What counts as an autonomous resolution? How nine agentic AI customer service vendors bill it

TL;DR

• No two agentic AI customer service vendors count an autonomous resolution the same way. They split on silence, handoffs, and customers who come back.
• The rule only sets your invoice if you pay per resolution. On session, conversation, or action billing, “resolved” is a dashboard word, so get the rule in writing either way.
• Call it the return test: of the nine rules we read, only Fin takes the charge back when a customer returns to the same conversation. Until a second vendor does, no two resolution rates measure the same thing.

Nobody had to define an autonomous resolution until somebody started billing for one. Gartner predicts that by 2029, agentic AI will resolve 80% of common customer service issues with no human help, but it doesn’t say who decides what “resolve” means. Right now, the vendor does, and if it bills per resolution, its definition is your invoice. We print no vendor’s resolution rate here, ours included: we quote nine vendors’ counting rules as of October 2026, ordered by who pays when the AI is wrong.

Which billing rule protects your budget?

Pick the rule before you pick the vendor. Here’s where each one fits, and the one thing to get in writing before you sign.

What is an autonomous resolution?

An autonomous resolution is a customer or employee request that an AI agent closes on its own, with no person taking over. In agentic AI customer service, the AI acts as well as answers, and each vendor writes its own rule for what counts, so one conversation can be resolved at one vendor and not at another.

Nine agentic AI customer service vendors, nine ways to count a resolution

This is desk research: we read each vendor’s own pages on September 21, 2026, and re-read the counting rules and prices on October 13, 2026. “Billed” means the rule sets the invoice, and “reported” means it only moves the dashboard.

Does the counting rule touch your invoice?

Start with the unit on your contract, because it decides how much of this article is about your money. If you pay per resolution, the counting rule is your invoice: Fin charges $0.99 per outcome, Salesforce $2 per Help Agent resolution, and Atlassian $1 per customer-side resolution from December 3, 2026. Zendesk’s tier platform draws verified resolutions from a dollar allowance per agent seat.

If you pay per conversation, session, or action, the rule only moves the dashboard. Ada bills per conversation by default, Freshworks sells sessions at $49 per 100, we meter 20 credits per support conversation, and ServiceNow counts assists per workflow. On those contracts, treat any resolution rate as a claim until you’ve seen the rule behind it.

The rate still matters to anyone who reports upward. In a Gartner survey of 1,303 senior leaders, published July 8, 2026, service and support leaders put a median 12% of their 2025 budget into AI, the most of the 10 business functions surveyed, yet only 24% of those service and support leaders showed positive financial returns across their AI use cases. When the dashboard runs on one rule and the invoice on another, that return gets harder to prove in front of a CFO.

An employee can’t churn, so silence proves nothing

Silence means something different at work. A customer who goes quiet may have left. An employee can’t, so staff “just text the technician directly,” one IT manager on Reddit warns. Atlassian’s IT agent still records that chat as resolved after a silent day.

Vendors wrote most counting rules for customers, and the IT-side rules copied them without asking what an employee’s silence means. Atlassian’s flow for its virtual service agent starts after five minutes without activity. If the AI answered anything or matched an intent and the employee then stays silent for another 24 hours, Atlassian records the chat as resolved. Freshservice counts a session whenever an employee talks to its agent in a 24-hour window, whatever happens next.

Two vendors’ own guidance points to a better test. Decagon’s glossary sets three tests for a confirmed resolution: the customer “did not come back about the same issue within a defined window, did not open a ticket through another channel, and did not escalate.” Ada’s September 2026 buyer guide warns that loose definitions risk “double-paying when customers re-engage through another channel.” Inside a company, the other channel is the technician’s phone.

Of the rules we read, Fin’s refund comes closest to that test, and it only covers a return to the same conversation. If your agent answers employees, ask the vendor how its count would ever learn about the text to the technician.

Amar Gautam, Head of DevRev Labs, makes the same point about ticket metrics in our post on employee self-service beyond the ticket: “None of them measure whether the employee’s problem got solved.”

How we read each vendor’s rules, and what we couldn’t check

Everything here comes from vendors’ own pages. We read each vendor’s pricing pages, help-center articles, product docs, and blog posts, plus a legal PDF for ServiceNow, and noted the page and the date behind every rule. We didn’t run live conversations or audit anyone’s invoice. The Reddit and G2 quotes are single users’ experiences, linked and dated. Where no Reddit or G2 conversation exists about a product, we left those blank.

Some doors stayed shut. Decagon’s docs need a login, so we read its public pages only. Five of the nine don’t publish a price for the unit they count: Zendesk’s resolution tiers, Decagon, Ada, ServiceNow, and our own credits. Where we didn’t find a rule or a price on a vendor’s public pages, the tables say “Not published,” so ask that vendor directly.

Four vendors didn’t make the list:

• Sierra, which helps businesses build customer experiences with AI, agrees on “clear, agreed-upon criteria for each outcome upfront” in each contract, so there’s no public rule to quote.
• Moveworks, an AI assistant platform that has been part of ServiceNow since December 15, 2025, and Forethought, an AI agent platform that has been part of Zendesk since March 26, 2026, sit with their owners here rather than as separate entries.
• Gorgias, an ecommerce help desk, sits outside this comparison of agentic AI customer service vendors.

Three moments split the nine rules: silence, a handoff, and a customer who comes back

Handoffs aren’t an edge case. In a Gartner survey of 3,566 customers (August 4, 2026), 87% said a company that uses generative AI for service must let customers reach a person. Here’s how each vendor handles all three moments:

Read down the returning-customer column: only one cell is a refund.

The nine vendors, and who pays when the AI is wrong

After our own entry, the order runs by how much proof a vendor needs before a closed or handed-off chat bills: from Zendesk, which bills only what a model verifies, to ServiceNow, where a handoff to a person adds assists. Fin sits eighth because it bills silence, even though it’s the only one that refunds a return.

It’s one axis, not a quality ranking, and the bottom of it can still be your best buy. Each entry closes with an AI nativeness check: is there a version without the AI, and does the AI act?

1. Computer, by DevRev – bills the conversation, audits every action

Our Customer Agent post promises “Pay for results, not seats.” The unit is a support conversation the agent handles “on its own, start to finish.” Our pricing page meters it at 20 credits for one AI agent session of up to 24 hours, and every action the agent takes in it lands in the audit trail. Ask us in writing how a session handed to a person bills, and how we count a customer who goes quiet.

Our docs publish a resolve rule for one mode: Computer resolves a conversation only once the customer marks the answer useful. That rule feeds reporting, not your bill. A resolved conversation reopens if the customer writes again, and the report updates with it.

What we add is what happens after the chat. To us, the clearest false “resolved” is a conversation that turns out to be a bug, however polite the customer was. Support teams attach engineering issues to tickets in DevRev, and our analytics report the share of tickets linked to issues, so that conversation becomes engineering work we track.

Feature overview – Computer’s Customer Agent works across chat, email, WhatsApp, and voice. It ships with skills to search your knowledge base, create tickets, and hand off to a person, and actions such as order lookups and refunds are wired to your systems during setup. Each skill has a “Needs Approval” setting, off by default, that makes the agent ask for confirmation before it runs.

The bill grows with – Support conversations, on Pro or Max plans priced on request.

Verdict – Our overall pick. Buy it if your tickets often turn out to be bugs.

See every action Computer takes on a conversation, and how a ticket becomes engineering work. Book a demo →

Best for – Teams who want the support ticket and the engineering issue that fixes it in one record. Every action, trace, and trigger lands in an audit trail you can check.

Limitations

• No published price per credit.
• We document the resolve rule for one mode only, and no rule for silence or handoffs.
• Our customer stories don’t say what counted as resolved.

AI nativeness

No version without the AI. It acts through skills, and you choose which ones ask before they run.

This DevRev video shows how a support ticket links to the engineering issue that fixes it.

2. Zendesk – bills only what an LLM verifies

On its tier platform, live since May 18, 2026, Zendesk bills only resolutions a large language model (LLM) verifies. A quiet conversation bills when that check passes and lands in the free Contained tier when it doesn’t. The catch: its reported rate now counts Contained, too, and Zendesk warns “you may notice an increase in your AR%” (its automated resolution rate).

The tiers sort every AI conversation three ways:

• Assisted escalation.
• Contained resolution.
• Verified resolution.

Only Verified draws on your allowance, and only after an LLM reads the conversation and agrees the AI solved the request. Zendesk shares that model check with Atlassian’s customer agent.

Not every account is on the tiers yet. Customers tend to move at renewal, with 30 days’ notice, and the older platform keeps its own logic until then, so Zendesk can count two accounts on the same plan two different ways. On legacy autoreplies with articles, a customer who clicks “Yes, close my request” counts as a resolution, and so does a click on the article within 72 hours with no reply. That feature retires December 10, 2026.

Timing counts, too, because Zendesk judges nothing until a conversation ends. As of October 2026, a messaging conversation ends two hours after the first message by default, and up to 72 hours on request, while email ends 72 hours after the first email and voice at hangup. Zendesk shortened the messaging default from 72 hours in an edit dated September 10, 2026, and edited the page again on September 30, 2026, so check it on the day you sign.

Among the per-resolution billers, Zendesk is the kindest to silence, because a model has to confirm the resolution before you pay. Its weak spot is the reported rate, which can now climb in a month when the agent got no better.

Feature overview – Agents that act in connected systems and follow plain-language procedures. They work across messaging, email, and voice, in 80 languages.

The bill grows with – Verified resolutions, drawn from a dollar allowance per seat.

Verdict – Buy it if you run Zendesk and want a model to approve every resolution you pay for. Skip it if finance needs a unit price first, because Zendesk hasn’t published one.

Best for – Support teams that want free escalations and a model check before any silence reaches the bill.

Limitations

• Dashboard and invoice count different things, so the rate you report isn’t the rate you pay for.
• No refund when a customer comes back. Deleting the ticket “doesn’t undo the consumption,” and disputes go through Zendesk Customer Support.
• Tier prices aren’t published.

AI nativeness

Included in every Suite and Support plan on top of seats. You can switch it off, and it acts.

3. Atlassian – two products, two opposite rules

Atlassian runs two opposite rules. Customer chats get an LLM verdict after 24 hours of silence, “permanent once set.” The IT agent in Jira Service Management bills any chat matched to an intent, whether the AI resolves it or escalates it.

On the customer side, a resolution needs a complete answer, no handoff to a person, no work item, and a “resolved” verdict from Atlassian’s LLM judge. A chat where the agent says it can’t help doesn’t count, and neither does a test run from the settings. A customer who comes back starts a new conversation, and the first charge stands. If you think a verdict is wrong, contact Atlassian Support.

On the employee side, the unit is an assisted conversation: any chat matched to an intent, plus any chat the AI answered that the employee confirms or abandons. Premium and Enterprise plans include 1,000 a month before the $0.30 rate starts. Read together, the two products show why the word needs a contract, because the customer side runs one of the stricter rules here and the employee side one of the loosest.

Feature overview – Customer and employee agents that answer from your knowledge base. Customers reach them on web, chat, email, and voice, and employees use Slack, Microsoft Teams, and portals.

The bill grows with – $1 per customer resolution from December 3, 2026. Extra employee chats start at $0.30 each. Resolution packs bring the customer price down to $0.85 at 200,000 or more.

Verdict – Buy it if your service desk already runs on Jira. Skip it if the agent has to change records, because Atlassian’s customer agent can read your systems but can’t write to them.

Best for – Jira Service Management teams that want an LLM, not a timer, to judge customer chats, and can accept a looser rule on the employee side.

Limitations

• The customer agent is read-only.
• The IT agent bills escalations.
• Customer-side billing starts December 3, 2026, so there’s no invoice history to check yet.

AI nativeness

Metered on top of seat-priced plans, and not on the Free plan. The customer agent reads but can’t write.

4. Salesforce Agentforce – bills some resolutions after a handoff

The Help Agent bills $2 for a session with two exchanges each way and no negative final feedback. A handoff cancels it unless the final feedback is positive: then it “consumes a Resolution, even if the user requests escalation,” and the agent asks for feedback just before it hands off. Salesforce’s announcement says a handoff means “there is no charge.” Salesforce now owns Fin, too, so one company sells two definitions of the same word: Fin refunds a resolution when the customer returns to the same conversation, and Help Agent can bill one after a handoff.

Salesforce says it uses “deterministic billing criteria rather than semantic interpretation of conversation quality,” so message counts and feedback decide, not a model. A session that ends before two messages each way counts as abandoned and isn’t billed. With no feedback, a session that wasn’t escalated counts as resolved when its window ends, two hours by default for text. A chat resolution covers up to two hours, an email one up to three days, and a voice one up to 10 minutes.

Deterministic rules are easy to audit, which is their point. They also can’t tell a solved problem from a polite “goodbye.”

Feature overview – Answers and manages cases across voice, web, portals, SMS, and WhatsApp. It can also act on orders and schedules.

The bill grows with – Your pick. Pay $2 per resolution, with no charge for the actions inside it, or pay $0.10 per action, or $2 per conversation.

Verdict – Buy it if your team lives in Salesforce and wants billing rules it can audit with a calculator. Skip it if positive feedback shouldn’t turn a handoff into a charge.

Best for – Service teams that want abandoned sessions kept off the bill and a rule they can check message by message.

Limitations

• A 12-minute call counts as two resolutions, because a voice resolution covers 10 minutes.
• Salesforce bills silence after two exchanges.
• A customer who returns after the window starts a new session, with no refund published for the first.

AI nativeness

Metered add-ons on seat-priced Service Cloud. It acts through flows.

What Reddit says

One team checked Agentforce’s dashboard against a manual review of nearly 700 conversations:

“We were seeing a 30% deflection rate shown in the dashboard, which really impressed people.” After the review: “It was actually deflecting 2%.”

r/salesforce. April 2026 ↗

5. Decagon – resolutions on request, conversations by default

Decagon bills per resolution if you ask, charging “for each fully resolved conversation, with no charge for escalations.” It says most enterprises choose per-conversation pricing “to avoid negotiating what counts as a ‘resolution.’” Its pricing post asks the question every buyer should: “If a frustrated customer stops responding, does that count as ‘resolved’?” A paragraph later, it answers: “You’ll get into situations where a user is upset, leaves, and that gets counted as a resolution.”

Decagon measures its resolution rate as conversations resolved “without needing to escalate to a human,” and its own blog counts every chat that didn’t ask for a person as “resolved.” Its glossary sets a higher bar, calling an unverified rate “a containment question wearing a resolution label.” We didn’t find a published rule for a customer who goes quiet or comes back, so ask Decagon for both.

Feature overview – Plain-language procedures across chat, voice, and email. It acts through application programming interfaces (APIs) and the Model Context Protocol (MCP) on top of help desks such as Salesforce, Intercom, and Zendesk.

The bill grows with – Conversations, or resolutions at a higher rate.

Verdict – Buy it if you want to write the playbook yourself and don’t mind agreeing on what “resolved” means in the contract. Skip it if you need a price before the first sales call.

Best for – Teams on per-resolution pricing that want escalations kept free and will agree on the definition in the contract.

Limitations

• We didn’t find a published silence or return rule.
• No published price.
• Its docs need a login, so check its rules with Decagon directly.

AI nativeness

AI-only, on top of help desks, acting through APIs.

6. Ada – a four-part quality test, kept off most bills

Ada’s AI calls a chat resolved only if it was relevant, accurate, safe, and contained, and customer feedback “does not override the automatic classification.” Ada still bills per conversation in most cases, and its blog calls counting disengagement “abandonment with a different label.”

Silence doesn’t count on its own. An idle conversation ends after 24 hours in chat and 72 hours in email, and only then does the AI judge it. Ada marks an escalated conversation “Not Resolved” as soon as the handoff happens.

Feature overview – One reasoning engine across voice, email, and chat, with Playbooks for multi-step work.

The bill grows with – Conversations, unless you negotiate resolutions. Ada offers those “for enterprises with specific needs.”

Verdict – Buy it if you want a grader that customer feedback can’t override. Skip it if you want to pay only for outcomes, because Ada sells that only to some enterprises.

Best for – Quality and customer experience teams that report on accuracy as well as containment. At Ada, silence ends a chat without resolving it.

Limitations

• On per-conversation plans, the resolution test never touches the invoice.
• No published price.
• A customer who comes back after the timeout starts a new conversation, and the docs don’t say whether Ada relabels the first one.

AI nativeness

AI-first, though legacy block workflows are still in the product.

7. Freshworks – a session, whatever happens next

Freddy AI Agent uses a session at its first reply. That holds even for a welcome message, “regardless of whether the customer replies,” so the greeting is billable. On one measure, it’s the plainest meter on the page: the outcome never reaches the invoice, so there’s no counting rule to argue about.

Feature overview – Takes actions like refunds and hands off with context.

The bill grows with – Sessions, $49 per 100, after 500 free for new customers.

Verdict – Buy it if finance wants a bill it can forecast in January and still believe in December. Skip it if you want to pay for outcomes instead.

Best for – Freshdesk and Freshservice teams that budget AI by the session, at a published price.

Limitations

• Unanswered greetings cost sessions.
• A chat session covers only 24 hours from the start of the conversation.
• Freshservice’s price for extra sessions isn’t published.

AI nativeness

A metered add-on that acts in connected systems.

8. Fin (Salesforce) – refunds the resolution when a customer returns

Fin counts a resolution when the customer confirms the answer or “exits the conversation without requesting further assistance,” so a customer who gives up and closes the tab costs $0.99. In return, Fin is the only one of the nine that takes the charge back if that customer returns to the same conversation, “even across billing periods.” Salesforce completed its acquisition of Fin, the company once called Intercom, on September 10, 2026. Intercom is still the name of Fin’s help desk.

The unit is an outcome, and a resolution is only one kind. Fin charges at most once per conversation, “even if Fin takes multiple actions,” and it moved from per-resolution to per-outcome pricing between March 9 and April 17, 2026. Fin also bills two outcomes that only sales teams use: qualifying a prospect and ruling one out.

The line between an escalation and a handoff decides the bill. You pay nothing when Fin escalates on its own default rules or your workspace rules, including when it detects frustration. But if it completes a Procedure you set up to hand off to a person, that’s a billable outcome.

Fin treats 24 hours of silence after its last answer as a resolution. Say Fin answers 100 customers in a month: 70 confirm the answer, 20 leave without asking for more, and 10 ask for a person. You pay for 90 outcomes, because the 20 who left count, too, and nothing for the 10. If five of the 20 come back to the same conversation, those five charges come off.

Fin’s rule is the only one here that admits a resolution can turn out wrong after the fact. It still bets on silence, though: Fin bills a customer who gives up and never comes back, and the report calls that chat resolved. If few of your customers return to the same thread, few refunds will fire.

Feature overview – Procedures for multi-step work, and actions through APIs, Data Connectors, or MCP, in dozens of languages. It runs on Intercom or on help desks such as Zendesk and Salesforce.

The bill grows with – Outcomes, $0.99 each, including handoffs you built into a Procedure. Fin itself costs nothing in a month you don’t use it, and Intercom’s seats cost extra.

Verdict – Buy it if your customers come back when an answer fails, because Fin takes the charge back. Skip it if they vanish instead, because Fin bills the vanishing.

Best for – Support teams that want the vendor to carry the cost of a false resolution, on Intercom or on Zendesk or Salesforce with Fin added on top.

Limitations

• Its own pages disagree on silence: 24 hours or three minutes.
• A new conversation about the same problem gets no refund, even inside 24 hours. Intercom only reports it.
• A handoff you build into a Procedure bills the same $0.99 as a resolution.

AI nativeness

A real agent, sold as a metered add-on. The help desk runs without it, and Fin runs without Intercom’s help desk.

What G2 reviewers say

A small-business founder on G2 describes staff stepping in to rescue conversations where Fin was unhelpful:

“Fin still claims these as its resolutions and Intercom charges us for these despite us having to pay a human to resolve them properly.”

Alex B., Founder and CEO, small business. G2, June 2025 ↗

9. ServiceNow – a handoff to a person costs extra

An agentic workflow costs 25, 50, or 150 assists by size, and a live-agent handoff adds 10. ServiceNow’s analytics call a chat resolved when “at least one AI asset, for example, AI agent, skill, topic was executed,” so activity stands in for an outcome.

The schedule, effective August 24, 2026, sizes a workflow by its actions: fewer than four is small, five to eight medium, and nine to 20 large. A workflow ends on completion, at 20 actions, or after one hour of inactivity. If the assistant only answered and nothing else happened, the analytics file the chat under “Response Provided,” so a quiet chat counts as resolved only when an AI skill ran.

Feature overview – Agents act through flows, scripts, and MCP.

The bill grows with – Assists from a contract pool, entitled in 12-month periods.

Verdict – Buy it if ServiceNow already runs your IT. Skip it if you need a dollar price for a wrong answer before go-live, because the assist table only tells you how many assists.

Best for – Platform owners who want to model assist consumption from a public schedule before go-live.

Limitations

• A handoff costs more than no handoff.
• The size bands skip a four-action workflow, so ask how one is counted.
• No published dollar price for an assist.

AI nativeness

Built into every tier and metered in assists. It acts.

What G2 reviewers say

A mid-market customer on G2, who still rated ServiceNow AI Agents five stars, on the assist meter:

“It is very expensive. Each ‘assist’ costs money. The licensing costs are very prohibitive for us to expand our implementation and use the capabilities to their full abilities.”

Jessica S., mid-market company. G2, May 2026 ↗

Seven questions a vendor should answer before you sign

Take these seven questions into the next demo. At least one vendor on this page publishes an answer to each, so a vague reply is a choice the vendor made. Ada’s own buyer guide puts it plainly: “Before you sign a contract priced on automated resolution, get the definition in writing.”

1. Who decides a conversation counts as resolved? A timer, the customer, or a second AI model, like Zendesk’s LLM check.
2. How long does silence last before it counts, and does it count at all? Fin’s pages say 24 hours or three minutes; Ada ends the chat and judges it.
3. Does a handoff to a person cancel the charge? At Salesforce, positive feedback before a handoff still bills.
4. What happens when the customer comes back? Only Fin refunds a return to the same conversation.
5. Does the dashboard use the same rule as the invoice? Zendesk’s reported rate counts conversations it doesn’t bill.
6. What does one unit cost, in writing? Five of these nine don’t publish a price.
7. Can you audit the conversations behind the number? Ada warns that per-resolution contracts “can require ongoing auditing.”

The return test: what happens when a customer comes back

The rule for a customer who comes back is the least exciting one on the page, and the one to watch. Fin deducts the charge when a customer returns to the same conversation. Atlassian’s verdict is “permanent once set,” Zendesk says deleting a ticket “doesn’t undo the consumption,” and Salesforce starts a new session and publishes no refund. It’s also the part customers feel.

Jennifer Heape, Voice AI Product Lead at DevRev, said it plainly when we launched Voice AI in our July 2026 announcement: “Customers do not care which system holds the answer. They care whether they have to repeat themselves, wait for a transfer, or call back again.”

Until a second vendor passes the return test, every resolution rate is a score on a test its own vendor wrote. When a second one does, you’ll at last have two numbers worth comparing.

Bring your hardest resolution question to a demo. We’ll show you the audit trail behind every action Computer takes, and how a ticket becomes an engineering issue. Book a demo →