Why are agent hours a bad unit?
- Runtime is not output. A slow or stuck agent bills more hours for the same result.
- The client cannot check it. Nobody can verify how long an agent "worked" or what it did in that time.
- It hides the AI savings. The client sees a new line item, not a lower cost for the same scope.
- It punishes better tools. A faster model earns the firm less.
How do firms bill AI work today?
| Model | What the client pays for | Can the client compare vendors? | Does the firm keep margin when AI gets faster? |
|---|---|---|---|
| Agent hours | Model runtime | No | No |
| Per task | A fixed price per task type | Partly, if tasks match | Yes |
| Cost plus AI spend | Tokens and seats, plus a markup | No | Only the markup |
| Per delivery point | Finished work, scored on one scale | Yes | Yes |
What goes on the invoice?
- The delivery points that the client accepted in the period.
- The price per point that you agreed in the contract.
- A link to the evidence: each point traced to the pull request or ticket behind it.
Do not add agent hours or token costs as separate lines. Your AI spend is your cost, the same as a laptop or an IDE.
How do you move a client off hours?
- Show the baseline. Score the last months of work in delivery points, so the client sees what an hour bought.
- Quote the next scope in points. Pass part of the AI savings to the client as a lower price per point.
- Review quality with the client. Show defects and rework next to the points, so the lower price does not look like lower quality.
Common mistakes
- Billing agent hours at a human rate. The client finds out, and the trust goes.
- Billing tokens at cost. You give away the know-how that made the agent useful.
- Hiding agent work inside human hours. It works until the client asks.
How Pensero helps
- Pensero scores every pull request and ticket in delivery points, for engineers, AI-assisted engineers and AI agents, and traces each point to the work behind it.
- Pensero shows AI spend per point by client, team and model, so you know your margin per point.
- Read how to price software outsourcing when AI agents write the code and how IT outsourcing firms win bids by pricing per delivery point.
- Book a scoping session to bill your next scope in delivery points.
Sources
- Keito, 20 March 2026: How to Bill Clients for AI Agent Work.
- BCG, 2025: Rethinking B2B Software Pricing in the Era of AI.
- The Wall Street Journal, 13 November 2026: Baker McKenzie Looks to Abandon the Billable Hour as Default.
Frequently asked questions
Should software vendors bill clients for AI agent hours?
No. Agent hours measure runtime, not delivered work. Bill the finished work at a price per unit.
How do you bill a client for work an AI agent did?
Score the finished work in one unit, such as delivery points, and bill the accepted units at the agreed price.
Should AI costs be a separate line on the invoice?
Not on an output-based contract. AI spend is the firm's cost. Only a cost-plus contract shows it.
Does output-based billing lower the firm's revenue?
Not if you price per point above your cost per point. When AI gets faster, your margin per point goes up.
