Problems · in production

The agent does the work. Then a person checks all of it.

Agents produce output faster than anyone can read it, and expert hours become the ceiling on what ships.

review-queue · manual QA example
Illustrative day, the week evaluation criteria replace the queue
Outputs produced today 412
Reviewed by a person 71
Hours of expert review 9.5
Runs auto-evaluated 0 → 412
Queue for an expert everything → failures + a sample
define good once per agent; people read only the runs worth their time
§01 / THE SYMPTOM you see: the signals
TL;DR

Manual review caps how much AI you can ship. Define what good looks like once per agent, evaluate every run against it automatically, and send people only the runs that fail or a sample worth auditing.

The symptom

What the review bottleneck looks like

The agent was meant to save time. Somewhere along the way, checking its work became a full-time job.

01
Experts do QA, not their job

The people who know the domain spend their days reading agent output instead of doing the work only they can do. Their available hours, not the agent, cap how much ships.

02
Double-checking eats the saving

Staff re-check everything the agent produces, so the hours the agent saved are spent verifying it. On paper the process is automated; in practice it has one more step than before.

03
Nobody can say what good means

Quality lives in each reviewer's head as a judgement call. Two reviewers disagree on the same output, and because the standard was never written down, nothing about it can be automated.

04
Accuracy plateaus below target

The aggregate number sits stubbornly under where it needs to be, and nobody can say which runs drag it down, because no one evaluates runs individually. The team tweaks prompts and hopes.

§02 / WHY IT HAPPENS cause: not carelessness
Why it happens

Why review became the bottleneck

Each of these was a reasonable call at the time. Together they put a person in front of every output.

01
Automated evaluation looked expensive

Building evaluation looked like a project with a budget line. A person reading outputs looked free, so nobody priced the person's hours, and the free option quietly became the most expensive one.

02
Good was never written down

The pilot was judged by feel: the expert read the output and nodded. That worked at ten outputs a week. A judgement that exists only in someone's head cannot be applied by anything but that head.

03
Every team measures differently

One team spot-checks, another keeps a spreadsheet, a third trusts complaints to surface problems. The numbers are not comparable, so there is no shared standard to automate against.

04
Checking everything felt like control

Reviewing every output feels safer than sampling, but a reviewer skimming their four-hundredth item of the day catches less than a rested one reading thirty flagged runs. Full coverage by tired eyes is thinner than it looks.

§03 / HOW YOU CATCH IT loop: observe → evaluate
How you catch it

How evaluation replaces the queue

Prefactor turns the checklist in the reviewer's head into criteria the platform applies to every run.

Define

Write down what good means, once per agent. The checks the expert applies by instinct, task completed, format met, policy respected, facts grounded, become evaluation criteria the platform can run. The standard leaves the reviewer's head.

Observe

Every run lands in one record. Outputs, tool calls, and decisions are captured as they happen, whoever built the agent, so there is a complete population to evaluate rather than whatever a reviewer happened to open.

Evaluate

Every run gets a verdict, automatically. Each run is checked against its agent's criteria: pass, fail, and why. Coverage goes from the fraction a person could read to all of it, without adding a reviewer.

Surface

Only the runs worth an expert reach one. Failures and a sample for audit route to a person; the rest do not. The queue shrinks from everything to the handful where human judgement actually changes the outcome.

§04 / HOW YOU FIX IT loop: act → improve
How you fix it

From checking everything to checking what matters

Automated verdicts clear the queue. The loop is what moves the accuracy number.

Act

Failed runs wait for a person. A run that fails its evaluation can be held or escalated before its output goes anywhere, so cutting manual review does not mean bad outputs reaching users unchecked.

Improve

The plateau becomes a diagnosis. Verdicts segmented by input type, prompt version, and model show exactly which runs fail and what changed the number. The team stops tweaking on instinct and starts fixing named failure classes.

Prove

Quality becomes a trend, not a feeling. Each agent carries its pass rate per version over time, measured the same way across teams, so "is it good enough yet" is answered with a chart rather than a meeting.

Two domain experts were reading roughly 400 agent outputs a week between them, and accuracy still sat at 81 per cent against a 95 per cent target. Writing their checklist down as evaluation criteria and running it on every run cut human review to failures plus one run in twenty, about 30 reads a week, and showed that most failures came from one input type the prompt had never handled. One prompt change moved the number more than six months of full manual review had. Illustrative, but this is the usual shape.

§06 / QUESTIONS faq: the common ones
Questions
How do I evaluate AI agent output automatically?
Define evaluation criteria per agent, the checks a reviewer applies by hand written as rules the platform can run, then evaluate every run against them as it happens. Each run gets a verdict, and people review the failures and a sample rather than the whole queue.
What if we cannot define what good looks like?
Start from what reviewers already do: the checklist they apply implicitly is the first draft of the criteria. Criteria can be rules, reference comparisons, or model-graded checks, and they are refined against the runs experts overturn, so the definition improves with use rather than needing to be perfect on day one.
Do subject-matter experts stop reviewing entirely?
No. They review the runs that fail evaluation and a regular sample of the ones that pass, which is how the criteria stay honest. Their judgement calibrates the automation; it stops being the throughput limit.
Why is our agent's accuracy stuck below target?
Usually because the aggregate number hides which runs fail. Without a verdict per run, a plateau is unreadable; with one, you can segment failures by input type, prompt version, and model and see exactly what drags the number down, which turns guesswork into a fix.
How do we standardise quality measurement across teams?
Put every agent's runs in one record and evaluate each against criteria defined for that agent. The criteria differ per agent, but the measurement, pass rate per version over time, is the same everywhere, so quality is comparable across teams for the first time.

See it in action on a fleet like yours

Book a demo and we will take one agent you check by hand, define its criteria, and show every run evaluated against them.

Agent Performance Platform
Unified performance platform for agents, authentication, and risk management
All Systems Operational
3Global Agents
7Instances
5Services
12%Human Intervene
4High Risk
$2,360Monthly Spend
Mission ControlLive agent health with 7-day activity heartbeat
Claims Proc...68
$330/moRed
Claims Proc...65
$160/moRed
Claims Proc...82
$170/moAmber
ChatGPT74
$150/moAmber

See how every agent performs, and make it better

Prefactor helps teams observe, evaluate, and improve their AI agents in production, across every framework and provider.