The AI agent eval metrics that matter are the ones that gate a deploy — intent accuracy, response accuracy against known-correct answers, policy compliance, escalation precision, and tone — scored on a golden set of real cases before traffic moves. Public leaderboards and demo vibes do not. In a published D2C support deployment, a suite built from 2,000 historical tickets with known resolutions hit 94.2% intent classification accuracy, 91.7% response accuracy, 99.1% policy compliance, and a 4.3/5.0 tone score before staged rollout; confidence gates of 0.85 (intent) and 0.80 (retrieval) kept low-certainty tickets with humans. That is the ruler. Without it you are debating feelings — which is why you don't have an AI strategy until you have an eval suite.
Which metrics gate a deploy — and which are vanity
Score the behavior that can hurt the business. Everything else is commentary.
| Metric | What it answers | Gate it? |
|---|---|---|
| Intent / task classification accuracy | Did the agent route the case correctly? | Yes — wrong route = wrong tools |
| Response / action accuracy vs known-correct | Did it do the right thing on labeled cases? | Yes — the core quality score |
| Policy / compliance hit rate | Did it break a documented rule? | Yes — refunds, PHI, contract language |
| Escalation precision & recall | Did it hand off when it should — and only then? | Yes — silent wrong answers and over-escalation both cost money |
| Tone / brand match (scored) | Would your best human send this? | Yes for customer-facing; lighter for internal |
| Public benchmark scores | How does the base model rank? | No — not your workflow |
| Demo "feels good" | Did stakeholders smile in the room? | No — dies on day-two traffic |
Resolution rate and CSAT belong in the operating dashboard after launch — they are outcomes, not pre-deploy gates. In the published support build, 64% autonomous resolution and CSAT moving 3.9 → 4.5 showed up after the suite passed and traffic ramped. Do not put outcome KPIs in the SOW as if they were eval metrics; put the accuracy and policy numbers that predict them.
The failure mode is mixing layers. Teams that skip the suite and watch live CSAT discover regressions from customers, not from a red CI run. 95% of enterprise generative-AI pilots deliver no measurable P&L impact (MIT, 2025) — the operational gap is almost always measurement and ownership, not the model card.
Build the golden set from real cases, not invented prompts
A golden set is a scored collection of representative inputs, defined correct outputs, and automated checks. Pull from production history — tickets, calls, form submissions — not from marketing scenarios. Label what "correct" means for each case: the right classification, the right answer or action, the right escalation.
Size and composition
Aim for hundreds to low thousands for a first production workflow, stratified by volume and by failure mode. The published support agent tested against 2,000 historical tickets with known resolutions — enough to cover the top categories and the nasty tail. Under-sampling awkward cases is how you ship 90% accuracy on happy path and blow up on the 10% that create refunds and chargebacks.
What every case must carry
- Input as the system will see it (raw text, structured fields, attached context).
- Known-correct label — intent, expected answer or tool call, or "must escalate."
- Policy tags — which rules apply (return window, discount eligibility, medical-advice refusal).
- Difficulty flag — easy / edge / adversarial, so you can report accuracy by band, not one blended number that hides the tail.
Who labels
Domain owners label; engineers wire the runner. If only the vendor labels "correct," you bought a vibe check. The suite becomes the spec for the production agent stack — evals sit next to retrieval, observability, and guardrails as a required layer, not a nice-to-have appendix.
Pass thresholds and confidence gates belong in the SOW
Write numbers into the statement of work before build week one. Vague "we'll tune quality" is how pilots never exit.
Thresholds that matched a live deploy
Use the published support suite as a concrete pattern — not a universal law for every vertical:
- Intent classification ≥ 94% on the golden set (they hit 94.2%).
- Response accuracy ≥ 90% against known-correct answers (91.7%).
- Policy compliance ≥ 99% (99.1%).
- Tone ≥ 4.0/5.0 against your best human baseline (4.3/5.0).
Your floors may differ for clinical-adjacent or payments workflows — raise them; do not invent softer ones after a bad demo week.
Confidence gates at inference time
Evals score the model offline. Confidence gates decide online whether this case may go autonomous. In the same deployment, the agent replied without a human only when intent confidence exceeded 0.85 and retrieval relevance exceeded 0.80; below either threshold the ticket routed to a human with the draft and context pre-loaded. That split is the difference between "64% autonomous" and "64% confidently wrong."
Put both in the SOW: offline pass criteria for the suite, online gates for runtime. No green suite run, no promote. No gate breach without escalation.
What changes after go-live
Launch is not the finish line. Vendor model updates, prompt edits, and retrieval corpus drift all move scores. The suite must run on every change — prompt, model, retrieval config — before production sees it.
Regression rhythm
- Re-run the full golden set on every candidate change; block deploys on red.
- Sample live traces weekly — especially escalations and low-confidence paths — and promote hard cases into the suite.
- When a model vendor deprecates a version, score the replacement on the suite before the forced cutover; keep rollback warm.
That ongoing discipline is what Managed AI Operations is for: $3,000–$20,000/month so quality movement is a number someone sees, not a ticket from an angry customer. The same eval layer that gates first launch is the layer that keeps the agent honest at month six.
What to do this week
1. Export 90 days of real cases for one workflow — tickets or forms — and pick 200–500 to label first; expand toward ~2,000 if volume supports it. 2. Define correct with the domain owner: classification labels, acceptable answers/actions, mandatory escalations, and policy tags. 3. Write four gate numbers into a one-pager: intent accuracy, response accuracy, policy compliance, and the confidence thresholds for autonomy. 4. Refuse a demo-only milestone — if the vendor cannot run your golden set, they cannot ship your agent. 5. Price the diagnostic: book the $25,000, 2-week AI Transformation Sprint to build the suite and a scoped pilot path that credits toward a build, or start from Gigabit Agents (from $8,000 flat) when the workflow and metrics are already clear.
You cannot ship what you cannot score. Put the metrics in the contract, build the golden set from your own ugly cases, and treat every model change as a deploy that needs a green run. The full support proof write-up shows how those numbers looked before traffic moved — and what autonomous resolution looked like after.
