Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
KPIs

Deflection vs. Containment Rate: Which Number to Trust

AI SNS CHAT EditorialPublished 6 min read

Deflection and containment sound alike but measure different things. Learn how to define, calculate and use both for website chat, WhatsApp and Messenger.

Your dashboard says the AI handled 80% of conversations, yet your support team still feels busy and customers still complain. The likely culprit is that "deflection" and "containment" are being used as if they were the same metric. They are not, and mixing them up leads to flattering numbers that hide real problems. This article defines both, shows how to calculate them, and explains which one to report to whom.

Why the two terms get confused

Both metrics describe the share of customer contacts that never reach a human agent. That is where the similarity ends.

  • Deflection asks: did a customer avoid creating a human-handled contact because of self-service? It is usually measured at the entrance, such as someone who searched a help page or opened a chat and left without a ticket.
  • Containment asks: of the conversations that actually started with the AI, how many finished without a human taking over? It is measured inside the conversation.

Vendors often use the words interchangeably, which is why a number quoted in one report cannot safely be compared with a number quoted in another. Before you benchmark anything, write down which definition you are using.

Definitions and formulas

Containment rate

Containment is the simpler of the two because it only needs conversation data.

Containment rate = conversations resolved by the AI without human involvement ÷ all conversations the AI started

A conversation counts as contained only if no human took over and no escalation was triggered. Note that "contained" is not automatically "resolved". A customer who gives up and closes the chat window is contained by this formula, but not helped.

Deflection rate

Deflection needs a denominator that includes people who never started a conversation.

Deflection rate = contacts avoided through self-service ÷ (contacts avoided + contacts that reached a human)

The hard part is the numerator. You cannot see a person who did not contact you, so deflection is always an estimate. Common proxies include help-page visits that ended without a ticket, or chat sessions where the customer left after reading an answer.

A worked example

These numbers are hypothetical. Suppose a shop receives 1,000 conversations a month on its website chat, WhatsApp and Messenger combined.

  • The AI starts all 1,000.
  • 700 end with no human involvement.
  • 300 are handed to a person.

Containment is 700 ÷ 1,000 = 70%.

Now suppose a further 500 visitors read the help page that month and left without contacting anyone, and you decide to count them as deflected. Deflection becomes (500 + 700) ÷ (500 + 700 + 300) = 80%.

Same business, same month, two different headline numbers. Neither is wrong, but they answer different questions.

Containment
Deflection
Measured inside the conversation
Measured before a contact is created
Denominator: conversations the AI started
Denominator: all demand, including silent visitors
Exact, from chat logs
Estimated, depends on proxies
Shows how well the AI finishes a job
Shows how much load self-service removes
Risk: counts abandoned chats as wins
Risk: counts people who simply gave up
Containment vs. deflection at a glanceIllustrative example

What each metric is good for

Use containment to judge the AI agent itself. If containment drops after you change the knowledge base or persona, you have a clear signal about answer quality and coverage. It is also the right metric for tuning hand-off rules.

Use deflection to judge the support operation as a whole. If you publish better help content and ticket volume falls, deflection captures that, even when the chat itself did not change.

A practical split:

AudiencePreferred metricWhy
Support leadContainmentTied to agent behaviour and tunable
Finance or operationsDeflectionCloser to workload and staffing
Product teamBoth, plus topic breakdownShows where documentation is missing

The traps that inflate both numbers

High rates are not automatically good. Watch for these.

  1. Abandonment counted as success. A customer who does not get an answer and leaves looks identical to a satisfied one in a simple containment count.
  2. Blocked hand-off. If the AI makes it hard to reach a person, containment rises while satisfaction falls.
  3. Channel mix. WhatsApp and Messenger customers often reply hours later, so a conversation may look unfinished. Decide in advance how long silence counts as resolved.
  4. Repeat contacts. A customer who returns the next day with the same question was not truly contained the first time.
  5. Topic mix. An average hides the fact that order-status questions may be contained at a very high rate while billing disputes should rarely be.

A measurement approach that holds up

Follow these steps to get numbers you can defend.

  1. Fix your definitions in writing. State what counts as a conversation, what counts as human involvement, and how long silence is treated as closed.
  2. Segment by channel. Report website chat, WhatsApp and Messenger separately before combining them.
  3. Segment by topic. Group conversations by subject so that a low rate on a sensitive topic does not get averaged away.
  4. Sample the transcripts. Read a small, random set of "contained" conversations every week and check whether the customer was actually helped.
  5. Track the reasons for hand-off. Each escalation is a lesson about a missing answer or a rule that is too cautious.
  6. Review trends, not single weeks. Small volumes make weekly percentages swing wildly.
  1. 01Conversation starts
  2. 02AI answers or hands off
  3. 03Outcome tagged and sampled
  4. 04Rate reported with quality check
From conversation to a trustworthy rateIllustrative example

Setting sensible targets

There is no universal "good" rate, and any figure quoted as an industry standard should be treated with caution. Your ideal level depends on your products, your regulations and the complexity of your questions. A better target than a single percentage is a direction: raise containment on routine topics while keeping or improving customer ratings, and keep hand-off easy for everything else.

Improving containment without hurting experience

If your containment is lower than you would like, work on causes rather than on the number.

  • Fill knowledge gaps. Look at questions the AI could not answer and add the missing content.
  • Teach by correction. When a human edits a reply, use that as the ideal answer for next time.
  • Tighten the persona. Clear guardrails stop the AI guessing at topics it should pass on.
  • Offer structured steps. For routine tasks such as capturing contact details, a form is more reliable than free-form chat.
  • Keep the hand-off honest. Tell the customer what happens next instead of trapping them in a loop.

Doing this in AI SNS CHAT

AI SNS CHAT does not present a single "containment rate" tile, so the practical approach is to build the metric from the data it records, using clear definitions.

Find the conversations. Chat widget, WhatsApp, Messenger and most other channels appear under Activity → Chat logs. Email conversations live in the Inbox. In Chat logs you can filter, flag, escalate, resolve or delete in bulk, and export the data for a spreadsheet. A simple weekly export lets you count total conversations and those that were escalated or taken over, which gives you containment per channel.

See the topics. Analytics includes Chats, Topics, Question history and Sentiment. Topics show which subjects dominate, so you can calculate rates by subject rather than one blended average. Advanced analytics require the Pro plan.

Check quality. Thumbs up/down ratings in Chat logs act as your quality signal. Review ratings on conversations that were not escalated before you celebrate a high rate.

Understand hand-offs. Staff can take over a chat, which pauses the bot. Under Helpdesk → Scheduling you set shifts per person and timezone, and hand-offs are assigned to whoever is on shift. Alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. The customer receives a "what happens next" message based on the rota, and the system never invents a wait time. Helpdesk needs Standard or higher.

Close the gaps. Where a human edits a reply into the ideal answer, the agent learns from it. Add missing content under Data sources, which accepts files, text snippets, websites, Q&A pairs and more. Use the Playground to test changes before they reach customers.

Capture leads. The Collect leads action and custom forms put details under Activity → Leads and create a contact, which is useful for conversations that should end in a follow-up rather than an answer.

For deflection, the estimate is yours to define: for example, count help-page visits that ended without a chat or email, using your own web analytics, and combine them with the chat figures. See /features for the full list of tools, and /pricing for plan details.

Frequently asked questions

Is a higher containment rate always better?

No. A high rate can result from customers abandoning chats or from a hard-to-find hand-off. Pair it with ratings, repeat-contact checks and transcript sampling.

Which should I report to leadership?

Report containment for the AI's performance and deflection for overall workload, but state the definition beside each number. Leadership usually cares most about total contacts reaching staff and customer satisfaction.

Can I measure deflection from chat logs alone?

Not fully. Chat logs show people who started a conversation, whereas deflection includes people who never did. You need an external source, such as help-page analytics, to estimate the silent portion.

How long should I wait before treating a WhatsApp or Messenger chat as resolved?

Choose a window that fits your customers' reply habits, such as a day or two, and apply it consistently. Because these customers often reply later, a short window will understate true resolution.

#KPIs#deflection rate#containment rate#support analytics

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.