Containment rate is the percentage of support conversations that finish inside an automated channel, a chatbot, IVR, or AI agent, without ever reaching a human. The formula is:
Containment Rate = (Contained Sessions ÷ Total Sessions Entering the Channel) × 100
A high containment rate can mean your AI agent genuinely solved the problem. It can also mean the player gave up before anyone helped them, and the dashboard on its own can’t tell you which.
This guide covers how to measure containment honestly, what good actually looks like by AI maturity tier, and how five AI support tools report it against CSAT.
What Is Containment Rate?
Containment rate is the share of conversations that enter an automated support channel and end there without escalating to a human agent. It’s a channel-level efficiency metric. Don’t treat it as a quality metric, since a session can be contained and still leave the player unhappy.
The number only tracks whether a handoff happened, not what the player actually got out of the interaction.
1. The Containment Rate Formula, With a Worked Example
The formula is:
Contained Sessions ÷ Total Sessions Entering the Channel × 100
Say an AI agent handles 2,000 conversations in a month and 1,400 finish without a handoff request. Containment rate for that channel is 70%. That’s the whole calculation: three numbers and a division. What actually counts as “finished without escalating” is where teams start disagreeing, and that disagreement is the entire honesty problem covered below.
2. Containment Rate vs Deflection Rate vs Resolution Rate
These three get used interchangeably, but they measure different things.
- Containment is channel-level: did this specific automated channel finish the session without a handoff?
- Deflection is portfolio-level: across every channel, what share of contacts never reached a human at all, counting help center articles and FAQ pages that stopped a contact before it started?
- Resolution is the only one of the three that’s an outcome statistic, since it asks whether the player’s problem actually got solved, not just whether a human touched it.
3. What Counts as an Escalation, and Why the Definition Changes Your Number
A loose definition counts any session that ended without the player pressing an escalation button, which overstates containment because it folds in frustrated abandonments.
A stricter, honest definition only counts a session as contained if the player confirmed resolution, or if no new contact on the same issue arrived within a defined window, commonly 24 hours. That single adjustment routinely pulls reported containment down by several points, and it’s the difference between a number that survives an audit and one that doesn’t.
Containment Rate Benchmarks by AI Maturity*
Containment benchmarks vary by AI maturity tier, and no single universal number exists. The intent mix drives your actual containment rate more than model quality does. A studio whose volume skews toward simple FAQ questions will contain higher than one fielding mostly account recovery, on the exact same AI agent. That’s why containment belongs on the same dashboard as the support KPIs gaming studios track, rather than standing alone as a single headline number.
| AI maturity tier | Typical containment rate |
|---|---|
| Rule-based, scripted bot | 25–45% |
| Retrieval-only AI (answers questions, can’t act) | 40–60% |
| Action-taking AI agent (executes changes, not just answers) | 55–75% |
| Best-in-class, mature deployment | 60–80% |
Source line: Benchmark tiers compiled from published containment research,
*Actual results depend heavily on your specific intent mix.
Why Gaming Containment Differs by Intent
Gaming containment isn’t one number; it is a spread across intent types that most published benchmarks never break out.
- Missing rewards and general FAQ questions (patch notes, system requirements, account settings) contain high, since the answer is static and the AI agent doesn’t need to touch a live system to close the loop.
- Purchase disputes and account recovery contain meaningfully lower, since both involve verifying a live account state, checking a payment record, or reversing an action.
These are exactly the intents where an answers-only bot has to hand off. Reading player analytics at the intent level, rather than as one blended number, tells a studio where its AI agent is strong and where it’s papering over a gap.
Why Containment Rate Alone Is a Vanity Metric
Containment rate alone is a vanity metric because it can’t distinguish a solved problem from an abandoned one. A player who asks a question, gets an unhelpful answer, and closes the chat in frustration is, by most measurement definitions, contained.
Gartner’s own 2024 survey found that only 14% of customer service issues were fully resolved through self-service channels, even though far more than 14% of those sessions never reached a human agent.
1. Pairing Containment With CSAT and Reopen Rate
Containment climbing while CSAT falls is the clearest signal that players are being trapped in automation rather than helped. Reopen rate solves the remaining issue. A session that was contained on Tuesday and reopened on Thursday was never actually resolved; it just took two days to admit it.
This is exactly why in-game support and CSAT should be read as a pair as opposed to two separate dashboard tiles.
2. The Three-Number Reading a CX Director Should Report
Containment, resolution accuracy, and CSAT, reported together, are the minimum a CX director can defend to finance. Containment alone answers “how much reached a human.” Resolution accuracy answers “how much of that automation actually worked.” CSAT answers “did the player feel good about it.” Report any one or two of the three, and you’re presenting a routing statistic as a quality claim.
5 AI Support Tools Compared on Containment and CSAT
These five platforms take genuinely different approaches to reporting containment, and the differences matter more than any single headline percentage. Some separate resolution from containment explicitly; others still blend the two in ways that flatter the dashboard.
| Platform | Containment reporting depth | Resolution vs containment separation | CSAT attribution | Intent-level breakdown |
|---|---|---|---|---|
| Decagon | Deep, publishes a full companion-metric framework | Explicit, treats resolution and containment as separate questions by design | CSAT is one of five companion metrics it recommends | Not detailed in public materials |
| Intercom Fin | Resolution rate paired with a per-conversation CX Score | Explicit, reports resolution alongside deflection | CX Score applied per conversation | Not detailed in public materials |
| Zendesk AI Agents | Automated-resolution count against a metered allowance | Blended; independent analysts report the advertised rate diverges from real-world performance | CSAT tracked separately inside Suite | Limited public detail |
| Ada | Automated Resolution (AR), built to replace raw deflection | Explicit, AR was designed specifically to not reward abandonment | CSAT tracked alongside AR | Not detailed in public materials |
| Helpshift | Native Care AI Agent Analytics dashboard | Explicit, Issues Resolved and Reopen Rate are reported as separate metrics | CSAT is a native metric in the same dashboard | Supported, Smart Intents routes and reports by intent |
1. Decagon
Decagon publishes one of the more rigorous public frameworks for this exact problem, recommending containment be read alongside recontact rate, escalation rate, and CSAT. The honest limitation, however, is that rigor lives in Decagon’s glossary content more than in a documented, intent-level reporting feature a buyer can point to in a demo.
2. Intercom Fin
Fin pairs resolution tracking with a CX Score on every conversation, and independent reviewers note it’s more transparent than most agent vendors about the deflection-resolution gap. But Fin is measuring Fin. A buyer still needs an independent read on whether the self-reported number holds up.
3. Zendesk AI Agents
Zendesk reports automated resolutions against a metered monthly allowance, with CSAT tracked separately elsewhere in Suite. However, third-party analysts have flagged that the advertised resolution rate and real-world performance diverge, which is exactly the crack a containment number can hide if resolution isn’t reported next to it.
4. Ada
Ada built its Automated Resolution metric specifically to avoid rewarding abandonment the way raw deflection does. But since it is an enterprise platform, its published reporting depth doesn’t break down to the intent level the way a gaming-specific deployment would need.
5. Helpshift
Helpshift’s Trust & Safety and Support solutions both run on the native Care AI analytics dashboard, reporting Issues Resolved, CSAT, Time to Resolve, Time to First Response, Reopen Rate, and Issue Touches as separate, named metrics.
Care AI (the AI agent) resolves 70%+ of player queries autonomously, and Helpshift’s 2024 Benchmark Report found 58% of all support interactions were fully automated. Because Issues Resolved and Reopen Rate are tracked separately, a studio can see containment and genuine resolution as two different lines.
However, this depth is built specifically for gaming and player-driven products. Helpshift isn’t a general-purpose contact center platform.
How to Improve Containment Without Hurting CSAT
Improving containment without hurting CSAT means fixing what the AI agent can actually do before tuning how it talks. The two most common containment ceilings are thin knowledge coverage and an AI agent that can only answer questions instead of taking action.
1. Fix Knowledge Coverage Before Tuning the Model
A containment ceiling is usually a knowledge issue, not a model problem. Long-tail intents, including edge-case refund scenarios and less common patch issues, fail because nothing in its approved knowledge covers them. Expand coverage before reaching for a different model.
2. Let the AI Agent Take Actions Beyond Just Answering Questions
An agent that can only answer questions hits a hard ceiling on purchase disputes and account recovery, since both require a real action, a refund issued, an account state changed, and not just an explanation. Action-taking agents consistently outperform answers-only bots on exactly these intents.
3. Set Escalation Triggers Deliberately
Escalation should be a designed path. It should never be an accident of what the AI agent couldn’t figure out. Set explicit triggers for the intents that need a human agent, including VIP accounts, contested cases, or anything touching real money. Don’t put off escalation until the AI starts running out of ideas.
4. Track by Intent and Pull Back Where Containment Is Bought With Frustration
If containment is climbing on an intent where CSAT is falling, that’s containment bought with frustration. It is not earned with a better answer. Pull the AI agent back on that specific intent rather than accepting the aggregate number at face value.
Start Reporting Containment the Honest Way
Containment, resolution accuracy, and CSAT, read together, are the reading that survives scrutiny from finance and from your own trust in the dashboard. Any one number alone is a claim about routing dressed up as a claim about quality.
If automation rate and CSAT are what you report up the chain, look for reporting that names these as separate metrics rather than blending them into one score. See the Helpshift platform reference for how the Care AI analytics dashboard reports Issues Resolved, CSAT, and Reopen Rate side by side, so containment never has to stand in for resolution.
Frequently Asked Questions
1. What is a good containment rate?
A good containment rate depends heavily on AI maturity and intent mix, so there’s no single universal target. As a range: rule-based bots typically land below 35%, retrieval-only AI reaches 40 to 60%, action-taking agents reach 55 to 75%, and best-in-class deployments reach 70 to 80%. Intent mix moves this number more than model quality does, so compare your rate against your own intent breakdown, not a single industry average.
2. What is the difference between containment rate and deflection rate?
Containment rate measures a single automated channel, whether a session that entered a chatbot or IVR finished there without a handoff. Deflection rate measures the whole support portfolio: the share of all contacts across every channel that never reached a human at all, including contacts stopped by a help center article before automation was ever involved.
3. How do you calculate containment rate?
Containment Rate = (Contained Sessions ÷ Total Sessions Entering the Channel) × 100. If 2,000 conversations enter an AI agent and 1,400 finish without escalating, containment rate is 70%. The honest version of this calculation defines “contained” strictly, counting only confirmed resolutions or sessions with no repeat contact within a set window, typically 24 hours.
4. Can a high containment rate be a bad sign?
Yes. A high containment rate paired with falling CSAT or a rising reopen rate usually means players are giving up rather than getting resolved, since most measurement definitions count an abandoned session as contained. Read containment next to CSAT and reopen rate before treating a high number as good news.
