What does an AI agent cost to run each month? Platform, usage, people and exceptions
An AI agent's monthly bill has five lines: platform, model usage, review time, exceptions and upkeep. How to model each and pick a pricing basis.
The short answer
An AI agent's monthly running cost has five lines: platform fees, model usage, the people who review its drafts, the exceptions it hands back, and upkeep as models and prompts change. For most back-office agents the people lines outweigh the model bill. Model all five per task before launch, then track cost per completed task every month.
Key takeaways
- Budget an agent as cost per completed task, built from five lines: platform, model usage, review time, exception handling and upkeep.
- Review time and exception handling are usually larger than the model bill for a back-office agent, because both are paid in staff hours.
- In KPMG's survey of 204 US leaders at $1bn+ firms (28 April–25 May 2026), only 26% had full, real-time visibility of their AI operating costs.
- Per-seat, per-task and per-outcome pricing move risk between you and the vendor, so pick the basis that matches how your volume varies.
- Set a monthly ceiling and a cost-per-task alert before launch, and review both with the agent's named owner every month.
In this article
Why the monthly bill surprises people
Most teams budget an AI agent's build and forget to budget its running cost. The build is a project with a quote. The running cost is five separate lines, billed by different people in different units, and only one of them looks like an AI cost.
Few companies can see those lines today. In KPMG's AI Quarterly Pulse, a survey of 204 US leaders at firms with $1 billion or more in revenue (28 April to 25 May 2026), 53% said their organisation was deploying AI agents.1 Only 26% reported full, real-time visibility of their AI operating costs.1 In the same survey, 66% had monitoring dashboards and 36% had direct token or usage controls.1
If firms of that size lack the view, a mid-sized company running its first agent almost certainly does too. Our note on the KPMG cost-visibility gap covered what the survey found. This post is the cost model we would build before launch.
AI agents, Q2 2026
Agents are spreading faster than cost control
The five lines of an agent's running cost
Every agent we would scope has the same five monthly lines. Name each one, assign an owner, and estimate it per task, not per month.
Platform fees
Platform fees are what you pay the vendor whose software runs the agent: a seat licence, a per-agent fee, a per-conversation charge or a share of an outcome. They are the line finance sees first, because they arrive as an invoice. They are rarely the largest line.
Model usage
Model usage is what the language model charges for the text it reads and writes, usually billed per million tokens. It scales with volume and with how much context each task drags in. An agent that reads a whole contract to check one clause costs many times more per task than one that reads the clause. Long prompts, retries and multi-agent hand-offs all add to it quietly, which is why usage controls matter.
Review time
Review time is the minutes a person spends checking what the agent drafted before it is submitted or sent. In a well-designed first agent, every action that is hard to undo passes through a person, so this line is never zero. It falls as trust grows and you widen what the agent may do alone, but that is a decision you take with evidence, not a saving you assume.
Exception handling
Exceptions are the items the agent cannot finish under your rules: a missing purchase order, a price outside tolerance, a request it was not built for. Each one goes to a person, and an exception often takes longer than doing the task by hand, because the person first has to read what the agent tried.
Upkeep
Upkeep is the monitoring, prompt fixes and model upgrades that keep the agent working as your data and the vendors' models change. Someone has to own it. Whether that is an internal engineer, the vendor or a managed service, it is a monthly cost.
Our view: for a back-office agent, review time and exceptions usually cost more than the model and the platform together. Both are paid in staff hours, which is why they vanish from AI budgets and reappear in a department's overtime.
A worked example: one month of an invoice agent
The example shows where the levers are. Cutting the exception rate from 1 in 8 to 1 in 12 saves more than halving the model bill. That is usually a data fix, such as cleaning the supplier master or the item codes, not an AI fix. Our guide to invoice matching on an ERP covers the controls that keep the review step cheap and safe.
Compare the per-invoice result with your current cost per invoice, measured the same way. If the agent does not beat it after the review and exception lines are counted, the business case rests on speed or accuracy, and you should say so plainly.
Seats, tasks or outcomes: which pricing basis?
Pick the pricing basis that matches how your volume moves. Seat pricing suits steady volume and predictable budgets. Per-task pricing suits volume that swings with the season. Outcome pricing suits a task where "done" is easy to define and hard to dispute.
Outcome pricing is spreading among the large platforms. From April 2026, HubSpot priced its customer agent per resolved conversation and its prospecting agent per qualified lead, as No Jitter reported.2 Buyers say they want this. G2's 2026 Buyer Behavior Report, from a survey of more than 1,000 software buyers (data period and region not stated), found the share preferring outcome-based pricing rose from 11% in 2025 to 23% in 2026.3
Outcome pricing moves risk to the vendor, but only if you agree what counts as an outcome. "Resolved" can mean the customer stopped replying. "Qualified" can mean a form was filled in. Our view: read the definition before the price, and ask for a monthly report of disputed outcomes.
Three ways to pay for an agent
Per seat or flat
- Predictable bill
- You carry the volume risk
- Fits steady queues
Per task
- Bill follows volume
- Failed tasks may still be charged
- Fits seasonal work
Per outcome
- Vendor carries more risk
- Depends on the definition of done
- Fits tasks with a clear end state
Who signs off the bill?
Finance does, increasingly. In the same G2 report (data period and region not stated), finance involvement in software purchases rose from 31% to 46%.3 A finance reviewer will ask for cost per task, the ceiling, and what happens if volume doubles. Have those answers before the meeting.
Three controls make the answers credible. A monthly spending ceiling, set in the model provider's console and the platform. An alert when cost per completed task drifts above the budgeted figure. And a named owner who reviews both lines every month alongside the exception rate. Our scoring model for the first process to automate uses the same inputs, so the cost model and the business case match.
On our published ladder, Managed AI-ops covers monitoring, cost control and model upgrades from $2,000 (₹1L) a month, after an automation sprint puts the workflow live; both are listed on our pricing page. You can run the same controls in-house. What matters is that someone owns them.
