- Intercom counts a resolution when the customer stops writing for 24 hours after Fin's last answer, and bills $0.99 whether or not anyone said it helped.
- Zendesk waits up to 72 hours on email and web forms, then asks a second model to confirm the reply was relevant before it charges $2.00.
- Salesforce sells the same work three ways, at $2.00 per conversation or $0.10 per standard action, and the two models cannot run in one org.
- Handing a ticket to a human costs nothing at either Intercom or Zendesk, so the outcome the marketing calls a failure is the cheapest one you can get.
- A resolution rate is a billing term before it is a quality measure, and the two are not the same number.
Three of the companies that sell software to answer your customers announced agents in the same week of September 2026. Zendesk put out a set of specialized agents built for specific sectors, including commerce, with a no code builder for the ones that do not fit a template and a claim that they automate up to 80 percent of workflows. Salesforce shipped a support console built around the same idea. OpenAI published a customer story about Ringg, whose voice agents it says resolve up to 65 percent of customer calls.
Every one of those numbers is a percentage of something. The interesting question for anyone running a shop is what sits in the denominator, and what event the vendor treats as the moment a customer was helped. Both answers are written down. They are in the billing documentation rather than the marketing, which is where they belong, because a resolution is first of all a thing you pay for.
What does a resolution rate actually count?
It counts conversations that ended without a human stepping in. It does not count conversations where the customer got what they needed, because no system can observe that directly, so every vendor picks a proxy and the proxy is silence.
Intercom is the clearest about this. Its documentation splits the outcome in two. A confirmed resolution is the customer saying something like "that worked". An assumed resolution, per the breakdown of Fin's per outcome pricing, is what happens when "the customer disengages for 24 hours after Fin's last answer without asking for more help". Both are billed at $0.99. The company is not hiding it. It simply prices the two identically, which tells you how it treats them.
Zendesk runs a longer clock and adds a check. On email and web forms an automated resolution is counted after 72 hours of inactivity, and the four conditions are spelled out: the agent gave a generative reply, the customer gave positive feedback or none at all, no human agent responded, and a separate large language model reviewed the transcript and confirmed the reply was relevant. On messaging the conversation is evaluated two hours after the first message by default, configurable up to the same 72 hour maximum.
That second model is a real difference and worth crediting, though verification still only checks that the reply stayed on topic, not that it was true, which is a separate problem covered in why a shop chatbot invents products it does not sell. It means a reply that wandered off topic can fail verification and cost you nothing. It does not mean the customer was satisfied. It means one machine read another machine's answer and thought it was on point.
What do the three billing units cost side by side?
Nobody publishes this comparison, because each vendor has a reason to describe its own unit as the natural one. Assembled from the current pricing pages, it looks like this.
| Vendor | What you are billed for | Price | What silence means | Window |
|---|---|---|---|---|
| Intercom Fin | Resolution, confirmed or assumed | $0.99 | A billable resolution | 24 hours of no reply |
| Zendesk AI agents | Automated resolution, verified by a second model | $2.00, or $1.50 on a committed rate | A resolution if verification passes | Up to 72 hours on email and web forms |
| Salesforce Agentforce | Conversation | $2.00 | Not the billing trigger | Windows vary by user type |
| Salesforce Agentforce | Standard action, on Flex Credits | $0.10, at 20 credits | Not the billing trigger | Per action |
The Salesforce rate card is worth reading closely if you are anywhere near it, because it is the one that varies most by what the agent is doing. The published breakdown puts Flex Credits at $500 per 100,000 credits, a standard action at 20 credits and a voice action at 30, an advanced prompt at 16 against 4 for a standard one, and a Help Agent resolution at 400 credits, which lands at the same $2.00. Unused credits do not roll over into the next term. Flex Credits and Conversations cannot run in the same org, so picking one is a commitment rather than a setting.
Why does going quiet count as success?
Because the alternative is asking, and asking costs more than it returns. A survey prompt at the end of every conversation depresses the completion rate and annoys the people who were already happy, so the industry settled on a default that reads absence as agreement.
The trouble is that two very different customers behave identically. One reads the answer, solves the problem and closes the tab. The other reads the answer, decides the thing is useless, and goes to your competitor, your Instagram messages or your phone. Neither of them writes back. Both of them land in the same bucket, and on a shop's bill they cost the same.
This is not a scandal and it is not a reason to avoid the tools. It is a reason to stop treating the vendor's headline percentage as a measure of customer satisfaction. It is a measure of conversations that did not escalate, which is a genuinely useful operational number and a different thing entirely.
Intercom deducts a resolution if a conversation marked resolved is reopened by the customer seeking more help, even across billing periods. Zendesk verifies once, at the end of the window. So the two rates are not directly comparable: one keeps correcting itself after the fact, the other closes the books.
What does this do to the number a salesperson quotes you?
Take a shop handling 400 support conversations a month, which is a busy one person operation or a small team during a peak. Suppose the agent handles 70 percent of them without escalation. That is 280 billable resolutions.
At Intercom's $0.99 that is $277.20 in outcomes, on top of the $49 base plan which includes the first 50. At Zendesk's standard $2.00 the same volume is $560, or $420 on the committed rate of $1.50. At Salesforce's $2.00 per conversation the arithmetic depends on whether an escalated conversation still counts as a conversation, which is exactly the kind of question the pricing page leaves to your account team.
Now suppose a quarter of those 280 were the second kind of silence, the customer who gave up. You paid for 280 and you were helped on 210. Your real cost per solved problem moved from $0.99 to about $1.32, and, more expensively, seventy people had a bad experience that never appeared in a report. The bill looked identical either way. That is the whole reason to measure this yourself rather than reading it off a dashboard.
The same arithmetic is why we keep coming back to the gap between what owners actually pay for AI tools and the headline price. Usage based pricing is honest, but it prices an event, and you have to agree with the vendor about what the event is.
Which tickets can an agent genuinely close?
The ones where the answer already exists in writing and the customer only needs it found. Order status, delivery windows, sizing, ingredients, opening hours, the returns policy, whether you ship to a country. These are lookups, and a machine that can read your help centre and your order table does them well.
The ones it closes badly are the ones where the answer depends on judgement you have not written down. Whether this particular late delivery gets a refund or a credit. Whether you make an exception for a customer who has bought from you for six years. Whether the damage in the photo is shipping damage or wear. Those are decisions, and a support agent that makes them for you is making policy on your behalf, which is a different risk from answering a question. We went through where that line sits when deciding what part of a returns flow is safe to automate.
There is a middle category that gets underrated: the conversation where the agent does the lookup and drafts the reply, and you press send. It never appears in a resolution rate, because a human touched it. It is also free at both Intercom and Zendesk, where an escalation with no outcome is not billed and assisted escalation carries no charge. For a small shop that reads every message anyway, that middle path can be the best value in the product, and it is the one nobody sells you.
What should you check before switching one on?
Five things, all answerable before you sign, none of them on the pricing page.
- What ends a conversation, and after how long. Get the silence window in writing. Twenty four hours and seventy two hours produce very different monthly bills on the same traffic.
- What happens when a customer comes back. Ask whether a reopened conversation refunds the resolution, and whether that holds across billing periods. Intercom says it does. Not every vendor does.
- What an escalation costs. If handing the ticket to you is free, the safest configuration is also the cheapest, and you should set the agent's confidence threshold accordingly rather than chasing the automation rate.
- Which tickets it is allowed to close alone. Write the list yourself. Anything touching money, health, safety or a legal promise belongs on the human side on day one, and can move later if the logs earn it.
- What you can see afterwards. You want the transcript of every conversation it closed, not a percentage. A weekly read of twenty of them will tell you more than any dashboard, and it is how you find the customers who left quietly.
How do you measure the rate yourself?
Count reopens and follow ups, not closures. A resolution that stays closed for a fortnight is worth something; one that generates a phone call two days later was not a resolution, it was a delay. Most help desks will show you conversations reopened within seven days as a filter, and that number is the honest denominator.
Then add the channel your agent cannot see. When a shop's chat agent starts answering badly, the traffic does not vanish, it moves. It turns up in Instagram DMs, in WhatsApp, in the phone. If your ticket count dropped 30 percent and your call volume rose, your resolution rate is measuring migration. This is the same measurement trap we wrote about when four shops measured their AI traffic and got four different answers, and it has the same fix: pick the metric that survives the channel moving.
How many resolutions does your plan already include?
More than most people check, and the allowance is per seat rather than per account, which catches out small teams. On Zendesk, the published allowances run at 5 automated resolutions per agent per month on Suite Team, 10 on Suite Professional and 15 on Suite Enterprise. Everything past that is the $2.00 standard rate, or $1.50 if you commit to a volume up front.
Read that as a one person shop and the number is stark: a single Suite Team seat carries 5 included resolutions a month. If your agent handles 280, you are buying 275 of them at the going rate. The included allowance is a sample, not a budget. Intercom is structured differently, with a $49 monthly base that carries 50 outcomes and a floor at 50, so a quiet month still costs $49 and a busy one scales linearly from there.
The practical consequence is that these products get cheaper per ticket as you grow and are proportionally most expensive for the smallest shops, which is the opposite of how the marketing is pitched. Work out your own cost per solved problem at your real volume before you compare it to an hour of your time.
Does any of this change for voice?
The units change and the stakes rise. OpenAI published a customer story this week about Ringg, whose multilingual agents it says resolve up to 65 percent of customer calls across voice, chat and web. On the Salesforce rate card a voice action costs 30 Flex Credits against 20 for a standard one, so the same work priced through a phone line runs at $0.15 rather than $0.10.
Voice also removes the silence signal that the text products lean on. A caller who hangs up has not gone quiet for 24 hours, they have ended the conversation in a single gesture, and a hang up can mean satisfaction or fury. If you are looking at a phone agent, the number to ask for is the callback rate within 48 hours, not the resolution rate, and we went through the rest of that decision in a piece on what an AI phone answering service can and cannot handle.
Is any of this a reason not to use one?
No. A support agent that closes order status questions at one in the morning is straightforwardly good for a business run by one or two people, and the pricing has moved in the buyer's favour over the past year. Zendesk's May 2026 changes stopped charging for assisted escalations and for contained conversations that failed verification, which is a vendor giving up revenue on the outcomes it cannot stand behind. That is the right direction.
What changed this week is only the volume of claims. When three vendors announce in the same week and each quotes a different percentage, the percentages are not comparable, because the units underneath them are not. Zendesk's 80 percent of workflows, a 65 percent of calls figure on a voice product, and a resolution rate on your own help desk are three different measurements wearing the same word.
If you are picking one of these tools for a commerce site, the integration surface matters as much as the rate: whether it can read your order table, your stock, your shipping rules and your returns policy without you pasting them in. That is a question about how your storefront is built, which is why we keep the MaShop plans and credit pricing in plain arithmetic rather than per outcome. Whatever you choose, ask the vendor what ends a conversation. If the answer is that the customer stopped writing, you now know what you are buying.
And keep reading the transcripts. The number on the dashboard is the vendor's best guess at your customers' state of mind, inferred from an absence. Twenty transcripts a week is the cheapest quality process in your business, and it is the only one that can hear the people who did not write back. If you want the longer version of what a bot can promise on your behalf, we covered what happens when an assistant commits your business to something.