How to Measure Whether Your AI Agent Is Actually Working
An AI agent can be live, answering questions, and technically “working” while still leaving you with no real idea whether it’s actually helping your business. That’s not a hypothetical risk. It’s the default state for a lot of AI rollouts, because businesses launch the agent, watch tickets flow through it, and assume that if nothing’s visibly on fire, it must be doing its job.
The problem is that “nothing’s visibly on fire” isn’t a metric. If you want to know whether an AI support agent is actually delivering value, and where it needs improvement, you need a small set of numbers you’re actually tracking, not a general sense that things feel better than before.
Start with the number everyone reaches for first
The most common metric people look at is deflection rate, the percentage of inbound questions the agent resolves without any human involvement. It’s a reasonable starting point because it’s simple and it directly reflects the core promise of the agent: handling volume so your team doesn’t have to.
For a biotech firm we worked with, that number landed at around 40 percent in the agent’s first version, which is a meaningful result for a brand new build. But deflection rate on its own doesn’t tell you the whole story, and treating it as the only number that matters is a common mistake. A high deflection rate paired with a lot of unhappy customers isn’t actually a win. It just means you’re resolving tickets badly at a faster rate than you were resolving them badly before.
The metric that matters more than the headline number
The more important question isn’t how many tickets the agent handled. It’s whether it correctly recognized which tickets it shouldn’t handle. This is where escalation accuracy comes in, and it deserves more attention than it usually gets.
You want to know two things specifically. First, when a question genuinely required a human, did the agent actually escalate it, rather than attempting an answer it wasn’t equipped to give. Second, when the agent did escalate something, was that the right call, or was it punting on questions it could have handled, which just creates unnecessary friction for customers who now have to wait for a person on something simple.
For a business in a sensitive space, like one dealing with test results and health-adjacent information, this metric matters as much as deflection rate, arguably more. An agent that deflects 60 percent of tickets but occasionally answers a clinical question it should have escalated is a bigger problem than one that deflects 35 percent but never crosses that line. Answer or Escalate: How AI Agents Decide What They Can Handle covers how that decision logic should actually be designed before you even get to measuring it.
Look for the gaps the agent is quietly telling you about
One of the more useful things you can measure isn’t about the agent’s performance directly. It’s about what questions are coming in that the agent can’t confidently answer. Every time that happens, it’s a signal, either that a knowledge gap exists and an article needs to be written, or that an existing article isn’t clear or specific enough to be retrieved correctly.
Reviewing this regularly turns your AI agent into an ongoing source of information about where your documentation is thin, rather than a static tool you build once and leave alone. For the biotech team, this meant that when a new type of question surfaced repeatedly without a good answer, that was the signal to write a new knowledge article, which then synced back into the system and became something the agent could use going forward. How Your AI Agent Actually Finds the Right Answer covers how that retrieval and sync process actually works.
Track this over time, not just at launch
A single snapshot of these numbers right after launch tells you how the agent performed against your initial knowledge base and your initial set of questions. It doesn’t tell you how it’s performing three months later, once your product has changed, new questions have started showing up, or your customer base has shifted.
The businesses that get the most value out of an AI agent revisit these numbers on a regular cadence, not just once. Deflection rate should generally trend upward over time as knowledge gaps get filled. Escalation accuracy should stay consistently high, not just high in the first month. If either number starts moving in the wrong direction, that’s usually a sign that your knowledge base has drifted out of sync with what customers are actually asking, which is worth catching early rather than discovering months later.
Don’t skip the qualitative signal
Numbers tell you what’s happening, but they don’t always tell you why. It’s worth periodically reading through a sample of actual agent conversations, not just the aggregate statistics. Reading real interactions surfaces things the numbers miss, like an agent that’s technically escalating correctly but doing so in a way that feels abrupt or confusing to the customer, or an agent that’s answering correctly but in a tone that doesn’t match your brand.
This kind of review doesn’t need to happen daily, but it should happen often enough that you’d notice a pattern forming before it becomes a real problem. A monthly spot check of a reasonable sample size is usually enough to catch what the aggregate numbers alone won’t show you.
What this looks like in practice
None of this requires a complicated dashboard to start. A simple, consistent review of three things, deflection rate, escalation accuracy, and recurring unanswered questions, gives you a real picture of whether your AI agent is actually earning its place in your support process or just quietly coasting on a good first impression. The specific tools you use to track this matter less than the discipline of actually looking at it regularly, rather than assuming that a quiet queue means a job well done.
If you haven’t built this measurement habit yet and you’re still early in deciding whether an AI agent is the right move for your team, How to Audit Your Support Data Before Adding AI is a good place to start, since the same discipline that helps you measure an agent after launch is what helps you scope it correctly before you build one.
The takeaway
An AI support agent that’s live isn’t the same thing as an AI support agent that’s working. The difference shows up in the numbers you choose to track, deflection rate, escalation accuracy, and the gaps the agent surfaces in your knowledge base, and in your willingness to actually look at those numbers on an ongoing basis rather than assuming a quiet queue means everything’s fine.
Listen to the full podcast episode here.