Almost nobody can tell whether their agents work: 5% report well-defined success metrics. Until a firm has its own numbers, promotion decisions run on the enthusiasm of the person who built the thing, and that enthusiasm is measurably wrong.
Why this matters
Two measurement failures compound here, and they are different problems with different fixes.
Does the agent do the task correctly? That is an eval, and vendor evidence will not answer it. Of 30 deployed agentic systems indexed in early 2026, four had agent-specific system cards and 25 disclosed no internal safety results, so the eval a firm writes is probably the first one its configuration has faced. Grounded accuracy also sits well below what a demo implies: on a benchmark of roughly 89,000 pages of Treasury bulletins, frontier agents with a standard retrieval pipeline plateaued below 45%. Both the benchmark and the ceiling are moving, but the practice does not change: the canonical treatment is evals.
Is the agent worth anything? That is analytics, and the honest state of the art is bad. Only 5% of 623 respondents report well-defined success metrics for agentic AI (AWS-sponsored, fielded July 2025). Where measurement does exist it is usually self-reported perception with a number attached: 72% of 801 enterprise decision-makers say they track structured business-linked ROI metrics, but the underlying question asks what ROI has been “based on internal conversations with colleagues and senior leadership.”
And perception is systematically wrong in a known direction. In a survey of about 5,000 US knowledge workers, 43% of C-suite respondents claimed 8+ hours saved per week while two-thirds of non-management reported under two hours and 40% reported none. The same seniority gradient shows up independently on ROI (45% of VP+ call it significantly positive against 27% of mid-managers). Users also overstate against their own measured baseline: in one vendor study, self-reported median speedup was about 25× against a measured figure near 7.5×. Assume internal ROI self-reports run roughly 3× high and design the measurement accordingly.
Two anti-patterns, both observed in the field rather than invented here. One firm tracks engineers’ agent interactions per day as a productivity proxy, a measure of activity rather than value. Another CEO reviews the agent spend dashboard and treats a tenfold jump as progress, telling staff they are not “going hard enough.” Spend and usage are the two metrics that rise automatically when nothing is working.
The metric that survives an examiner is different from the one that survives a CFO, and both differ from the eval. Keep them separate: eval scores gate promotion, business metrics decide whether the agent lives, and neither is evidence for the other. On the CFO side, the objection arrives in its native form: nobody can eat hours. Time saved is the least bankable currency there is, and the pre-agent citizen-automation programs that reported hundreds of thousands of hours saved mostly could not convert them into a line anyone booked.
Where you stand
| Level | Looks like | Cheapest next move |
|---|---|---|
| Crawl | No evals. Success is “the demo worked.” No agent has a defined output quality bar. | Write ten test cases with known-correct answers for your most-used agent, including three you expect it to fail. Run them today. |
| Walk | Ad-hoc spot checks by the builder. Usage and spend are the reported metrics. | Move the checking to someone who did not build it, and replace usage metrics with one outcome metric per agent. |
| Run | A regression suite per tier-2+ agent, run before promotion and after model changes; eval results are the gate’s evidence. | Add adversarial cases from real incidents, and re-run when the vendor changes the model under you. |
| Fly | Continuous scoring against production traces, drift alerts, and business metrics reconciled against a pre-registered baseline rather than a recollection. | Audit one agent’s claimed benefit against the general ledger. If it cannot be found there, say so out loud. |
Concerns this dimension covers
- Unreliable output acting on your systems — confident fabrication, the failure class needing no attacker.
- Quality debt and orphaned apps — what accumulates when nothing measures whether an agent still works.
- Grounded in: the jagged frontier — why aggregate accuracy hides task-level failure, and why a firm’s own tasks are the only benchmark that matters.
Controls that answer them
- Logging and audit — eval scores mean nothing without the traces that explain them.
- Human approval gates — the backstop while eval coverage is thin, and the control evals eventually make it safe to relax.
Who sells it
- Observability and eval platforms — scorers, trace stores, and the question of where the traces live.
Sequencing and where this is checked
- Day 3 sequencing — when eval investment stops being optional.
- Evals — the recurring practice: acceptance thresholds, regression suites, and what to do when the model changes underneath a running agent.
- 90 — where the business baseline gets set, before anyone can remember it differently.
Open questions
- What does a passable eval suite for a citizen-built agent minimally contain? Nobody has published a floor, and “write ten cases” is our judgment, not a standard.
- No published measurement connects eval coverage to incident rates. If the connection is weak, much of this dimension is expensive hygiene, worth knowing either way.