The resolution number is not the highest on our customer-stories page. The satisfaction number is. That pairing is what earned resolution looks like.
Across deployments, Aissist reports an 83% average resolution rate and 4.8/5.0 CSAT on its benchmark page. We'd rather be measured on both together; either one alone is easy to game.
The Verdict
The gap is a measurement problem before it is an AI problem
Twilio's 86-versus-51 split shows companies and customers grading the same AI on different tests. The company test counts who stayed; the customer test counts who got helped. As long as dashboards reward containment, AI customer service satisfaction will look better from the inside than it feels from the outside.
The fix starts with the scoreboard, not the model: count abandonment, discount re-opens, survey the people who left, and make reaching a human easy. Resolution earned that way doesn't trade against CSAT. Resolution bought with friction always will.
FAQ
Frequently asked questions
- Are customers satisfied with AI customer service in 2026?
- About half are. Twilio's 2026 study of 7,652 consumers found 51% consider AI-powered service satisfactory, while 86% of companies believe their customers are satisfied.
- What is the AI expectations gap?
- It is Twilio's name for the 35-point difference between companies' belief that customers are satisfied with AI service (86%) and consumers' actual satisfaction (51%), from its November 2026 Customer Insights Series.
- Does a higher AI resolution rate always lower CSAT?
- No. Resolution earned by solving problems raises CSAT. Resolution raised by blocking or delaying human escalation lowers it, because the deflected customer is counted as resolved but leaves unhappy.
- Why is my AI CSAT score higher than customer sentiment?
- Post-chat surveys mostly reach customers who stayed to the end. Customers who abandoned, escalated or re-opened a ticket are under-sampled, so CSAT on AI-handled conversations skews high.
- How many consumers try to avoid AI agents?
- Twilio reports 78% of consumers have tried to bypass an AI agent to reach a human, and 63% want an easy way to reach a live person at any time.
- What resolution rate vs CSAT should a support team aim for?
- Track them together, never one alone. Aissist reports an 83% average resolution rate with 4.8/5.0 CSAT; the useful target is the highest resolution rate at which CSAT holds steady.
Resolution and CSAT, measured together.
Aissist's deployments average 83% resolution with 4.8/5.0 CSAT — what the two look like measured together when escalation is part of the system, not something the customer has to fight for.
The two metrics decoded, and why a closed chat is not a solved problem.
Why containment lumps resolved, abandoned and silently failed conversations into one number.
Resolution rate, CSAT & cost across 6 industries — every vendor claim linked to its source.
The two-curve chart and the 1,000-chat example are illustrative, not measured datasets. Third-party figures are attributed to their publishers. Product figures are Aissist's own; results vary by channel mix, intent complexity, and how each team defines resolution.
Rob Jiang
Chief AI EngineerRob is chief AI engineer at Aissist.io, with two decades of experience building conversational and agentic AI systems.