Enterprise AI KPIs: The Metrics Leaders Should Track

Sunset boardroom with a holographic AI KPI dashboard, with icons for adoption, quality, oversight, and cost

A reference glossary of AI KPIs, with a definition, a formula, and a common mistake for each.

A KPI is only useful if two people calculate it the same way. This glossary gives twelve AI metrics a definition, a formula, and the most common mistake made when reporting them. They fall into four groups: adoption and value, quality and reliability, oversight, and cost and governance.

Active Usage Rate

The share of employees with access to AI tools who use them regularly.

In Gallup’s tracking of U.S. employees, 47% said their organization had integrated AI tools, while 30% used AI a few times a week or more (as of May 2026).

Workflow Coverage

The share of a defined workflow’s volume that runs through AI.

  • Formula: Cases handled with AI ÷ total eligible cases.
  • Common mistake: Counting deployed use cases instead of volume, or counting a case as covered when AI only assisted one step of it.

Cycle Time Change

The change in elapsed time from request to completion, measured against a baseline.

  • Formula: (baseline median time minus current median time) ÷ baseline median time, over the same period and mix of work.
  • Common mistake: Relying on self-reported time savings.

In METR’s randomized trial, developers believed AI had made them 20% faster, while the measured result was a 19% slowdown. METR later called that result outdated for newer tools, but the lesson is the gap between belief and measurement.

Business Outcome Change

The change in a metric the company already tracks, such as revenue, margin, cost, or loss rate, that can be attributed to the AI workflow.

  • Formula: The metric after launch minus the metric before launch for that workflow, compared with a control group or a baseline period.
  • Common mistake: Crediting AI with the whole change, or reporting expected gains as realized ones.

An NBER survey of nearly 6,000 executives found that 89% reported no productivity impact from AI over the past three years, while the same executives forecast a 1.4% boost over the next three.

Task Success Rate

The share of tasks completed correctly against a defined standard.

  • Formula: Tasks completed correctly ÷ tasks attempted.
  • Common mistake: Scoring each task once, which hides how often the same task fails on a repeat run.

Consistency (pass^k)

The probability that an agent succeeds on all of k repeated attempts at the same task.

  • Formula: The share of tasks solved on every one of k runs.
  • Common mistake: Reporting pass@k, the chance of at least one success in k tries, for work where every attempt reaches a customer.

The τ-bench paper introduced pass^k for that reason. Function-calling agents such as GPT-4o succeeded on under 50% of tasks, and in retail, fewer than 25% of tasks were solved even after 8 tries. Those were 2024 models with simulated customers, so the figures show the gap, not current levels.

Escalation Rate

The share of tasks an agent hands to a person.

  • Formula: Escalated tasks ÷ total tasks, split by reason.
  • Common mistake: Treating a low rate as success.

Anthropic’s analysis of Claude Code found that agent-initiated stops, such as asking for clarification or approval, are an important form of oversight. A near-zero rate can mean the agent isn’t stopping when it should, and a high rate can mean the use case isn’t ready.

Human Intervention Rate

The share of agent turns or runs that a person interrupts or corrects.

  • Formula: Human interruptions ÷ agent turns.
  • Common mistake: Reading a rising rate as falling quality.

In the same Anthropic study, experienced users interrupted Claude Code in about 9% of turns against 5% for new users, while also auto-approving more often, which the study reads as a shift from approving every step to active monitoring.

Review Time per Output

The time reviewers spend checking an AI-produced output.

  • Formula: Total review hours ÷ outputs reviewed.
  • Common mistake: Counting the time saved producing an output while leaving out the time spent validating it.

HBR research published in June 2026 found that middle managers took on validating AI outputs and finding errors while delivery pressure stayed the same or rose.

Cost per Successful Task

The total cost of a type of work divided by the number of times it succeeded.

  • Formula: (inference, tooling, infrastructure, and review costs) ÷ tasks completed successfully.
  • Common mistake: Tracking cost per token or per request alone.

The FinOps Foundation defines cost per inference as total inference costs ÷ inference requests, and notes that cost can also be expressed per anticipated outcome. It also cautions that a token-only view leaves out the fixed and semi-fixed costs that decide whether an initiative is viable at scale. Retries and failed attempts raise cost per successful task without changing cost per request.

Agent Inventory Coverage

The share of agents in production that are registered with a named owner.

  • Formula: Agents on the register with a named owner ÷ agents discovered in the environment.
  • Common mistake: Measuring against a manually kept list, which only counts agents someone remembered to register.

A Cloud Security Alliance survey commissioned by Strata Identity found that only 21% of organizations maintain a real-time registry of their agents and only 28% can trace agent actions back to a human sponsor across all environments.

Overdue Permission Reviews

The share of agent permission grants that are past their review date.

  • Formula: Grants past review date ÷ total active grants.
  • Common mistake: Counting reviews held instead of privileges revalidated or removed.

NIST’s control for reviewing privileges calls for validating the need for each privilege at a defined frequency and reassigning or removing any that can no longer be justified.

A scorecard does not need all twelve. A workable set takes one metric from each group and ties each to an owner, a baseline, and a review date.

Les Ottolenghi Avatar

Keep reading

More from the blog.

Booking for 2026–27

Bring an operator into the room.

Keynotes, board briefings and advisory for leaders who need AI and cybersecurity to be a strategy, not a headline.