What does each rate actually count?
Containment comes from call centres, where it meant the share of calls the phone menu finished without transferring to an agent. For an AI agent the idea is the same: conversations that ended without a live handover to a person, divided by all conversations. It is the easiest of the three to produce and the easiest to flatter, because a customer who gives up and closes the window counts as contained.
Deflection asks a slightly different question: how many conversations never became work for your team? A message the agent took for a callback is contained, since nobody was put through, but it is not deflected, because somebody still has to call back. Deflection is usually reported gross, as everything that ended without a person, and sometimes net, after removing the customers who came back about the same thing within a few days.
Resolution is the narrowest. The customer's question was answered or their problem fixed. It can be confirmed, when the customer says so or rates the answer, or inferred, when there was no handover and no return on the same issue within a set window. It is the only one of the three that measures the customer's outcome rather than the state of your queue.
Why does the definition matter more under per-resolution pricing?
Because then 'resolved' is what you pay for. As of October 2026, published rates for this kind of billing sit between about $0.90 and $2.00 a unit: Intercom lists Fin at $0.99 per outcome, Zendesk lists automated resolutions at $1.50 committed or $2.00 pay-as-you-go, and Gorgias lists its AI agent at $0.90 to $1.00 per automated interaction. Each vendor publishes its own definition of the unit, and they are not the same unit. Intercom's outcome, for example, covers procedure handoffs as well as resolutions.
The questions to ask are concrete. Is a conversation the customer abandoned billed as resolved? Is one that ends with a ticket? How long does the vendor wait before deciding a customer won't come back, and does a return on another channel count? None of these are trick questions. They are the difference between paying for answers and paying for silence.
Usage pricing sidesteps the argument rather than winning it. WireDesk bills $0.02 per agent reply on chat and email and $0.20 per started voice minute, whatever the conversation turned out to be, so the outcome label never moves the invoice and you are free to define resolution as strictly as you like.
What do the three rates look like on one month?
Take 10,000 conversations across chat, email and phone. The agent answered 5,600 with no person involved. 1,800 were handed to a person live. In 1,200 the agent couldn't help and took a message for a callback. 1,400 were abandoned: the customer left before an answer, or opened a conversation and never said anything.
Containment is everything without a live handover: 10,000 − 1,800 = 8,200, or 82%. Gross deflection is everything that ended without work for a person, the answered and the abandoned together: 5,600 + 1,400 = 7,000, or 70%. Now look a week later. 500 of the answered customers came back about the same problem, and 600 of the abandoned ones turned up on the phone or by email. Net deflection is 7,000 − 1,100 = 5,900, or 59%. Resolution, meaning answered and stayed answered, is 5,600 − 500 = 5,100, or 51%.
Price that month per resolution at a dollar and the definition decides the bill. Count 'didn't reach a person' as resolved and the month is $8,200. Count 'answered, and no return within seven days' and it is $5,100. That is $3,100, or about 60% more, for identical work.
Which numbers should a support lead actually track?
Report net resolution rate as the headline, with a fixed recontact window. Seven days is a common choice, and the choice matters less than never changing it. Then keep four diagnostics next to it. Abandonment rate, counted separately and never as a success. Handover rate, with how often a person actually answered and how long that took: a containment rate that rises because nobody picked up is a failure dressed as a win. Message-taken rate by topic, which is the list of things your knowledge doesn't cover yet. And satisfaction on AI-handled conversations, so a falling score shows up before the churn does.
Finally, cost per resolved conversation: the AI bill for the month divided by net resolutions. It is the number to set beside your cost per human-handled ticket, and it is the one that stops a cheap-looking unit price from hiding a lot of units.
What does WireDesk record for each conversation?
When a conversation ends (a call hangs up, or a chat or email thread has been quiet for half an hour) WireDesk reads the transcript once and files four things on it: a one- or two-sentence summary, a short topic, the customer's sentiment, and one of four outcomes. Resolved means the agent answered what they came for. Handed to you means a person took over, by live transfer or by replying from the dashboard. Message taken means the agent couldn't help and took details. Abandoned means they left before it could help, or never said anything.
The analytics page shows the share of each outcome over 7, 30 or 90 days, the topics where the agent most often took a message instead of answering, how many conversations asked for a person, and what share of those requests a person answered, with the typical wait. Website chat can also collect thumbs-up and thumbs-down ratings on answers; that is off by default.
Two caveats, said plainly. The outcome is a model's reading of the transcript, not the customer's confirmation, so treat 'resolved' as a gross figure. And WireDesk doesn't compute recontacts for you. The conversations export carries the channel, contact, outcome and topic of each conversation, so the net figure is a join on contact and date in a spreadsheet.