01 The €1.27 draft still needed a human
Sebastian Mueller is a founding partner of MING Labs, where we build and run AI agents alongside our human team.[S4] In his September 2026 operating report, he described a revealing pair of numbers: €1.27 in computing costs per proposal draft, plus 15–25 minutes of human review.[S2]
A proposal draft is one step towards an offer you can send to a client. Someone still has to check whether it answers the brief and whether the company can stand behind it. The computing cost tells you what it took to generate the draft; the review time tells you part of what it took to make that draft useful.
That is where an agent budget should begin: follow the work until your team can use the result. The example above shows how the costs can add up. Below, we walk through the calculation so you can replace our assumptions with your own.
02 The software bill is one of three costs
Sebastian separates running the system, reviewing its work and owning its operation.[S2] Each needs a line in the budget, because each pays for a different part of getting the work done.
- Runtime: the software bill. Model calls, tools, hosting and other services the agent uses. Include retries as well as successful runs.
- Review: the time spent on its work. Reading drafts, checking facts, making corrections and deciding what is ready to use.
- Ownership: the time spent keeping it useful. Maintaining access, testing changes and adjusting the workflow when the job changes. Keep these hours separate from reviewing individual tasks.
For a worked example, imagine 100 proposal drafts in a month, all checked and accepted for use. Each gets 15 minutes of review. We have chosen the other rates below to make the calculation easy to follow; they are illustrative assumptions, not MING’s actual operating costs or a service quote.[S1]
| Cost line | Assumed inputs | Monthly cost |
|---|---|---|
| Runtime | 100 attempts × €2 + €120 fixed | €320 |
| Review | 100 attempts × 15 min ÷ 60 × €50/h | €1,250 |
| Ownership | 4 hours × €80/h | €320 |
| Total | 100 accepted tasks × €18.90 | €1,890 |
In this example, the software costs €320 and the people around it cost €1,570. Together, that is €1,890 a month, or €18.90 for each accepted draft.[S1] Changing the model price would affect only one part of that budget.
03 A rejected draft still costs money
A draft that gets rejected still takes resources to produce and time to check. Suppose the same 100 attempts yield only 80 results your team accepts, with no change in spending or review time. The cost per accepted result rises from €18.90 to €23.63.[S1]
| Accepted tasks | Monthly spend | Cost per accepted task |
|---|---|---|
| 100 of 100 | €1,890 | €18.90 |
| 80 of 100 | €1,890 | €23.63 |
| 0 of 100 | €1,890 | No accepted outcome |
Define what “accepted” means before the trial. For a proposal draft, it could mean that the account lead has checked it and approved it for the intended next step. Counting drafts alone would miss that distinction. Our guide to when to shut down an AI agent looks at what happens when a team gets plenty of output but uses none of it.
04 Measure a real workflow before you scale
Pick one recurring task and follow it from request to accepted result. Start with a week that includes ordinary work and awkward cases, then repeat if the workload varies. A few easy drafts will tell you less about the budget than a period with corrections, retries and handovers.
- Record every attempt. Keep the software costs for rejected work and retries in the total.
- Time the human work. Track review for each task and shared maintenance separately, so the same hour is counted once.
- Mark what was accepted. Use a named reviewer and the quality standard you would expect from a human doing the same task.
- Compare equivalent work. Work out the cost of completing the same job manually, then record what the team does with any time released.
An hour freed up can help someone clear a backlog or spend more time with a client. It does not automatically reduce the payroll. Show that extra capacity separately from expenses you actually reduce or avoid. Keep setup and integration costs visible too, even when you spread them across several months.
05 If review eats the gains, revisit the role
When checking an agent’s work takes almost as long as doing the job, examine the assignment. Is it too broad? Does the reviewer have a clear standard for accepting the result? Has the agent been given work that requires a person’s judgment?
The ABC Framework helps separate human judgment, shared expert work and routine operations. Use that distinction to narrow the role and make the handover clear. A named human owner still decides whether the result is good enough.
Budget for the people around the agent, then check whether the work is worth that time.
A simpler technical approach can help too. Anthropic’s guidance on building agents recommends adding complexity when the performance gain justifies the cost and latency, with limits and human checkpoints where needed.[S3] The right change depends on what is taking the time: running the software, checking the output or maintaining the workflow.
06 Start with one role and its full cost
The €1.27 proposal figure is useful once the human work sits beside it. For your own agent, start with one task, a clear definition of a usable result and a person who owns the outcome. That gives you a budget you can test before adding more agents.
MING Labs’ Hybrid Organisation work helps teams design those roles and the systems around them. Bring a recurring task, its current workload and the time your team spends checking results. Talk to us about the first role you want to put through this exercise .
About the numbers
- The worked example: one illustrative month, dated 1 October 2026, version 1.0. We assume 100 attempts, 15 minutes of review each, €2 runtime per attempt, €120 fixed runtime, €50 per review hour and four ownership hours at €80 per hour. The base case accepts all 100 tasks.[S1]
- What the example leaves out: initial implementation, integration, tax, exceptional incidents and business losses from errors. Failed attempts stay in the cost total. With zero accepted results, cost per accepted result is undefined.
- Sebastian’s report: he describes €1.27 compute and 15–25 minutes of review as measured, but does not publish the sample size, measurement dates or underlying logs. His four operator hours per agent per week are an estimate; the model above separately assumes four hours per month.[S2]
- Reproduce the example: open the inputs, formulas and results , then replace the assumptions with your own observations.
