.png)
Last updated: July 31, 2026
Every support leader who has deployed an AI agent eventually gets asked the same question by their CFO: is it actually working? The answer usually arrives as a single percentage — the containment rate — and that number carries an uncomfortable amount of weight. It shows up in board decks, renewal negotiations with your AI vendor, and headcount planning conversations. Yet most teams calculate it differently, compare it against benchmarks that were never meant for their business, and draw conclusions their customer experience data flatly contradicts.
Containment rate is worth measuring. It is also the single easiest support metric to game, and gaming it quietly destroys customer trust. This guide covers what containment rate really measures, what a healthy number looks like by channel and use case, why a rising containment rate is sometimes a warning sign, and how to build a scorecard that survives scrutiny from both finance and CX.
What's in this guide
- What containment rate actually measures
- What a good containment rate looks like
- Why a high containment rate can still be a failure
- Five ways to raise containment without hurting CX
- Building a containment scorecard finance and CX both trust
- Frequently asked questions
What containment rate actually measures
Containment rate is the share of customer conversations that an automated system handles from start to finish without a human agent joining. If 1,000 people started a chat last week and 640 of them finished without a handoff, your containment rate was 64%.
The formula
Contained conversations ÷ total conversations entering the automation = containment rate. Simple arithmetic, three decisions hidden inside it:
- What counts as "entering the automation." If you exclude conversations where the customer immediately typed "agent," your denominator shrinks and your number inflates.
- What counts as a handoff. A customer who abandons the chat and calls your phone line has not been contained — but most dashboards record it as contained.
- What window you measure. A conversation "contained" on Tuesday that generates a new ticket on Wednesday should not count. Many measurement systems never connect the two.
Containment vs. deflection vs. resolution
These three terms get used interchangeably and mean genuinely different things. Vendor analyses of containment versus deflection make the distinction clear: deflection asks whether the ticket was prevented from reaching an agent, containment asks whether the conversation stayed inside the automation, and resolution asks whether the customer's problem is gone. Only the third one is a customer outcome.
The gap between them is where credibility is lost. If you already track ticket deflection across your support channels, containment is a narrower, more honest cousin: it only describes what happened inside the conversations you actually routed to AI, and it says nothing about whether those customers walked away satisfied.
Key takeaway: Containment is an efficiency metric, not a quality metric. Reported on its own it is almost meaningless. Paired with a resolution or repeat-contact metric, it becomes one of the most useful numbers in your operation.
What a good containment rate looks like
There is no universal target. Containment depends almost entirely on how repetitive your inbound mix is and how much your AI agent is allowed to actually do. A bot that can only answer questions is capped far lower than one that can look up an order, issue a refund, or reschedule a delivery.
| Setup | Typical containment | What limits it |
|---|---|---|
| Scripted FAQ chatbot | 25–45% | No system access; breaks on any phrasing it wasn't trained on |
| LLM agent, answers only | 40–60% | Understands the request but cannot act on it |
| AI agent with system actions | 55–75% | Integration coverage and permission limits |
| Mature enterprise deployment | 60–80% | Genuinely complex or regulated cases that should reach a human |
Ranges compiled from published benchmark work by Bookbag's ecommerce containment benchmarks and Notch's AI resolution rate benchmarks.
Read those numbers as a diagnostic, not a goal. If you are sitting at 30% with an agent that has full CRM and order-system access, your problem is intent coverage. If you are at 75% with an answers-only bot, your problem is probably measurement.
Why a high containment rate can still be a failure
Containment counts conversations that ended inside the automation. It does not care why they ended. A customer who gives up, closes the window, and posts a complaint on social media is, by most definitions, contained.
This is not a hypothetical failure mode. Gartner expects agentic AI to autonomously resolve 80% of common customer service issues by 2029, cutting operational costs by 30% — but the same firm also predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, largely on unclear value and escalating cost. Teams that optimise the containment number instead of the customer outcome end up in the second group.
The three signals that expose fake containment
- Repeat contact rate within 7 days. If contained conversations generate follow-up tickets at a higher rate than agent-handled ones, containment is hiding unresolved work.
- Channel switching. Track whether a customer who was "contained" in chat appears on your phone queue or in your inbox within 24 hours.
- Abandonment mid-conversation. Separate customers who got an answer from customers who simply stopped replying. These should never be in the same bucket.
Teams that already measure first contact resolution with AI in the loop have most of the instrumentation they need. The move is to report containment and resolution as a pair, always, and never let one appear in a deck without the other.
Five ways to raise containment without hurting CX
1. Fix your top ten intents before adding an eleventh
In most support operations a small number of intents drive the majority of volume — order status, password resets, billing questions, delivery changes, returns. Pull the last 90 days of transcripts, cluster them, and check the containment rate of each intent individually. A blended 55% often hides one intent at 90% and three at 15%. Fixing the weak ones moves the aggregate far more than adding new coverage.
2. Give the agent permission to act, not just to answer
The single largest jump in containment comes from connecting the AI agent to the systems where the answer lives and the action happens: order management, CRM, billing, logistics. An agent that can read an order status contains a "where is my order" conversation; an agent that can also reschedule the delivery contains the follow-up too. This is why automating "where is my order" requests end to end tends to be the highest-leverage first project for retail and logistics teams.
3. Design the handoff as a feature, not a failure
Counterintuitively, a fast and graceful escalation path raises containment over time. When customers trust that they can reach a human, they stop opening every conversation with "agent" — which means more conversations get a real chance at automation. Building a clean AI-to-human handoff workflow with full context transfer is a containment investment disguised as an escalation project.
4. Set a confidence threshold and honour it
Let the agent escalate when it is uncertain rather than guessing. Every confident wrong answer costs you a repeat contact, a trust hit, and often a refund. Tuning the threshold down to force containment is the classic way teams turn a good deployment into a bad one.
5. Review failed containments weekly
Take a sample of escalated conversations every week and label the reason: missing knowledge, missing integration, ambiguous request, genuinely human. That label set becomes your roadmap. Most teams find that two or three missing integrations account for a third of all escalations.
This is the operating model we built Eva, Darwin AI's customer experience agent, around — she works across WhatsApp, email, and chat with direct access to the systems where orders, subscriptions, and billing records live, so containment comes from actually completing the request rather than from a well-worded deflection. The distinction matters more than the percentage.
Building a containment scorecard finance and CX both trust
The reason containment debates get heated is that finance and CX are optimising different things. A shared scorecard ends the argument. Report five numbers together, every month, with no exceptions:
- Containment rate — conversations completed inside the automation.
- Resolution rate — contained conversations with no repeat contact in 7 days.
- CSAT on contained conversations — measured separately from agent-handled ones.
- Cost per resolution — not cost per contained conversation.
- Escalation reason mix — the weekly labels from step five above.
Cost per resolution deserves particular attention. Gartner now predicts that GenAI cost per resolution for customer service will exceed offshore human agent costs by 2030, which means the economics of automation will increasingly hinge on resolution efficiency rather than raw volume absorbed. A deployment that contains a lot of conversations expensively and resolves few of them will not survive that shift.
One more discipline: hold your containment target flat while you improve resolution. Most teams discover that resolution climbs several points while containment stays still, because the agent stops "containing" conversations it was never going to solve. That is progress, and a scorecard built this way makes it visible instead of looking like a regression. If you also run AI-assisted SLA management, feed escalated conversations into the same clock so a handoff never resets the customer's wait.
Frequently asked questions
What is a good containment rate for an AI support agent?
It depends on capability. Published benchmarks put scripted FAQ bots at roughly 25–45%, answer-only AI agents at 40–60%, and action-taking AI agents at 55–75%, while mature enterprise deployments commonly land in the 60–80% band. Compare yourself to your own trend line first and to benchmarks second.
What is the difference between containment rate and deflection rate?
Deflection measures tickets prevented from reaching a human at all, usually including help-centre visits. Containment measures conversations that entered the automation and finished there. Deflection is a volume metric and containment is a conversation metric, and neither one tells you whether the customer's problem was solved.
How do I calculate containment rate correctly?
Divide conversations completed without human involvement by all conversations that entered the automation, including the ones where the customer asked for an agent immediately. Then subtract any conversation that produced a new contact on another channel within 24 hours. That second step is what separates a defensible number from a flattering one.
Can a containment rate be too high?
Yes. If containment rises while CSAT on contained conversations falls or repeat contacts increase, the automation is trapping customers rather than helping them. Some conversations — cancellations, complaints, complex commercial negotiations — should escalate by design.
How long does it take to improve containment?
Intent-level fixes to knowledge and prompts show up within two to four weeks. Integration work that lets the agent take real actions takes longer to ship but produces the larger and more durable gain, because it converts answered questions into completed requests.
Raise containment by completing requests, not by deflecting them. Eva resolves support conversations end to end across WhatsApp, email, and chat — with a clean handoff when a human is the right answer.
See how Eva works →
