AI & Strategy • September 2026

New AI Agent Metrics: What to Track in Self-Service Journeys

AI agents are in production for real now. The old IVR KPIs — containment, AHT, abandonment — still matter, but they no longer tell the whole story. Agentic self-service introduces new failure modes, so it needs new observability and reporting metrics. Here's the direct list, and why each one earns its place.

When self-service was deterministic, monitoring was simple: did the caller reach the right place, and how long did it take? Agentic AI changes that. The agent now reasons, calls tools, and generates responses — which means it can be fast and contained but still wrong, unsafe, or expensive. You need metrics that see inside the agent, not just around it.

Here's how these metrics might look on an AI agent observability dashboard. The visuals below are illustrative mock-ups with sample data to show how you'd surface each data item.

AI Agent Observability — Self-Service (last 24h) · sample data
Hallucination Rate
0.8%
▼ 0.3pts vs last week
Off-Policy Breaches
12
▲ 4 vs last week
Tool Call Success
97.4%
▲ 1.1pts vs last week
Cost / Interaction
$0.041
▲ $0.006 vs last week
Containment quality vs. raw containment
100% 0 62% Raw 48% Resolved
Response groundedness
91% grounded

Example dashboard: headline KPI cards with week-on-week movement, plus containment quality and groundedness. Green/amber/red framing surfaces quiet failures at a glance.

Quality & Safety

1 Hallucination Rate

The percentage of agent responses containing fabricated, unsupported, or factually wrong information.

Why: This is the number-one risk of generative agents. A contained interaction that gave the customer wrong information is worse than no automation at all. Track it, sampled by an LLM-as-a-judge, on as close to 100% of interactions as you can.

2 Off-Policy / Guardrail Breach Rate

How often the agent says or does something outside its defined scope — giving advice it shouldn't, discussing restricted topics, or ignoring a rule.

Why: In regulated sectors, a single off-policy response is a compliance incident. This metric is your early-warning system before it becomes a fine or a headline.

3 Response Groundedness

The share of answers that can be traced back to an approved source (knowledge base, retrieved document, system data) rather than the model's own memory.

Why: High groundedness is the strongest defence against hallucination. If answers aren't grounded, you're trusting the model's training data — which you don't control.

4 Faithfulness

Of the answers that are grounded in a source, how accurately they represent that source — without distorting, exaggerating, or contradicting it.

Why: Groundedness asks "did it use the right source?"; faithfulness asks "did it tell the truth about that source?" An agent can cite the correct policy document and still misstate what it says. Faithfulness catches that subtle, dangerous failure — and it's the core RAG-evaluation metric for exactly this reason.

5 Sentiment Trajectory

Not just final sentiment, but how it moved across the interaction — did the customer end calmer than they started?

Why: An agent can complete the task and still leave the customer angry. Trajectory reveals whether the experience actually felt good, at a scale no human QA team could measure.

Agent Behaviour & Tools

6 Tool Call Success Rate

The percentage of tool/API calls the agent makes that succeed (correct tool, valid parameters, successful response).

Why: Agentic resolution depends on tools actually working. A high failure rate here is the hidden cause of "the AI couldn't help me" — even when the language understanding was perfect.

7 Tool Selection Accuracy

When the agent chose to act, did it pick the right tool for the customer's actual need?

Why: The agent can call a tool successfully but call the wrong one — processing a change the customer never asked for. This catches reasoning errors that success rate alone hides.

8 Conversation Turns to Resolution

The average number of back-and-forth turns needed to resolve an intent.

Why: Replaces raw handle time for agents. Rising turns signal the agent is looping, misunderstanding, or over-clarifying — a direct measure of conversational efficiency.

9 Escalation Rate & Escalation Reason

How often the agent hands off to a human — and, crucially, why (out of scope, low confidence, customer request, repeated failure).

Why: Escalation isn't failure; a clean handoff is good design. But the reason breakdown tells you exactly where to improve the agent next.

10 Containment Quality (not just Containment Rate)

Of the interactions the agent contained, how many were genuinely resolved versus abandoned or wrongly trapped in self-service.

Why: Containment rate can be gamed by simply not offering an agent. Containment quality is the honest version — resolution, not entrapment.

Cost & Operations

11 Cost per Interaction (Token & Model Cost)

The real compute cost of each interaction — input/output tokens, model tier, and tool calls.

Why: Agentic cost is variable, not fixed. A verbose prompt or a chatty agent can quietly multiply your bill. This metric didn't exist in the deterministic world — now it's essential for ROI.

12 Latency per Turn (and Time-to-First-Token)

How long the customer waits for the agent to respond, including model inference and tool-call time.

Why: Reasoning and tool calls add latency that scripted IVR never had. Silence kills voice experiences — you have to watch response time as closely as accuracy.

13 Model & Prompt Version Attribution

Every interaction tagged with the exact model version, prompt version, and config that produced it.

Why: When quality shifts, you need to know what changed. Without version attribution, you can't tell a model update from a prompt tweak — and you can't roll back with confidence.

14 Model Drift / Regression Rate

Change in quality scores over time against a fixed evaluation set, especially after a model or prompt update.

Why: Foundation models and prompts change behaviour subtly. Drift detection catches silent degradation before customers do.

Trends & Diagnostics · sample data
Hallucination rate over time (with model update marker)
3% 0 model v4.2 spike W1 W8
Escalation reasons (share of handoffs)
Out of scope Low confidence Customer ask Repeated fail 41% 28% 19% 12%

Example diagnostics: a trend line with a version marker makes model drift obvious (note the spike after v4.2), while the escalation-reason breakdown shows exactly where to improve the agent next.

The one-line takeaway: Deterministic IVR failed loudly and visibly. Agentic AI fails quietly — a confident, fast, contained answer that happens to be wrong, unsafe, or expensive. These metrics exist to make quiet failure visible.

Where to Start

You don't need all fourteen on day one. Start with the three that catch the most dangerous failures: hallucination rate, off-policy breach rate, and tool call success rate. Add cost per interaction and latency as soon as you're at any scale, and layer in the rest as you mature. Evaluate as close to 100% of interactions as you can — with an LLM-as-a-judge, that's finally affordable.

Keep your traditional KPIs too — containment, CSAT, FCR, AHT still matter. The point isn't to replace them. It's to add the new lens that agentic AI demands, so you can trust what you've deployed.

Designing an agentic self-service journey and want to map out its logic and test cases first? Try the free IVR Design Tool.