AI Agents Took Over Support Chat. Email Is 62% of the Work and Still Waiting.

Most of the noise about AI in customer service is about what AI agents could do. A new benchmark looks at what they actually did, inside 316 real support teams, over one quarter.

The Dixa Customer Service Benchmark 2026 is built from platform data, not a survey. It covers 316 customer service teams on Dixa's platform from July to September 2026, compared with the same months in 2025 and 2024. The headline: AI agents have gone from pilot to standard tool in a year. But they mostly answer chat, and chat is a small part of the job.

AI Agents Now Run in a Third of Support Teams

In the third quarter of 2026, 97 of the 316 teams ran an AI agent, up from 53 a year earlier. That is close to double in twelve months.

Usage per team is still small, though. At the typical team that runs one, the AI agent handled 7.6% of all conversations, up from 3.8% a year ago. The top quarter of teams gave it 19.7%. The top tenth gave it 46.3%.

So the gap is not "AI or no AI." It is how much of the work a team is willing to hand over.

The Channel Problem: Email Is 62% of the Work

This is the most useful number in the report. Here is how conversations split by channel, and how much of each an AI agent answered:

AI has taken over chat, which is about one conversation in eight. Email, which is six in ten, is still mostly answered by people. Voice is untouched.

Email is also where customers are least happy. Typical satisfaction was 88 for phone, 81 for chat and 78 for email. The typical team's first human reply to an email took 20.2 hours, down from 26.1 hours a year earlier. The fastest quarter of teams replied in 11.2 hours.

Dixa's own read is blunt: "The next gain isn't a better AI. It's giving the AI the email."

A Realistic Resolution Rate Is About Six in Ten

Vendors like to quote high resolution rates. This benchmark gives a grounded one. Across the 76 teams whose AI agent had at least 50 conversations with an outcome, the typical AI resolved 61.7% of what it was given, up from 55.6% a year earlier. The top quarter passed 78.4%. The top tenth reached 87.4%.

Email is harder than chat here too. The typical AI resolved 62.2% of chat but only 47.8% of email. Hand-offs to a person fell from 42% to 35% of AI conversations.

Dixa's planning advice: build the business case on 60% and the roadmap on 78%. One caveat matters. "Resolved" here means the AI closed the conversation without a person. It is not a customer-confirmed fix.

Did Satisfaction Drop When AI Switched On?

Not measurably. Dixa followed 37 teams that turned on an AI agent between April 2025 and March 2026. Three to six months later, their customer satisfaction was 1.1 points lower than before. Teams that never used AI moved 1.7 points over a similar period. Teams already using AI moved up 1.1. None of those moves is big enough to call an effect.

There is a wrinkle. Teams that give AI a bigger share of the work score lower today: 82 with no AI agent, 81 with AI under 10%, 78 at 10 to 30%, and 74 at 30% or more. Because switching AI on did not move scores, Dixa thinks that gap was most likely there before the AI. Either way, it is a reason to watch quality closely as AI takes on more.

More Conversations Per Seat, but Not Fewer Seats Yet

The typical team handled 546 conversations per licensed seat per month in Q3 2026, up from 512 a year earlier and 436 two years earlier. The top quarter handled 856.

Here is the surprise. The gains do not line up neatly with AI use. Teams that gave AI 10 to 30% of their conversations gained the most per seat (+21.6%). Teams that gave it 30% or more lost ground (-20.8%). Dixa's conclusion: for most teams, headcount is still the main lever, because most teams have not given AI enough of the work to change that.

Sign up for our newsletter

Get the latest updates, product launches, and helpful tips delivered straight to your inbox.

Email addressSubscribe

Peak Season Is Where Teams Already Lean on AI

This matters right now. In 2025, the typical team's busiest month (November or December) was 1.15 times a normal month. For a quarter of ecommerce and retail teams, it was 1.68 times or more.

Teams with an AI agent pushed three times their usual share to it during the peak: 6.9% of conversations, against 2.1% the rest of the year. The typical team also added 7% more licensed seats for those months.

What the Data Can't Tell You

Be honest about the limits, as Dixa is. These are Dixa customers: mid-market, mostly European, mostly ecommerce. Every split covers fewer than 200 teams, so read the splits as direction, not precision. Monthly data hides week-level spikes like Black Friday. And the data cannot separate satisfaction for AI-handled versus human-handled conversations.

Still, it is rare to see measured numbers rather than opinions, and the pattern is clear.

Why Email Is Harder Than Chat for AI Agents

There are practical reasons email lags:

• Emails carry more at once. One message can hold an order number, a complaint, a photo and two separate questions.
• Emails need action, not just answers. A refund, an address change or a replacement means touching the order system, the payment tool and the shipping tool.
• Emails arrive with history. The right reply depends on the last three messages and what the customer was promised.

A chat widget that answers FAQs does not solve any of that. An AI agent that can read the full thread, look up the order, draft the fix and ask a person to approve anything risky can.

Where Workmate Fits

The benchmark's biggest gap, email answered mostly by people and a 20-hour typical first reply, is exactly where a connected AI agent earns its keep. Workmate builds custom AI agents around a team's own workflows and connects them to the apps that team already uses, so an agent can read a support email, check the order, draft the reply and route it, rather than just answering a FAQ.

Because the risky part of email is the action, not the words, control matters. Workmate gives admins control over which apps and integrations each agent can reach, and stores credentials in a vault that can require human approval before an agent acts. That lets a team move from 7.6% of conversations toward the top-quarter's 19.7% one workflow at a time, with a person still signing off on refunds and exceptions.

What This Means for Support Teams Going Forward

The lesson from 316 teams is simple. Chat was the easy win, and most teams have taken it. The next gain is the inbox.

• Measure AI's share of all conversations, not just its resolution rate. A 62% resolution rate on 7.6% of the work is a small dent.
• Start with your highest-volume email types. Order status, returns and address changes are repetitive and rule-based.
• Plan on about 60% resolution, aim for 78%. And track satisfaction by channel as you go.
• Use peak season as the proving ground. Teams with AI already lean on it hardest in November and December.

FAQ

How much of customer service do AI agents handle today? At the typical Dixa team running an AI agent, 7.6% of all conversations in Q3 2026. AI handled 61.3% of chat but 11.4% of email.

What is a good AI resolution rate? About 62% is typical; the best quarter of teams pass 78%, per Dixa's measurements across 76 teams.

Does AI lower customer satisfaction? Not measurably in this data. Teams that switched on an AI agent moved about one point, similar to teams that did not.

What is a typical email first-reply time? About 20 hours at the typical team, and 11.2 hours for the fastest quarter.