Deflection rate, containment rate, resolution rate: which one tells you if your AI agent works?

What is deflection rate?

Deflection rate is the share of customer contacts that were handled without reaching a human agent. It comes from help-desk reporting, where every question answered by a help-centre article, a bot or another self-service channel is one ticket your team didn’t have to open.

The formula: deflection rate = contacts handled without an agent ÷ all contacts × 100. If your store receives 10,000 questions in a month and 6,000 never reach your team, your deflection rate is 60%. Count each customer’s attempt once – adding self-service contacts to agent contacts double-counts everyone who tried self-service first and then escalated.

The catch is in what “handled” means. A contact is deflected if it didn’t reach a person – for whatever reason. A customer who couldn’t find an answer, closed the tab and went to a competitor is a deflection. So is one who will write again tomorrow about the same order. Deflection measures how much work stayed away from your team, not how many customers got what they came for.

What is containment rate?

Containment rate is the share of conversations with a bot that ended without being handed over to a human. The term comes from call centres, where an automated phone menu “contained” a call if the caller never pressed through to an operator.

The formula: containment rate = bot conversations with no handover ÷ all bot conversations × 100. It’s narrower than deflection – it only looks at conversations the bot actually had – but it has the same blind spot. A shopper who asks a question, gets an unhelpful answer and leaves the chat was never handed over, so the conversation counts as contained. That’s why containment can mislead: it tells you the bot kept the conversation, not that it solved anything.

What is resolution rate?

Resolution rate is the share of conversations in which the customer’s question was actually answered or their request actually completed. It’s the strictest of the three, because it asks about the outcome, not about where the conversation ended.

The formula: resolution rate = resolved conversations ÷ all conversations × 100. The hard part is deciding what “resolved” means, and it’s worth being strict. In an AI agent for e-commerce customer service, a conversation should count as resolved only if the question got a correct answer from the store’s own data, or the action the customer asked for – an order lookup, a change, a cancellation – was done. A conversation the agent couldn’t answer shouldn’t count, and neither should one that only looked finished because the customer left. In AskSpot, the AI marks every conversation as resolved or unresolved after it ends, against definitions you can edit, and you can open any of them to check the transcript. That resolved share is the figure the dashboard shows.

Deflection vs containment vs resolution, side by side

	Deflection rate	Containment rate	Resolution rate
What it counts	Contacts that never reached a person	Bot conversations with no handover	Questions answered or requests completed
Customer who gave up	Counts as a success	Counts as a success	Doesn’t count
Handover to your team	Doesn’t count	Doesn’t count	Doesn’t count
What it tells you	How much work stayed away from your team	How often the bot kept the conversation	How often customers got what they came for

All three are legitimate numbers. The problem starts when one is presented as another – when a containment rate is labelled “resolved by AI” and compared with someone else’s genuine resolution rate.

Why the same AI agent can report 98% and 55%

We see this gap in our own data. Across the stores AskSpot runs on, counting conversations as “handled” whenever they weren’t handed over typically reads 10 to 32 percentage points higher than counting the conversations that were actually resolved – on the very same chats. On one marketplace account the gap was wider still: the “no handover” view said 98%, the resolved view said 55%. Same agent, same period, two honest-looking numbers.

That’s why we publish the stricter one. On the AskSpot case studies that report it, resolution ranges from 55% to 81% – a believable range for real stores with real catalogues, complaints and edge cases, and a long way from the 95%+ claims you’ll see in ads.

A good handover isn’t a failure

Chasing deflection pushes in the wrong direction: the fastest way to raise it is to make it harder to reach a person. But some questions should reach your team – a damaged parcel, a payment problem, an angry customer (more on what to automate and what to keep human). Handled well, these conversations don’t count as resolved by AI, and that’s fine. What matters is that the customer got to the right person quickly, with the details already collected.

Where the handover goes depends on your setup – your store’s own live chat, a person joining inside the same chat window, or a ticket. We compare the options in live chat vs AI chat.

How to measure your AI agent every week

One number is never enough. A short weekly routine tells you far more than a single headline rate:

1. Resolution rate – the main number. Watch the trend week to week, not one snapshot.
2. Handover rate and the reasons. Which topics go to your team, and should they? A rising share of handovers on one topic usually means a gap in the knowledge base.
3. A sample of unresolved conversations. Read twenty a week. Most fixes come from here – a missing answer, an outdated policy, a product the agent couldn’t find.
4. Ratings, with care. Only a small share of shoppers rate a chat, and one-star ratings tend to cluster on the conversations that weren’t resolved anyway. Use ratings as a signal, not as proof.
5. For a sales chat, sales metrics too. Add-to-cart and purchase rate from chat tell you whether the agent sells, not just whether it answers. Track them in Google Analytics – see how to track chat in GA4.

AskSpot’s analytics and reporting marks every conversation as resolved or unresolved, tags its topic and keeps the full transcript behind it, so you can filter, read and run this routine without exporting anything.

Questions to ask a vendor about their numbers

• Is this a deflection, containment or resolution rate – and how exactly is it calculated?
• What’s in the denominator: all conversations, or only the ones the bot “understood”?
• Does a customer who abandons the chat count as a success?
• If the same customer writes again about the same issue, is the first conversation still “resolved”?
• Who decides that a conversation was resolved – the system, the customer or your team?
• Can I read the transcripts behind the number?

See real resolution rates on live stores: GadziSklep resolves 81% of its exotic-pet chats with AI, and Konesso resolves 68% while 21.6% of its chats end in a purchase.

FAQ