How to Set a Price per Result for an AI Agent

Start from what the work costs you today

When a client asks what one result should cost, I ask them to time 30 items first, with a stopwatch rather than an estimate. A supplier invoice that "takes a couple of minutes" usually takes eight once you count finding the purchase order, chasing the receipt and typing the variance note. The number from the stopwatch is the one the whole price depends on.

Multiply the minutes per item by the monthly volume and by the loaded cost of a staff minute, and you have the cost of the work today. For 1,500 invoices a month at eight minutes each and $45 an hour, that is $9,000 a month. Loaded cost should include benefits and overhead, so ask finance for it rather than the team lead.

Run the agent on real work before anyone names a price

A two-week acceptance test gives you the other two numbers. The first is the pass rate: the share of items where the agent's output clears every check and the approver accepts it without rework. The second is review time, because a person still looks at each passing result before it goes anywhere that matters. In the invoice example, 85 percent passed and review took about a minute each.

Teams are tempted to skip the test and price from a vendor demo. The demo ran on clean samples. Your pass rate on scanned PDFs from 400 suppliers will be lower, and the price has to hold up at that rate.

Work out the break-even price

After launch your team still handles the items that fail and still reviews the ones that pass. Subtract both from today's cost and divide what is left by the number of passing results. The answer is the most you could pay per result and come out even.

In the example, 225 failed invoices still take eight minutes each, which costs $1,350. Reviewing 1,275 passing invoices at a minute each costs $956. That leaves $6,694, and spread across 1,275 passing invoices the break-even price is $5.25.

Set the price well under that line

A fair price leaves most of the saving with the buyer. At $1.50 per passing invoice the bill is $1,913, the client keeps about $4,800 a month, and a $10,000 build is paid back in a little over two months. At $4 a result the vendor takes most of the value while the buyer still carries the risk of a pass rate that slips.

The monthly minimum matters at low volume. Ours is $1,200, which covers monitoring and model updates in a quiet month. At $1.50 a result, the minimum stops mattering once the agent passes 800 results a month. If your volume is well below that, a fixed build without a per-result plan may suit you better, and we will tell you so.

A second example, where the minimum does the work

Low volume changes the picture, and it is worth seeing what that looks like. Take an independent insurance agency that handles 600 certificate of insurance requests a month at about 10 minutes each. The BLS puts the median wage for insurance policy processing clerks at $23.97 an hour, and its June 2026 compensation survey has benefits at 30 percent of total pay, so a loaded hour costs roughly $34. The work by hand costs about $3,400 a month.

Say the acceptance test shows 80 percent of certificates passing every check, with a minute and a half of CSR review each. The 120 failed requests still cost $680 of staff time and the review costs about $408, so the break-even price is around $4.82 per certificate. At $2 a certificate the result fees come to $960, which is under the $1,200 minimum, so the bill is $1,200 and the agency saves about $1,100 a month.

That is still a saving, but a $10,000 build would take about nine months to pay back. For an agency at that volume we would point to our agency sprint and smaller agency builds first, and keep per-result pricing for when volume grows or a second workflow joins the first.

Know which number moves the price most

Pass rate moves the economics more than anything else. Each extra point takes failed items off your team and adds billable results, so a vendor paid per result has a reason to keep raising it. Minutes per item today come second. If a task is genuinely quick by hand, there is less saving to share, and the price per result has to be small to make sense.

Model usage is the number people forget. When the agent runs in your own cloud tenant, the tokens appear on your cloud bill rather than the vendor's invoice. For document work it is usually small next to staff time, but ask whether any price per result you are quoted includes it.

What happens when the numbers change

The price is set from a test, and real volume rarely matches a test for long. If the pass rate climbs from 85 to 92 percent, more items are billed and fewer land on your team, so both sides gain and the price per result can usually stay where it is. If volume doubles, the minimum stops mattering and the bill grows with the work, which is the point of pricing this way.

The case to plan for is a pass rate that falls, usually because suppliers change their invoice layouts or a new document type starts arriving. Fewer items pass, so the bill drops and your team picks up the slack. We would rather fix the agent than reprice it, and the contract should say that either side can ask to reset the price from the last quarter of logged results if volume or pass rate moves a long way.

How we do it at QueryNow

We build the agent in two weeks and run the acceptance test on your own work, and the $10,000 build fee is due only if it passes. The test numbers then set one fixed price per passing result, written into the contract with the $1,200 monthly minimum. Anything that fails a check, or that your team finishes, is not billed. The calculator on our Results as a Service page runs these sums on your own numbers.