1st Annual
The State of AI Customer Support
AI is now answering your customers before your team even sees the ticket. This year, we analyzed data from thousands of Crisp workspaces to understand what AI-powered support actually changes, and what it doesn't (yet).
Introduction
Support teams adopted AI tools faster than they redesigned the workflows around them. In most workspaces we studied, an AI agent is already first in line, drafting, resolving, or routing, while the operating model behind it still assumes a human reads every conversation.
That mismatch is the story of this report. Where AI is measured as a feature toggle, the gains look uneven. Where it is measured as a share of volume handled end-to-end, a pattern appears: speed is the obvious win, and capacity is the quiet one.
The pages that follow are a benchmark, not a product tour. Figures are workspace-level medians from Crisp production data. Product mentions are limited to where a mechanism helps explain a result.
Questions this study answers
- How much support can AI resolve without a human?For teams running an AI Agent in production, AI now closes around 40% of conversations without human involvement.
That share has grown significantly over the period studied. AI-alone conversations rose from 22% in January 2026 to 40% in August, showing that AI is moving beyond answering the first message and increasingly resolving complete support interactions. - How much faster is AI than human customer support?
- Do smaller support teams benefit more from AI?
- Does AI reduce the workload handled by human agents?
- What happens to human response times after AI takes over simpler requests?
- How much does an AI support conversation cost?
Methodology
Who's behind the data
Sixteen months of conversations from Crisp workspaces on the Plus and Essentials plans, May 2025 to August 2026. No figure is cited from fewer than 30 teams.
27,318
workspaces analyzed
484.8M
conversations analyzed
16
months
Team size
How the workspaces in this study break down by support team size.
- 1–5 agents28%
- 6–15 agents24%
- 16–30 agents18%
- 30–50 agents14%
- 50–100 agents10%
- 100+ agents6%
Sector
How the workspaces in this study break down by industry.
- Technology35.3%
- Retail & ecommerce11.4%
- Marketing & media7%
- Professional services6.8%
- Education6.2%
- Other33.3%
Geography
How the workspaces in this study break down by country.
- United States24.4%
- France16.8%
- United Kingdom6.3%
- Brazil4.1%
- Canada3.3%
- Other45.1%
Flagship finding
Four in Ten, and Climbing
In a workspace running an AI Agent, close to 4 in 10 conversations now close without anyone on the team seeing them, and that share is still rising.
This is the adoption story hiding in plain sight: AI support is not a slow rollout waiting for a plateau. Among teams running AI Agents, the share of conversations resolved without a human is still climbing fast, and August is not a peak. The number worth repeating is not only how high it sits today, but that it has not stopped rising.
January to August is the window that matters. These teams were already looking for more automation than the rest of the market. What AI Agents changed is the rate: conversations closed without a human go from 22% to 40% in eight months, and the curve steepens from February onward as more of the workload transfers to AI every month.
The rest of the market is frozen, which is what makes the slope real. Teams without an AI Agent still close 11% of conversations without a human, unchanged from a year earlier, because rule-based automation had already found its ceiling and AI has not been bound by it. Two support economies now run on the same chart: one where AI already handles more than three times as much of the queue and is still taking more, and one stuck at one conversation in ten. August is a snapshot of a curve that is still rising, and nothing in the latest months says the first group is done.
40%
Conversations closed without a human in August 2026.
Share of conversations handled with no human
The two lines compare, month by month, the share of conversations closed with no human on teams with an AI Agent in production and on teams without one.
- Teams with an AI Agent in production
- Teams without an AI Agent
Seconds, Not Minutes
The median AI first response is 7 seconds, while the median human first response is 14 minutes.
Customers do not experience support as a median on a chart. They experience the wait before anyone answers, and that wait is what decides whether they feel heard or open a second channel, abandon the purchase, or write the complaint while they are still waiting. When AI Agents take the first turn, that wait is measured in seconds instead of minutes, which for a company is the difference between acknowledging a customer before frustration starts and leaving them in a queue that a person has not even opened yet.
That gap did not close over the period we studied. AI first responses stayed in seconds while the human median moved from about 11 minutes to about 14, and the work that still needs a person is not getting faster. Hiring more agents can shave minutes off a queue, but it cannot move first response into seconds, so speed in customer support is no longer a staffing target but a decision about whether AI is allowed to answer first. Companies that make that decision stop competing on wait time, while companies that do not are still selling a minutes-scale experience in a market that has already moved to seconds.
7s
Median AI first-response time in August 2026, versus 14 minutes for humans.
Speed is the first gap customers feel, and closing it is a workflow problem: who takes the first turn, and how fast the handoff happens when AI cannot finish.
Median first-response time, AI
This line shows, month by month, how long it takes AI to send a first reply, in seconds.
Median first-response time, human
This line shows, month by month, how long it takes a human agent to send a first reply, in minutes.
The Smaller the Team
The smaller the team, the more customer support AI actually handles.
Across every team size we studied, AI took on a larger share of support over time. But the effect is strongest at the smallest end of the market. By the end of the period, teams with five agents or fewer were closing roughly 4 in 10 conversations without human involvement. The data does not show small teams waiting for more headcount before automating. It shows them using AI to extend the capacity they already have.
The biggest shift came from teams of two to five agents. Their autonomous resolution rate rose from 12% to 39%, a 27-point increase. Single-agent teams followed closely, moving from 20% to 44%. In practical terms, AI is approaching the workload of an additional support layer for these teams: a substantial part of the queue can now be handled before the only agent, or handful of agents, needs to step in.
Larger teams in the sample followed the same direction, but at a lower level. Teams of six to fifteen agents moved from 9% to 30% autonomous resolution. AI-powered customer support increases across every group; what changes with team size is how much of the queue gets delegated.
44%
Conversations closed without a human on one-agent teams, August 2026.
Teams of two to five rose from 12% to 39%. Teams of six to fifteen reached 30%. The smaller the roster, the more of the queue AI takes.
Share of conversations handled with no human, by small-team size
The three lines compare, month by month, the share of conversations closed with no human on teams of one, teams of two to five, and teams of six to fifteen.
- Teams of one
- Teams of 2–5
- Teams of 6–15
Where the Work Went
On teams running AI Agents, conversations handled entirely by humans fell from 45% to 16%.
The work did not disappear, it moved upstream. Sixteen months ago, almost half of conversations on these teams still started and ended with a human and no AI at all. That share is now 16%, so the operating model has flipped: most of the queue no longer waits for a person to pick it up. For a company, that is a change in who handles demand first, and the shift is still underway.
The hybrid path is the tell that this is not a greeting layered onto the same human load. Conversations that start with AI and then need a person stayed around a quarter of the mix, while conversations closed by AI alone rose from 13% to 40%. Capacity is not coming from humans working faster, it is coming from conversations that never enter the human queue, and that is why the performance gap is widening. Teams that have not made AI the default still have humans closing 42% of conversations on their own, close to where they stood a year earlier.
Coverage makes the same split visible from the other side. Across the market, unanswered conversations keep rising, and one in three now gets no answer at all, while teams running AI Agents moved the other way, from 19% unanswered to 16%. Customer support is splitting into companies where AI has become the default and more of the same demand gets handled, and companies still running a human-first queue that leave more of it unanswered.
16%
Conversations handled entirely by humans in August 2026.
How a conversation is handled, teams with an AI Agent
The three lines show, month by month, the share of conversations handled by AI alone, by AI then a human, or by a human alone, on teams with an AI Agent.
- AI alone
- AI then human
- Human alone
Do Humans Get Better, or Just Less Busy?
Human response time rises as AI takes more of the frontline workload.
AI escalation itself does not appear to create a meaningful response-time advantage for human agents. Once an AI Agent hands a conversation over, operators typically take around one hour to respond, and that figure remains broadly similar across teams regardless of whether they have deployed an AI Agent in production. The handoff, in other words, is not where the productivity effect shows up.
The more revealing pattern appears in conversations humans handle without AI. Before the AI Agent release, future adopters behaved much like the rest of the sample. After February 2026, their first-response time rose sharply, reaching roughly 32 minutes by August, compared with around 13 minutes for teams that never adopted an AI Agent.
That divergence suggests AI is changing the composition of human support work. As AI Agents resolve more of the repetitive and straightforward conversations upstream, support agents are increasingly left with cases that require investigation, judgment, or action. The data cannot tell us that those individual conversations are objectively more complex, but the timing is consistent with a queue being filtered: less work reaches humans, while the work that does reach them demands more attention.
32 min
Median first human response on conversations handled without AI.
Humans are handling fewer conversations, but the remaining queue appears to demand more of them. Longer response time here may reflect a shift toward cases that cannot be resolved automatically.
Median first human response, conversations handled without AI
The two lines compare, month by month, how long humans take to send a first reply on conversations they handle without AI: teams with an AI Agent in production versus teams that never adopted one.
- AI Agent in production
- Never adopted
What an AI Conversation Costs
The longest conversations are not the expensive ones.
These figures are what an AI Agent costs per conversation, not the fully loaded cost of a human reply. On the median, every industry in this dataset sits in cents. Professional services is the peak, at $0.14. Entertainment and marketing follow at $0.13. Government and transportation sit at $0.06.
Conversation length does not explain that ranking. Telecommunications runs the longest AI conversations, at 4.15 median messages, and still costs $0.10. Education is next, at 3.05 messages, for $0.09. Entertainment is among the shortest, at 2.18 messages, and among the more expensive, at $0.13. What gets automated, a short status check versus a longer thread, shows up more clearly in message count than in cost.
$0.14
Median AI cost per conversation in professional services, the highest sector in this dataset.
Looking ahead
Where support is headed
AI already replies in seconds and closes about 4 in 10 conversations on teams that run an AI Agent. The leftover queue did not get easier. The next advantage is putting AI on the first turn, then staffing and measuring the work that still needs a person.
Four moves the data supports
- Let AI take the first turn, then make the handoff explicit.The speed gap is 7 seconds versus 14 minutes. That only compounds when the first reply can resolve, not just greet. When AI cannot finish, a late or context-free escalation wastes the work already done.
- Staff and measure the queue AI leaves behind.On conversations humans still handle without AI, first-response time rose to about 32 minutes for teams with an AI Agent, versus around 13 minutes for teams that never adopted. Fewer tickets reach people. The ones that do need more attention. Track that queue on its own.
- An AI Agent is the difference between 11% and 40%.Teams without an AI Agent still close about 11% of conversations without a human, the same as a year earlier. Teams with an Agent are at 40%. Rule-based automation had found its ceiling. Small teams sit higher still: one-agent teams reached 44%.
- Treat the drop from 45% to 16% as a staffing decision.Human-only conversations fell from 45% in May 2025 to 16% in August 2026 on a comparable volume. That capacity is only an advantage if it is spent, on the harder leftover queue, on coverage, or on not hiring yet. Unspent, it disappears into a busier calendar.
AI is not replacing support teams. It is taking the first turn and filtering the queue. The teams who win will let an AI Agent close what it can, then rebuild the desk around the conversations that still need a person.
Get the full report
The complete State of AI Customer Support 2026 includes the charts on this page, the methodology appendix, and the sector cuts that do not fit a single screen. One PDF, built to be cited.
See the same metrics in your workspace
A customer support AI Agent sits behind a large share of the first-response times in this dataset. The report is the benchmark; your inbox is the live version.
Explore AI Agent
