AI agent performance comes down to four numbers you can read straight off the work: accuracy, approvals needed, turnaround, and rework. Measure these four in an agent's first month, and together they tell you whether the agent should get a bigger job or stay where it is.
CG Landscape shows how it plays out. It started with finance agents on contract renewals, which now finish 6x faster, and on check entry, where 2 to 3 hours of work that used to start every morning went to the agents. Marketing agents came after that. Argaux started with marketing agents and added finance and operations. Neither company handed over every job at once. The rest of this piece is how you decide when a job is ready to be added.
The first month is the measuring period
Every agent we build is a teammate with one job, a manager, and an approval step. A person approves everything that leaves the building, and in month one that approval step is also the measuring instrument. Each time someone approves, edits, or rejects a draft, they leave a data point, and by the end of the month you hold a record instead of an impression. The first three weeks of that month are laid out in your first three weeks, start to finish.
Without that record, "how's the agent doing?" gets answered from memory of the last time someone looked, and a wrong entry can sit there until a customer finds it.
Some of the measuring is done by other agents. On finance work, an auditor agent reviews every deliverable before it goes out. On marketing, an editor-in-chief agent reviews every piece. When one of those reviewers sends work back, count it under approvals needed.
The four numbers
| Number | The question it answers | What it looks like for a payment-entry agent |
|---|---|---|
| Accuracy | Was the work right when it was delivered? | The check was read correctly, matched to the right invoice, and entered for the right amount |
| Approvals needed | How often does a person step in, and how often do they change something? | Entries a person approved untouched, against entries a person corrected |
| Turnaround | How long from the trigger to a finished deliverable? | Time from a check arriving to an entry waiting for approval |
| Rework | How much comes back after it was approved? | An entry reopened a week later because it landed on the wrong invoice |
Accuracy gets checked against the source, never against the agent's own confidence. For checks, the source is the check. Accuracy is the obvious number and the least useful on its own, because an agent can be right on every easy case and wrong on the hard ones. So also watch what the agent does when it isn't sure. The payment agents at CG Landscape send anything uncertain to a person instead of guessing. Count a flagged check as a good outcome and a confident wrong entry as the worst one.
Approvals needed is really two counts. How many items needed a person at all, and of those, how many did the person change. A click-and-move-on approval costs almost nothing. An approval where someone rewrites half the draft means the agent is still doing somebody else's job, and the reviewer is doing the rest. Edits that fall week over week mean the agent is learning the role. Edits that stay flat in week four usually mean the instructions are wrong, and more time won't fix that.
Turnaround needs one split. When it looks slow, find out whether the agent was slow or the approval queue was. The fixes differ. One is a tooling and instructions problem. The other is a conversation about who approves and by when.
Rework is the number scorecards often skip, and it catches the failures that show up late. Accuracy is measured at the gate. Rework is what happens after: the entry reopened, the post corrected after it went live, the renewal sent with the wrong year's pricing. If rework climbs while accuracy at the gate looks fine, your reviewers have usually started rubber-stamping, and the review is the place to look.
When an agent earns more
The rule is that the four numbers have to agree. Picture two payment agents at the end of week four. The first has accurate entries, edits that dropped every week, a queue that clears the same day, and nothing reopened. The second has accurate entries at the gate and quick turnaround, but a few entries came back a week later on the wrong invoice. On the approval screen they look identical. On the scorecard, the first one has earned something and the second one has a rework problem that nobody has noticed yet.
There are two kinds of "more." One is a quicker approval on the same job. The other is a new job. Do them in that order. At the end of month one, go through the record with whoever manages the agent: every miss, and every place a person had to step in. A weekly meeting is a good place for that review; how we partner describes the rhythm we use with clients. The approval step stays on every job. As the edits stop, approving a payment entry or a draft post goes from a rewrite to a glance and a click.
What a mistake costs sets the bar, so we don't publish one magic threshold. A weak headline on a draft blog post and a wrong amount on a payment are different problems, and the second needs a longer clean record before the agent gets more. Sensitive changes like budgets or pricing keep needing approval from both teams no matter how good the numbers get.
The new job comes after that. At CG Landscape, the marketing agents that followed finance got the company showing up in ChatGPT. At Argaux, marketing agents doubled inbound leads in 90 days, and finance agents later replaced the fractional CFO with a daily cash flow forecast. Operations agents now keep inventory data current. An agent handling one task becomes a digital worker running a whole process. Most of the value we see comes from that expansion.
When an agent should not get more
If edits are still flat in week four, look at the job description before the agent. Usually the process was never written down. At CG Landscape the renewal cycle ran on spreadsheets and know-how that lived in people's heads, which is why we document how the work gets done before we build anything. You can't score an agent against a standard nobody has stated.
If an agent never flags anything uncertain, be suspicious. Real work has edge cases, and an agent that reports a perfect run on messy inputs is guessing quietly.
And sometimes the answer is that the job is too small. If a task happens once a month and takes a person twenty minutes, a better scorecard won't make it worth giving the agent more. You don't need to expand everything. An agent that stays at one job and does it reliably is a good outcome.
What to ask for
If you already run an agent, from us or from anyone, ask for these four numbers for the last 30 days. If nobody can produce them, you have a tool you hope is working.
If you want to see what this would look like for your first agent, book a planning session. Real work starts within 2 to 4 weeks.
FAQ
What is AI agent performance?
AI agent performance is how reliably an agent does its assigned job, measured against its work rather than its demos. It comes down to four numbers: accuracy, approvals needed, turnaround, and rework.
How do you measure AI agent performance?
Measure AI agent performance through the approval step. A person approves every output, and in the first month each approval, edit, or rejection becomes a data point. Check accuracy against the source material, count how often a person changes the work, time the work from trigger to finished deliverable, and track what comes back after approval.
When does an AI agent get more responsibility?
There is no fixed date. The end of month one is when to review the full record. If the four numbers agree, approvals on the same job get quicker and the agent can take on a new one. The bar is higher where a mistake costs more, and the approval step stays on every job.
What if the numbers are bad?
Start with the job description and the review process before blaming the agent. Flat edits usually point to unclear instructions, and rising rework with clean approvals usually points to rubber-stamping. Either way, the agent stays at its current responsibility until the numbers improve.