Chatbot Containment Rate: What It Measures, What It Hides, and the Number That Actually Matters

A vendor tells you their bot runs an 85 percent containment rate. Three weeks after launch, your inbox looks about the same as it did before. Both of those things can be true at once, and that is the entire problem with chatbot containment rate as a number.

Containment rate measures how many chat sessions ended inside the chat window without a human agent joining. That is all it measures. It does not know whether the customer got an answer. It does not know whether they closed the tab, opened their email, and typed the same question to you instead. A session that ends in a solved problem and a session that ends in a shrug produce the same row in the log.

This matters more in 2026 than it did two years ago, because the definition of a handled conversation stopped being a dashboard curiosity and turned into a line item. Intercom's Fin bills $0.99 per resolution. Zendesk bills $1.50 per automated resolution. Both define that resolution, in large part, as silence.

So the number on the sales deck, the number on your invoice, and the number of customers who actually got helped are three different figures. Only the first two are easy to see.

Below: what containment rate counts, a worked example with 100 chats so the gap shows up in plain arithmetic, what a realistic number looks like for a small Shopify store, how per-resolution pricing quietly profits from the difference, and the three numbers to ask for before you sign anything.

What is containment rate, exactly?

Containment rate is the percentage of chat sessions that ended without a human agent joining the conversation. The formula is as simple as it looks.

Containment rate = (sessions with no human agent involved / total sessions) x 100

If 100 people opened the chat widget last month and 80 of those conversations ended without you or anyone else stepping in, your containment rate is 80 percent.

Where the metric came from

Containment is not a chatbot metric by birth. It comes from phone support, where the question was whether a caller finished inside the automated phone menu or pressed zero to reach a person. Keeping the caller "contained" saved a real, measurable cost: one agent, several minutes, per call.

The metric moved to chat unchanged. On the phone, pressing zero is a loud, obvious escalation signal. In a chat window, the customer has a third option the phone menu never offered: leave. That third option is where the number starts lying.

Containment, deflection, and resolution are three different things

These get used interchangeably on vendor pages, and they should not be. Containment is session-level: this conversation ended without a human. Deflection is channel-level: a ticket that would have been created was not created. Resolution is outcome-level: the customer's actual problem got solved.

Only resolution says anything about the customer. I wrote a longer breakdown of how resolution rate and deflection rate differ, and containment sits in the same family: a measure of what your team did not have to do, dressed up as a measure of what the customer received.

One hundred chats: contained, resolved, abandoned

Arithmetic is more convincing than argument here, so take a month with exactly 100 chat sessions on a small Shopify store. Twenty of those conversations ended with a handoff: the customer asked for a human, or the bot escalated on its own. That leaves 80 sessions that ended inside the widget. Containment rate: 80 percent.

Now open those 80 and sort them by what actually happened.

Containment rate: 80 percent. Resolution rate: 40 percent. Same month, same bot, same log file. The two numbers are not in conflict, because they were never measuring the same thing.

Look at the middle three rows, because that is where a small store's real support problem lives. Forty conversations ended in the widget without the customer's problem being solved. Fourteen of those people went and emailed you anyway, which is why your inbox did not shrink. Sixteen just left, and you will never know what they wanted.

The ten no-question sessions deserve their own note. Greetings, curious browsers, someone poking the widget to see if it is real: none of these were ever support conversations, and all of them end without a human. They inflate containment for free. On a new store with low chat volume, this group can be a quarter of your sessions.

Why can 80 percent containment mean 40 percent resolution?

Three things open the gap, and all three are invisible on a containment dashboard.

1. Silence is ambiguous, and the metric reads it as a win

When a conversation stops, containment records one thing: no human joined. A satisfied customer and a frustrated one both stop typing. Unless something explicitly asks the customer whether the answer worked, the metric has to guess, and it guesses in the vendor's favor every time.

2. Customers switch channels instead of escalating

The escalation button is not the only exit. Your email address is on the contact page, and for most people it is the more reliable option. Gartner surveyed 3,566 customers in February and March 2026 and found that only 27 percent would try a chatbot again after one bad experience, while 87 percent said access to a human agent is essential when a company uses AI for service. People who have been burned once do not press the handoff button. They leave and find your email.

Every one of those customers is contained. Every one of them is also a ticket you still have to answer, plus a person who now trusts your store slightly less.

3. Questions the bot was never going to answer still close cleanly

A return outside your policy window, a damaged item, a custom order question, a billing dispute: these need a decision only you can make. If the bot cannot escalate reliably, the conversation ends without a human and gets counted as contained. Forrester's 2026 predictions put this bluntly, forecasting that a third of companies will harm customer experiences with frustrating AI self-service, largely by shipping bots into situations where they were always going to fail.

Containment cannot tell the difference between a bot that answered well and a bot with no exit door. The second one scores higher.

What is a good containment rate for a small Shopify store?

Honest answer first: containment rate is not a number worth setting a target for, because you can raise it by removing the handoff button. Set targets on resolution instead. But you will see containment quoted at you, so here is roughly what to expect by store type, and what the resolution figure underneath it usually looks like.

Notice the pattern: the harder the questions, the wider the gap between the two columns. A store selling one simple product to one country has a narrow gap. A made-to-order store has a canyon.

For context on the high end, Intercom publishes a 76 percent average resolution rate across its Fin customers, with top performers at 80 to 84 percent as of March 2026. That is a real number, but it is an average weighted by large software companies with big documentation libraries and dedicated support teams tuning the bot weekly. It is not the number a store doing 80 orders a month should plan around.

For a well-built chatbot on a small Shopify store, 40 to 70 percent genuine resolution is the honest range. If a vendor quotes you 85 percent and does not immediately explain how they define the word, that number is containment wearing a better jacket. Resolution is one of the six chatbot metrics worth watching; containment is not on that list for exactly this reason.

How per-resolution pricing profits from the gap

Until recently this was an argument about dashboards. In 2026 it is an argument about your bill, because the major platforms moved to charging per outcome, and they each had to write down what an outcome is.

Here is what they settled on.

Read those trigger conditions again. All three are definitions of silence. Fin's pricing page puts it as no further help being requested after its last answer. Zendesk uses a 72-hour quiet period. Gorgias documents a billable automated interaction as one where the customer does not require a human within 72 hours.

Now put your frustrated customer through that. She asks about a return, gets a policy answer that does not cover her situation, closes the tab, and emails you from her personal address. She never reopens the chat. She never asks for a human inside the widget. Under all three definitions, that is a billable resolution.

Run the earlier month through it. Eighty contained sessions at $0.99 is $79.20 for the month. Only 40 of those were real resolutions, so your true cost per solved problem is $1.98, not $0.99. At 500 chats a month the same gap costs you roughly $396 in billed resolutions for about 200 solved problems. The pricing is not dishonest, and nobody is hiding the definition. It is written on the pricing page. The issue is that almost nobody reads what a resolution means before signing, and the definition is the most important sentence in the contract.

This is why Studio Niza chatbots are flat at $99/month rather than per-resolution: the cost math should not get better for me when the bot answers worse. If you are comparing pricing models more broadly, I broke down all four layers in what a Shopify chatbot actually costs in 2026.

The three numbers to demand before you sign

You do not need a metrics program. You need three answers in writing, ideally in the same email thread as the quote.

1. The resolution rate, plus the sentence that defines it

Ask directly: what is your resolution rate, and how does your system decide a conversation was resolved? If the answer involves a quiet period, no reopen, or no human requested, you have been quoted containment. That is not a deal breaker. It is just a different number than the one you thought you were buying, and you should price it accordingly.

2. The escalation rate, and what happens when escalation is unavailable

A healthy small-store bot escalates something like 20 to 35 percent of real questions. A vendor reporting 5 percent escalation alongside 90 percent containment is describing a bot with a jammed exit door. Ask what the customer sees at 11pm on a Sunday when no agent is available, and ask to see that message.

3. The repeat-contact rate within 72 hours, across every channel

This is the number that catches everything else, and it is the one almost no vendor volunteers. Of the customers whose chats were contained, how many emailed, messaged, or opened a new chat within three days? That figure is your abandonment rate with a friendlier name. If the vendor cannot measure it because the chatbot cannot see your email inbox, that is a real and honest answer, and it tells you to measure it yourself.

You can approximate it in an afternoon. Export last month's contained chat sessions, pull the email addresses, and check how many of those addresses also appear in your support inbox in the following three days. On a store doing under 500 chats a month, that is a spreadsheet job, not a project.

One more thing that is not a number: ask for 20 real transcripts from a comparable store, redacted, chosen at random rather than hand-picked. Reading them takes 30 minutes and tells you more than any dashboard. Your own transcripts work the same way, which is why I treat them as free market research you should read monthly.

Wrapping up

Chatbot containment rate measures one thing accurately: how often a conversation ended without costing you agent time. That is a genuinely useful operations number, and it was never designed to describe customer experience. The trouble starts when it gets quoted as though it were a success rate, which is now the industry norm, and priced as though it were one, which is new.

Three things to carry out of this. Containment counts silence, and silence includes everybody who gave up. An 80 percent containment rate sitting on top of a 40 percent resolution rate is a completely normal month, not a broken bot. And when a platform bills per resolution, the definition of that word is a pricing term, so read it before you sign rather than after your third invoice.

The work left to you is small but real. Ask for the three numbers. Read 20 transcripts. Run the repeat-contact check on one month of data. If your containment is high and your inbox is unchanged, you already know which direction the answer is going to point, and the spreadsheet just tells you how far.

A bot that resolves 45 percent of questions honestly, and hands the rest to you cleanly, is worth more than one that contains 90 percent and quietly loses a customer a day. The second one looks better in every report you will ever be shown.

Or email contact@studioniza.com if you have a specific question about your store. I read every one.

---