Blog

Definitions and measurement

Resolution vs deflection vs containment: what each one counts

Three words, three different counts. On vendors' own dashboards, "resolution" often means the second one with a timer attached, and this year Zendesk and Intercom both changed how their headline rate is worked out.

The short version: deflection counts customers who never reached a person, containment counts AI conversations that ended without a hand-off, and resolution is supposed to count problems that were actually solved. The three meet at silence. A customer who reads the AI's answer and leaves without replying is deflected and contained; whether they are "resolved" depends on the vendor. As read on 28 October 2026, Intercom's Fin counts that customer as resolved after 24 hours, and HubSpot's customer agent and Gorgias after 72, and all three bill for it. Intercom's own buyer's guide says containment "cannot distinguish between a customer who left satisfied and one who gave up in frustration"; Fin's billing FAQ says a frustrated customer who just leaves "will be charged". Zendesk calls the same conversation a "Contained resolution" and does not bill it unless an LLM check of the transcript, run when the conversation ends, confirms the request was resolved. Salesforce's agent analytics files a session that times out as "Abandoned".

The headline rates also moved this year without a single conversation changing. On 18 May 2026 Zendesk began counting contained conversations in its automated resolution rate and told customers they "may notice an increase". Between 1 and 8 July 2026 Intercom took "Fin Constrained" conversations out of the denominator of Fin's Resolution Rate; its own worked example goes from 33% to 50%. Both vendors announced the change in their help centres, and neither change touched what is billed. A resolution rate compared across those dates may be comparing two definitions.

Deflection, containment and resolution: what each one counts

The three terms as they are generally used. Vendors vary, which is the subject of the rest of this piece.
Term What it counts Usual formula What it cannot tell you
Deflection Customers who did not reach a person. Some teams count contacts that never became a conversation, such as a help article read instead of a ticket opened Self-service sessions with no ticket, chat or call on the same issue within a set window, over all self-service sessions Whether the customer got what they needed, gave up, or tried another channel
Containment AI conversations that ended without a hand-off to a person AI conversations with no hand-off, over all AI conversations The same: a satisfied customer and one who gave up look identical
Resolution Conversations where the customer's problem was solved Conversations solved, over AI conversations or over all conversations Nothing, if it means that. In practice it means whatever the vendor's rule says, and the rule usually includes silence

Each word is meant to be stricter than the one before. On vendor dashboards the first two are often the same count. HubSpot's "Deflections" are "conversations handled by the customer agent only, without any involvement from a human agent", and the Deflection Rate in Salesforce's Agentforce analytics is the "share of sessions that ended without escalation". Both are containment by the usual definition. The difference that matters to a buyer is between those two and the third.

Measured on Fin's own scale, an assumed resolution sits at containment

Intercom has published one of the clearest explanations of the difference. Its guide "Resolution Rate vs. Deflection Rate" on fin.ai, dated 10 March 2026, puts four measures on a scale from least to most rigorous: deflection, containment, automated resolution, and resolution with experience measurement. It says containment "cannot distinguish between a customer who left satisfied and one who gave up in frustration". It tells buyers to ask every vendor whether its resolution rate includes deflections, and how many "resolved" conversations come back within a day.

Fin's billing rule sits lower on that scale. Intercom's help article "Fin AI Agent outcomes" counts a resolution when, after Fin's last answer, the customer either confirms it helped or "exits the conversation without requesting further assistance". The second case is an assumed resolution, and fin.ai's article "Fin pricing: Outcomes" sets the wait: a customer who disengages for 24 hours after Fin's last answer. Asked whether a frustrated customer who just leaves is billed, the help article's FAQ says it "is counted as an Assumed Resolution and will be charged". Measured on the guide's scale, the silent case sits with containment, not automated resolution: Fin must have given an actual answer, but nothing checks whether it was right, and nothing in the rule tells the satisfied customer from the one who gave up.

This is a point about a definition, not about Intercom's honesty: everything in it comes from Intercom's own pages, and the rule has protections around it. A resolution needs "an actual answer", so a reply to a greeting does not count. A confirmed resolution is recorded separately from an assumed one. If Fin detects frustration and escalates, nothing is billed, and the FAQ suggests escalation rules to make that more likely. A clarifying question that nobody answers is not billed. And a resolution is deducted if the customer returns to the same conversation for more help, "even across billing periods" (a new conversation about the same problem is not covered, as our piece on reopened conversations sets out). What the guide's scale shows is that one word on the invoice covers two of its rungs. We set out how long each vendor waits before silence counts in our piece on assumed resolutions.

What five vendors call each count, and which one they bill (October 2026)

Each row is from the vendor's own pages, read in a browser on 28 October 2026. Prices for each billed unit are in the AI support billing index.

The words on five AI support dashboards, and what each vendor bills. From their own pages, read 28 October 2026.
Vendor The words on its dashboard A customer reads the answer and leaves What is billed
FinIntercom Resolution Rate, Involvement Rate and Automation Rate An assumed resolution, after 24 hours Outcomes, assumed resolutions included
AI agentsZendesk Assisted escalation, Contained resolution and Verified resolution; one automated resolution rate covering the last two When the conversation ends (by default 2 hours after the first message on messaging, 72 hours after the first email), a Contained resolution unless an LLM check of the transcript confirms the request was resolved Verified resolutions only
Customer agentHubSpot Deflections and Resolutions, reported separately A resolution 72 hours after the customer's last message, if the agent shared a source or took an action Resolutions, in HubSpot credits
AI AgentGorgias Automated interactions and automation rate An automated interaction, once 72 hours pass without a human handover Automated interactions
AgentforceSalesforce Deflection, Escalation, Abandonment and Task Resolution rates Abandoned, if the session times out (2 hours by default) Help Agent: resolutions. No charge if the customer "walks away unhappy"; silence otherwise not stated

HubSpot says the important thing plainly. Its help article on analysing the customer agent reports deflections and resolutions as two numbers and warns that "Deflections do not always indicate resolution", giving the example of a customer who left because the agent could not answer. Its resolution still includes silence: the status is set 72 hours after the customer's last message, and HubSpot says later negative feedback does not change it.

Zendesk is the only one of the five whose pages say silence alone is not billed. A conversation where the customer asked for nothing more counts against the allowance only if an LLM's reading of the transcript confirms the request was satisfactorily resolved, which makes it a Verified resolution; if not, it is a Contained resolution and costs nothing. That check is an LLM inside Zendesk's own product, so the vendor grades its own agent, the kind of self-grading Fin's guide warns about. But it is a check, and a conversation that fails it costs nothing.

Gorgias's automation rate has the widest denominator. Its metrics glossary divides billable automated interactions by the "total billable workload": every automated interaction plus every ticket a person handled. Because every ticket is in the denominator, the rate cannot rise by narrowing which conversations count. Its AI Agent report also shows a Success rate, which divides by only the tickets AI Agent handled, handovers included, so ask which of the two you are being quoted. The automation rate's numerator is the billed unit itself, so the rate and the invoice describe the same thing.

Salesforce gives abandonment its own name. Its Trailhead module on Agentforce analytics lists four session outcomes: deflected ("Ended at user's request (user resolved)"), escalated, abandoned ("Timed out or dropped") and ambiguous, plus a separate Task Resolution Rate scored by a detector in its Trust Layer. That separates a session that times out from one the customer ended, though a customer who closes the chat in frustration still ends "at user's request", which the module labels "user resolved". The module does not say whether abandoned sessions also count towards the Deflection Rate, which it defines as sessions that "ended without escalation".

  1. The customer reads the AI's answer and leaves without replying
  2. Is the vendor Zendesk, and did its LLM check of the transcript fail to confirm the request was resolved? Yes Contained, not billed
  3. Is it a Salesforce Agentforce session that timed out? Yes Abandoned
  4. Fin, HubSpot, Gorgias and Zendesk: counted as resolved, and billed, once the wait passes Fin: 24 hours after its last answer. HubSpot: 72 hours after the customer's last message, if the agent shared a source or took an action. Gorgias: 72 hours without a human handover. Zendesk: only when the check passed, as a Verified resolution. Salesforce: a session the customer ended counts as deflected, and how Help Agent bills it is not stated.
One silent conversation at five vendors, per their own pages, read 28 October 2026. Salesforce does not state how Help Agent bills a session that times out. Fin deducts a resolution if the customer returns to the same conversation for more help.

Two 2026 changes that raised the headline rate and left the bill alone

Zendesk, 18 May 2026. Its announcement "Announcing changes to AI agent reporting" gives the formula before and after. Before, the automated resolution rate was Verified divided by all conversations (unassisted, assisted escalations, contained and verified). After, it is Contained plus Verified, over the same denominator. Zendesk says customers "may notice an increase" in the rate, that they "will continue to be billed only for Verified resolutions", and that history is re-mapped up to two years back. So the headline rate now includes conversations Zendesk does not bill because it could not confirm them. Zendesk records the tier in a Resolution tier field on the ticket, so a buyer can still separate the two.

Intercom, July 2026. fin.ai's article "Update to Fin performance metrics" says conversations where Fin was active but "didn't have the opportunity to answer" are no longer counted as Fin Involved. Its reasons include escalation guidance, a workflow that runs first, usage limits, an idle customer, and a teammate stepping in. Fin's Resolution Rate is resolutions divided by Fin Involved, so it rises. The article's own example: 250 resolutions out of 750 conversations was 33%, and out of the 500 that remain it is 50%. Existing workspaces moved between 1 and 8 July 2026. Data under the new definitions is available back to 1 October 2025, historical trends are not re-generated, existing custom reports keep the old definitions, and the article says "Billing is not affected."

Intercom's reasoning is defensible: a conversation Fin was never allowed to answer says little about Fin. And it kept a stable number, the Automation Rate (resolutions over all conversations), which did not change. But its March guide tells buyers to "always ask what types of queries are included in the resolution rate denominator", and that is now the question to ask of Fin's own rate before and after July.

How to tell which one a vendor is quoting

  1. Ask for the word and the formula. Get the numerator and the denominator in words, and the date the formula took effect.
  2. Ask for counts, not a percentage. A rate without its two numbers cannot be checked or compared.
  3. Ask what happens when the customer leaves without replying. Is it counted as resolved, after how long, and is it billed?
  4. Work out a rate over every conversation. Divide the AI's resolutions by all conversations in the period, including those the AI never answered. Fin's Automation Rate and Gorgias's automation rate are already built this way.
  5. Read forty of the resolved conversations at random. Mark each resolved, not resolved or ambiguous, and count ambiguous as not resolved. What passes is resolution; the rest was containment.
  6. Compare the result with the bill. The billing index sets out what each vendor's billed unit includes.

Where Varley stands

Varley sells a competing AI support agent, so read this section as a vendor talking about itself. Everything above it can be checked on the vendors' own pages.

Varley's definition page, How we count a resolution, says a resolution is counted "only where there is positive evidence the customer's issue was dealt with": the customer said so, or Varley completed an action that settles the request and can be checked afterwards. It says "Silence never counts, however long it runs and however warm the last message was." On rates, it says the denominator "is every conversation Varley was given, including the ones it handed off and the ones it got wrong". Varley does not publish a resolution rate yet; the page says it does not have the volume for one to mean anything. The early-access price for direct use is $0.50 per resolution, as of 28 September 2026 on the pricing page (read 28 October 2026).

Get early access to Varley.

Right of reply

None of the companies named here was approached before publication. Every claim is taken from their own public pages, read on the date given in the sources. If anything on this page is wrong or has changed, write to support@dragapp.com and the correction will be published here with a date, including if it is inconvenient for us.

Questions

What is the difference between containment rate and deflection rate?

Containment is measured inside the AI conversation: of the conversations the AI handled, the share that ended without a person. Deflection is wider and less consistent. Teams that measure it outside the conversation usually count a help-centre visit as deflected only if no ticket, chat or call on the same issue follows within a set window, such as a few days.

Is a high containment rate a good sign?

Not on its own. A contained conversation ended without a person, and that includes customers who gave up, switched to email or phone, or will write again tomorrow. Read it next to a repeat-contact rate and a sample of transcripts. Salesforce reports abandoned sessions as a separate outcome, which makes the customers who timed out easier to see.

Do AI support vendors bill for deflections?

None of the five vendors compared here bills a unit called a deflection. They bill a resolution, an outcome or an automated interaction, and two of those units reach past resolution. Fin's outcomes include a procedure hand-off, where Fin follows steps you configured and passes the conversation to a person. Gorgias's automated interactions include its flows and order management features, not only its AI Agent.

Why does a Fin custom report show a different resolution rate from the dashboard?

Because the two can use different definitions. Since July 2026 Fin's Support Performance dashboard leaves Fin Constrained conversations out of the Resolution Rate denominator, but Intercom says existing custom reports keep the previous definitions and only new report templates use the updated ones. Check which definition a report uses before comparing figures, or compare Automation Rates, which did not change.

Changes to this piece

28 Oct 2026
First published.

Sources