Skip to content
be-human.ai
← Back to the journal

Testing & reliability

Agent Reliability Metrics: Four Measures for a Company Agent Layer

Track task success, escalation rate, correction rate and recovery time to judge whether a company AI agent is reliable. Definitions and a weekly review routine.

Be Human typographic cover with the title: Agent Reliability Metrics: Four Measures for a Company Agent Layer
Conceptual illustration created with AI · Be Human

Agent reliability metrics are the measurements that show whether an AI agent finishes real business work correctly, when it hands work to a person, how often people must fix its output, and how fast the operation recovers after something breaks. For a company agent layer, four measures do most of the work: task success, escalation rate, correction rate and recovery time. Tracked weekly with written definitions, they tell an owner whether an agent should receive more authority, less authority, or a narrower job.

Key takeaways

  • Define task success against a written outcome, not against whether the agent produced a reply.
  • Read escalation rate and correction rate together, because a falling escalation rate can hide errors that staff quietly fix.
  • Recovery time runs from detection to a verified fix, including any queued work that must be reconciled.
  • Set provisional thresholds for each task type before the first review, and change them only through a logged decision.
  • Review the four numbers in one short weekly sitting with real examples open beside them.

What a reliability metric measures for a company agent

Model benchmarks answer a narrow question: how a model performs on a fixed test. A company agent is more than a model. It reads a request, consults memory, calls tools, respects permissions and writes to a system of record. A high benchmark score does not tell you whether the refund it issued matched your policy, or whether the reminder it sent reached the right contact. Reliability metrics measure the agent in operation, on the work your business actually runs.

The NIST AI Risk Management Framework treats measurement as an ongoing activity tied to the risks a system creates, not as a single score produced at launch. That framing fits an operations leader's job. You need a small set of numbers that move when something important changes and that point you toward the cause. The four measures below were chosen for that purpose. Each has a clear denominator, a named owner and a clear next action when it moves.

Task success: define it against the outcome

A task is successful when its outcome meets the acceptance criteria written before the agent ran. For a billing reminder agent, that might mean the correct invoice was identified, the message used the approved template, the send was logged, and the record in the system of record matches what was sent. Task success rate is successful completed tasks divided by tasks attempted in the period. Escalated tasks are not counted as successes, and tasks abandoned because a tool failed are counted separately.

The most common mistake is measuring whether the agent responded at all. An agent that writes a fluent answer to the wrong question is busy, not successful. Anthropic's engineering guidance on evals for agents makes a similar point: grade the outcome the agent produced in its environment, not only the transcript of what it said. Write acceptance criteria as observable facts, such as a status changed, a field populated or a message sent to a named address. If two reviewers would disagree about whether a run passed, the criteria are not yet specific enough. A practical method for building those criteria is described in How to evaluate AI agents on real business tasks.

Escalation rate: how often the agent knows to stop

Escalation rate is tasks handed to a person divided by tasks attempted. It answers a question owners often skip: does the agent recognize the cases it should not decide alone? A well-designed agent escalates when a policy requires a human, when inputs are missing or contradictory, when a tool returns an unclear result, or when the requested action exceeds its approval boundary. An agent that never escalates is not reliable. It is guessing, and the guesses tend to surface later as corrections or incidents.

Read the rate by reason. Split escalations into policy required, uncertainty, tool failure and out of scope. A rising share of tool failure escalations usually points to an integration problem rather than an agent problem. A falling rate of uncertainty escalations after a prompt or permission change deserves a closer look, because the agent may have become more confident for the wrong reasons. The handoff matters as well. A useful escalation sends the decision and the evidence behind it, not a raw transcript. The guidance in Human escalation for AI agents: send the decision, not the mess covers that handoff in more detail.

Correction rate: what people fix after the agent acts

Correction rate is completed tasks that a staff member edited, reversed or re-performed, divided by completed tasks reviewed. The denominator comes from reviewed tasks, not from all tasks, so the number depends on how you sample. Most teams cannot review every output, so they review a risk-weighted sample and report the rate for that sample. Record the sample size and the selection method alongside the rate so the figure can be compared across weeks.

Correction rate is the metric that catches quiet errors. Escalation rate counts the cases the agent flagged. Correction rate counts the cases it did not flag but a person still had to change. Tag each correction with a cause, such as wrong field, wrong tone, missing context, outdated policy or permission misuse. The pattern across causes is more useful than the total. If most corrections trace to outdated policy, the fix belongs in the agent's knowledge and memory, not in a longer prompt. Feed recurring corrections back into the agent's rules so the same fix is not made by hand every week.

Recovery time: how long a failure stays open

Recovery time is the elapsed time from detection of a failure to a verified return to normal operation. Define detection as the first moment someone or something flags the failure. Define normal as the state in which the workflow runs, its outputs pass the agreed checks, and any work queued during the failure has been reconciled. Stopping the agent is not recovery. A paused agent with a backlog nobody reconciled has not recovered.

Report recovery time per incident category, and keep the longest case visible rather than only the average. A single long recovery usually reveals a missing runbook step, an unclear owner or a dependency nobody mapped. Recovery time also depends on whether the agent can resume without repeating completed work, because retries that create duplicate records lengthen recovery and damage trust. Track the time to a verified fix as well as the time to a first response, since teams often mistake a quick acknowledgment for a resolution. The steps for running that resolution are set out in How to respond to an AI agent incident: a practical playbook.

Putting the four metrics into a weekly routine

Start with one agent role and one task type. Broad rollouts make the numbers hard to interpret. Then work through these steps:

  • Write the acceptance criteria for each task type as observable facts, and name the person who can change them.
  • Log every run with its outcome, any escalation reason, any staff edits, and timestamps for detection and resolution. The logging fields are described in AI Agent Observability: What to Log, Trace and Alert On.
  • Tag each escalation and correction with a short cause category from a list agreed in advance.
  • Review a small, risk-weighted sample each week, and increase the sample only when the findings justify it.
  • Set provisional thresholds for each metric, record them, and change them only through a logged decision.
  • Hold a short weekly review with the four numbers, two or three raw examples and one recorded action.

Illustrative scenario: a small company uses an agent to draft payment reminders for overdue invoices. Over one month, the task success rate holds steady, but the correction rate rises, mostly for tone and for reminders sent to contacts who had changed roles. The escalation rate also falls after a permissions update. Read together, the numbers point to the cause: the agent has stopped escalating contact changes, and staff are catching them by hand. The response is to restore the escalation rule, add those cases to the test set, and keep reviewing the sample for two more weeks before changing any threshold.

Be Human designs these agents with the metrics defined from the start. The first engagement sets one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations then keep the numbers in view through monitoring, recurring evaluations, permission reviews and incident response, and every controlled change is recorded in a change log the owner can read. If you are not yet sure which task should become the first agent role to measure, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Which of the four metrics should a small team measure first? Start with task success and escalation rate, because together they show whether the agent completes work and whether it knows when to stop. Correction rate can follow once someone reviews a sample each week, and recovery time becomes meaningful after the first real incident. Keep the definitions short enough that one person can apply them without a meeting.

What is a good target for these metrics? The right thresholds depend on the task, the cost of an error and how much authority the agent holds, so no universal number is worth copying. Set provisional thresholds for your own task types, record them, and tighten or relax them only through a logged decision after you have seen real results. A threshold that never changes is usually a threshold nobody reviews.

How is correction rate different from an error rate? An error rate usually counts failures against a fixed test set, while correction rate counts completed tasks that people had to change during real operation. The second measure captures quiet errors the agent never flagged, which is why it belongs in the weekly review next to escalation rate. Each measure answers a different question, and both are needed.

Do these metrics need special software? A spreadsheet or a simple log can hold them at the start, provided each run records an outcome, an escalation reason, any staff edits and two timestamps. Dedicated observability tools help as volume grows, but the definitions matter more than the tool. Without clear definitions, any dashboard will display numbers that nobody knows how to act on.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →