An autonomous AI agent in customer support resolves conversations on its own and hands over to a person only when it has to. Salesforce says its agent resolves more than 63% of Help-site conversations instantly, and other Salesforce pages say 85%. ServiceNow says its agent handles more than 90% of employee IT requests. All of it is self-reported, and the definitions of resolved differ or are missing.
This guide covers what each company has published about running an autonomous agent on its own desk, what the figures count, what went wrong first and how a first 90 days could run. Every figure names who says it and in which document, with the date where the page gives one; undated pages were read on 17 November 2026. One disclosure: we make CoSupport AI Agent, so it gets the last section, with one customer case. Nothing else here is a CoSupport result.

Two companies, their own desks
Salesforce runs Agentforce on its public Help site, for customers. ServiceNow runs its Level 1 Service Desk AI specialist on its own employees' IT requests. The table keeps each source's wording.
| Salesforce, Agentforce on its Help site | ServiceNow, Level 1 Service Desk AI specialist | |
|---|---|---|
| Who asks | Customers, on the public Help site | Employees, through ServiceNow's own IT desk |
| Volume | Over five million customer conversations and "tens of thousands of conversations every week" (Agentforce Customer Zero page); 45,000 a week (Salesforce News, 24 July 2025) | Not stated in the CIO or CX Today reports |
| Resolution figure, in the source's words | "more than 63%" resolved "instantly" (Customer Zero page); "in the 85% range" (Salesforce News, 24 July 2025); 84% (VentureBeat, 18 July 2025); "85% of Salesforce's customer service requests" (AI for Customer Service & Support page) | "more than 90% of employee requests" (CIO, 26 February 2026); "over 90% of employee IT requests" (CX Today, 26 February 2026) |
| What is counted | 63%: resolved "instantly"; the page's weekly counters define resolved as conversations "where Agentforce answered the customer's inquiry", with abandoned and handed-off conversations counted separately. 85% (AI for Customer Service & Support page): "of customer queries without a human". The News story and VentureBeat don't define their rates | Not defined in those reports |
| How it started | A pilot with "just 200 authenticated users within a four-week period" (Customer Zero page) | Not described in those reports |
| Quality measure | Answer Quality of 60% in October 2025 and 76% in its most recent test (Salesforce's answer quality guide) | None in those reports |
| Who vouches for it | Salesforce's own pages; VentureBeat reported it from interviews with Salesforce executives, six days before the News story | ServiceNow's own claim at its February 2026 announcement, as CIO and CX Today report it |
Salesforce's pages carry figures between 63% and 85% for the same desk. ServiceNow's 90% covers employee IT requests, a different desk from customer conversations. Before you compare any of these numbers, check what each one counts as resolved.
sbb-itb-97114f1
What autonomous support means, and what counts as resolved
Autonomous support, also called autonomous customer service, means the AI agent carries a conversation to its end. It reads the question and the customer's account, uses the knowledge base, takes an action such as a refund or an address change, and closes the conversation. A person steps in on a handoff. A chatbot follows a script. A copilot drafts replies that a person still sends. "Agentic AI" is the wider term for systems that act on their own. Agentic AI explained and AI chatbots versus AI agents go deeper, and AI for customer support covers the basics.
The word that decides every percentage is resolved. A conversation can count as resolved under three rules: the agent gave an answer, the customer stayed quiet for a set period after the answer, or a person confirmed it, the customer or a reviewer. Salesforce's weekly counters use the first. They add resolved, abandoned and handed-off conversations into one total, and a resolved conversation is one where "Agentforce answered the customer's inquiry." ServiceNow's reports don't say which rule they use.
Gartner's number belongs in a different bucket. In a press release of 5 March 2025, Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, with a 30% reduction in operational costs. That's a forecast. The two desks above are reports of what already happened. For more forecasts, see the future of customer support.
How Salesforce started: four weeks and 200 users
In week one, Salesforce's team could read every conversation by hand. Its support team spent months planning a staged, four-week rollout of the Help-site agent, "baby steps with limited exposure and extensive monitoring," according to Salesforce News on 24 July 2025. The Customer Zero page says the pilot opened to "just 200 authenticated users within a four-week period." In week one, the rollout covered only 10% of authenticated users, which, in the words of Zachary Stauber, a senior director on the launch team, "gave us fewer than 150 conversations to sift through manually."
A few weeks in, mistakes were rare enough that the CEO encouraged the team to scale more quickly. Ten months after launch, the July 2025 story reports 45,000 conversations a week, more than 1 million in all, "with resolution rates in the 85% range." VentureBeat reported 84% for the same rollout on 18 July 2025, plus 87% resolution in Japanese after three weeks, a handoff rate raised on purpose from 1% to about 5%, so customers reach a person faster, and a 5% reduction in support case volume.
| Stage | What Salesforce states | What a team can take from it |
|---|---|---|
| Plan | Months of planning for a staged, four-week rollout with "limited exposure and extensive monitoring" (Salesforce News, 24 July 2025) | Decide the stages and the monitoring before the first customer meets the agent |
| Pilot | "Just 200 authenticated users within a four-week period" (Customer Zero page); 10% of authenticated users in week one, fewer than 150 conversations to read by hand | Start with people you can identify, at a volume one person can read in full |
| Scale | A few weeks after launch, mistakes were rare enough that the CEO encouraged faster scaling (Salesforce News) | Widen when the transcripts stop showing the same errors |
| Mid-2025 figures | 45,000 conversations a week, more than 1 million in all, resolution "in the 85% range" (Salesforce News); 84% and 87% in Japanese after three weeks (VentureBeat, 18 July 2025) | Report a rate together with its weekly volume and its definition |
| Customer Zero page, read 17 November 2026 (the page shows only "week of Nov 8th", no year) | More than five million conversations, "more than 63%" resolved, seven languages (Customer Zero page) | Date every figure and name the page it came from |
What went wrong first
Salesforce's own account of its early weeks is more useful than its headline rate, because it names what broke and what the team did about it.
| Symptom | Cause the story gives | Response the story describes |
|---|---|---|
| The agent read "live chat" or the product name "Digital Engagement" as a request for a person | Not stated; the agent "grew confused" | Not described; the story says the early work was continued training and fine-tuned instructions |
| It refused questions about integrating Microsoft Teams | An instruction not to discuss competitors, added after it recommended a competitor's product | Moved from "don't mention competitors" to "act in the best interest of Salesforce" |
| It couldn't answer developers' technical questions | The technical documentation hadn't been given to it | Brought the developer-portal content in |
| It cited release notes from 2018 | Likely "content collisions": too many articles on the same topic | Agents culled the content, and anything unused for a year was deleted |
Salesforce draws two lessons from it. First, the team coached the agent. After the competitor mix-up, the team's instruction became "act in the best interest of Salesforce," which the story describes as letting the LLM be an LLM. Second, content is a cause of wrong answers. Bernard Slowey, an SVP at Salesforce, says the problem was likely "content collisions": "When there's too many articles about the same thing, you have poor content hygiene." Stauber puts it more briefly: "more curated content equals better answers."
The other half is handoff. The agent watches for explicit requests for a person and for subtler signals such as profanity or frustration, then hands over with the context attached. Salesforce says an escalation is often "exactly the right outcome" when a customer needs nuanced problem-solving or prefers a person. The mechanics are in AI to human handoff and in working with a human agent.
How Salesforce measures answer quality
Resolution rate says how many conversations ended without a person. Answer quality says whether the answers were any good, and Salesforce tracks both. Its guide to measuring Agentforce defines Answer Quality as "the percentage of answers from Agentforce that meet our acceptance criteria" after a single question, with the assessment focused on the first turn of a conversation. A good answer is "Relevant + Correct + Complete."
The guide gives a benchmark of 60% in October 2025, a target of 75% by the end of the year, and a result: the team "just hit 76%" in its most recent test. It adds that 60% sounds low but is rather high once off-topic questions, small talk and requests to open a case on the first turn are counted.
The test set is built from synthetic questions, which Salesforce calls synthetic utterances: around a hundred at launch, several hundred today and a goal of near a thousand. Agents test the agent, and a new role, the Evaluation Manager, owns the framework, runs the tests and shares the results. The guide puts resolution rate and escalation rate next to answer quality, because answer quality on its own isn't enough.
Accuracy figures depend on the test set and the acceptance criteria, so one company's percentage says little about another's. On your own desk, write down what a good answer is, fix a list of real questions and give one person the job of running it. See measuring and controlling AI hallucinations and reducing inaccuracies.
ServiceNow: an internal IT desk
ServiceNow's figure comes from a different kind of desk. In its 26 February 2026 announcement of the Autonomous Workforce, ServiceNow said its Level 1 Service Desk AI specialist already ran internally. CIO writes that ServiceNow "claims that it is handling more than 90% of employee requests." CX Today, the same day, gives "over 90% of employee IT requests" and says the headline figures are "largely drawn from ServiceNow's own internal deployment."
An internal IT desk and a customer desk differ in ways that move the number. CIO describes the work as common IT requests such as password resets, software access and network troubleshooting, for ServiceNow's own staff. Salesforce's figure covers customers on a public site. The sources don't say how routine the requests were, so neither number predicts yours. Both reports attribute the figure to ServiceNow itself, so treat it as a claim made at the announcement, like Salesforce's: a vendor reporting on its own desk.
How fast can an autonomous agent answer
Three published speed claims show the problem. Salesforce's AI for Customer Service & Support page shows a "65% reduction in response time for 90% of users." CIO writes that ServiceNow claims the agent is "almost twice as fast as human agents in performing these tasks". CX Today, on the same announcement, reports that early data puts it at resolving cases "99% faster than human agents." None of the three says what was timed or over which period, and the two ServiceNow figures compare with "human agents" without saying which agents, which hours or which measure.
"Almost twice as fast" and "99% faster" could describe the same result (1.99 times the speed) or very different ones (half the time against a hundredth of it). Neither report says which.
Track four numbers on your own desk:
- First response time: from the customer's message to the first AI reply, by channel.
- Time to resolution: from the first message until the conversation counts as resolved under your definition. If the definition includes a quiet period, the clock includes it.
- Time to a person: from the handoff to the first human reply. This is the clock a customer feels when the agent can't help.
- The human baseline for the same channels and hours, taken before launch, as medians and 90th percentiles, because averages hide the slow tail.
Without that baseline, a speed claim has nothing to be faster than.
A first 90 days, as a plan
The sources give one hard anchor, a four-week pilot on a defined group. The rest of the table is a plan built from what Salesforce says it learned, and the week numbers are a frame to adjust. For a ticket-level test before you commit, see how to run a pilot on your own tickets in our comparison of leading AI agents. For connecting the agent to the systems you already run, see integrating AI into your workflows.
| Stage | What to do | What to watch |
|---|---|---|
| Before launch | Clean the content the agent will read: fix or remove outdated and overlapping articles. Write the escalation rules. Define "resolved" and record the human baseline for response and resolution time | Topics that have more than one current article; the baseline, written down |
| Weeks 1 to 4, pilot | Open the agent to a defined group, as Salesforce did with 200 authenticated users. Start with high-volume, low-risk questions. Read every conversation | Handoff rate and the reason for each handoff; wrong or off-topic answers |
| Weeks 5 to 6, fix | Fix what the transcripts show: instructions that overcorrect, missing documents, product names mistaken for requests for a person | Repeat errors; conversations where the agent refused a topic it should answer |
| Weeks 7 to 10, widen | Add users, channels or languages one step at a time. Hand over fast whenever a customer asks for a person | Resolution rate under one fixed definition, with weekly volume, by channel and language |
| Weeks 11 to 12, routine | Fix a test set of real questions, a written definition of a good answer and one owner for answer quality, the way Salesforce set up its Evaluation Manager role | Answer quality on the same questions each run, next to resolution and escalation rates |
| Day 90, decide | Report the figure with its definition, period and volume. Decide what to widen, what to keep with people and what to take back from the agent | Customers who got a "resolved" answer and came back about the same issue; share of conversations that ended in a handoff |
The pillars of AI in customer support give the wider frame, and AI resolution rate benchmarks by industry show what a good rate looks like in yours.
How CoSupport AI Agent counts and bills a resolution
CoSupport AI Agent is an autonomous AI agent for customer support. A conversation counts as resolved when the customer hasn't replied within 72 hours after the AI's answer, and a customer who writes back later opens a new ticket. In the terms used above, that's the second rule: a quiet period after the answer. You pay $0.50 per resolved AI conversation, and conversations that go to a person are free. The first 1,000 AI responses are free, there's a 30-day pilot, and the agent answers in 120+ languages. CoSupport is ISO 27001 certified and GDPR and CCPA compliant.
Shelterluv, a software platform for animal shelters, runs CoSupport AI Agent on Zendesk Suite. Its customer page shows chats resolved by AI going from 69% to 71% to 73% over three months, and 61% of emails resolved by AI by month three. The Zendesk Suite integration took one month.




