What does an AI agent cost? Budget per usable outcome

An AI agent makes economic sense for your business when its assignments produce usable work and the total effort is appropriate. A low usage cost per call does not answer that question by itself. You also need to include setup, human review and rework. The useful budgeting question is therefore: what does an outcome cost when our team can actually use it to move the work forward?

This article develops a calculation template for a small pilot. Every amount, duration, quantity and result in the example is explicitly hypothetical. These are neither webRichtung prices nor promised performance figures. Replace the assumptions with the terms of your specific arrangement and your own observations. The focus is the decision before implementation, rather than a general comparison of licensing and usage-based payment models.

Define the work you want to pay for first

Choose an assignment whose result a responsible person can assess. “Support our office” is too vague as a cost unit. In the example, the agent should create an internal status brief for a supplied customer case. It describes the supported facts, open questions and the person responsible for the next action. Sending it to customers, making new commitments and changing an order sit outside this pilot’s scope. That is a chosen working boundary, not a claim about a preinstalled standard process.

Define when the brief is usable. The correct customer case must be identifiable, statements must agree with the supplied documents, and missing information must remain visible as open questions. The next colleague should see which task has already been accepted and where a decision is still required. An elegant text attached to the wrong customer does not meet those conditions. Several versions of the same brief still represent only one possible work outcome.

Name the recipient before the first attempt. In our example, a team coordinator checks whether the brief lets her prepare the next internal action. Her review belongs in the calculation. Merely confirming that some text was produced is insufficient for acceptance. If she must reconstruct the whole case herself, that is evidence about the assignment or inputs, rather than a free extra to be added invisibly to the agent’s work.

webRichtung agent supports assignments using your data, together with rules and tools you have authorised. Explore agent and choose a specific, assessable use case for your pilot. Bring the intended result to the module page. You can then start with a bounded piece of work and subsequently judge its practical value for yourself.

Separate setup, ongoing usage and maintenance

Setup includes the work that makes the pilot usable in the first place. You describe the assignment, clarify necessary sources, determine allowed actions and check suitable test cases. Include internal coordination where it is actually needed for this use case. Record that time separately. It does not disappear economically just because nobody sends an external invoice. Equally, charging its full value again in every later round would be misleading if the setup remains reusable.

Ongoing usage covers the consumption actually incurred for the work being assessed. Establish the applicable billing unit and scope from the concrete arrangement or visible usage information. One business assignment may require several technical steps. Do not silently treat one call as equivalent to a finished customer brief. Also check whether any additional costs for the required tools are relevant before treating your own calculation as complete and ready for a business decision.

Recurring human work includes review, correction and maintenance. Review decides whether an outcome meets the agreed conditions. Rework addresses a specific defect. Maintenance keeps sources, rules and the process usable. Separate these activities so you can identify what needs improvement. If much of the time is spent finding missing documents, changing one sentence in the agent’s instruction may help less than creating a clearer way for the required information to reach the task.

Avoid counting time twice. If a recorded review already includes a correction, those same minutes cannot also appear as additional rework. An internal hourly rate is a valuation assumption, not automatically a cash payment or salary that can immediately be saved. Explain whether you are considering direct expenses, valued working time or both. Our template combines hypothetical direct usage costs with working time valued at one consistent internal rate.

Count failed attempts and approvals appropriately

Keep submitted assignments, execution attempts and usable outcomes separate. When an assignment is repeated after a correction, another attempt occurs, but a second customer case does not. An interrupted run may have incurred effort without delivering a usable result. You must establish whether and how individual steps are billed for your specific use case. The economic calculation captures total actual effort instead of assuming that an unsuccessful status automatically means work was free.

Classify rejected outcomes by cause. A brief containing an unsupported promise is a different finding from one missing necessary input information. In the first case, the output boundary needs better enforcement; in the second, someone must obtain the source or narrow the assignment. Record the cause clearly enough to change the next attempt deliberately. Ten unfocused repetitions otherwise offer mainly ten more opportunities to pay for or review the same mistake again.

Handle approval time just as concretely. If the responsible person actively reviews for six minutes, record six minutes of work. If the case then waits two days for their decision, that is additional elapsed time, not automatically two days of active labour. Both measures are useful but answer different questions. Record waiting as a cause of delayed usability and assess its real consequences rather than multiplying it by an hourly rate without justification.

With agent, approvals are set for risky steps, while routine work can continue within your rules. Our complete human review of results is a deliberately chosen pilot method. It does not imply that every subsequent routine case must permanently receive an individual manual confirmation. Whether you can later adjust review depends on the specific assignment, observed findings and applicable working rules.

Calculate the pilot through to usable outcomes

For hypothetical setup, assume six hours at an internal valuation rate of €50 per hour. That gives a one-off €300. The pilot covers 40 distinct customer cases. Ten additional execution attempts bring the total to 50 attempts. For all usage in this round, we assume a total of €100 solely for the calculation. It is not a price per call or an offer for completing an agent assignment.

The team coordinator reviews all 40 cases for six minutes each. That is 240 minutes, or four hours; at €50 per hour, the value is €200. Add two separately recorded hours of rework, worth €100. A further hour maintains rules and sources, worth €50. These additional activities are not already included in the six minutes of review in our example. The arithmetic therefore remains traceable without counting the same working time twice.

Under these assumptions, the ongoing round costs €100 usage plus €200 review plus €100 rework plus €50 maintenance. Together that is €450. Including one-off setup, the first round costs €750. Of the 40 submitted customer cases, 32 produce a brief that is usable under the criteria set beforehand. Eight remain unusable or unresolved at the end of the measurement window. Their effort so far is included; completing them later has not yet been costed.

For the ongoing round, €450 divided by 32 usable briefs gives approximately €14.06 per outcome. In the first round including setup, €750 divided by 32 gives approximately €23.44. Dividing by 50 attempts instead would produce a different measure describing cost per attempt. It would not answer what the team coordinator had to spend, including valued effort, to obtain a customer brief she could actually use.

The eight open cases must neither disappear nor count as finished work. Record which are still needed and how they will be completed. Our outcome measure already carries their pilot effort to date. A full budget for processing all 40 cases would need to add the later completion of those eight. Claiming that the entire workload was finished for €450 would therefore go beyond what this example demonstrates.

Compare the same usable work

As a second explicitly hypothetical assumption, creating an equivalent brief entirely manually takes 20 minutes. For 32 outcomes, that is 640 minutes, or ten hours and 40 minutes. At the same internal rate of €50, the value is approximately €533.33. This comparison must match the scope and quality of the 32 accepted results. A short, unchecked note would not be an appropriate counterpart to a supported brief with a clear next owner.

Compared with €533.33, our ongoing round at €450 is approximately €83.33 lower. The first round including setup, at €750, is approximately €216.67 higher. This values the 32 outcomes delivered in the example; it does not prove total savings for all 40 incoming cases. Further work on the eight open cases remains unknown and must be added before producing a complete calculation for the whole workload.

The time benefit needs its own explanation. Review, rework and maintenance add up to seven human hours in the ongoing round. Compared with ten hours and 40 minutes of manual reference work, that represents three hours and 40 minutes less committed time. Its valuation is approximately €183.33; subtracting the assumed €100 usage leaves the €83.33 already calculated. Released time becomes useful to the business when the team can actually apply it to other worthwhile work.

Test how sensitive the decision is. With unchanged effort of €450 but only 24 usable outcomes, each costs €18.75. Manual comparison work for 24 briefs would take eight hours and be valued at €400, making the round €50 more expensive. If 32 outcomes remain usable but assumed usage rises from €100 to €200, the round costs €550, or approximately €17.19 per outcome. That too exceeds the approximately €16.67 value of one manually produced brief.

These variations test the calculation; they do not predict failure rates or consumption. They reveal which observations matter: the number of genuinely usable outcomes, additional human work and actual usage. A single attempt cannot establish a reliable payback period for setup. Such planning only becomes meaningful once the process is repeatable with a comparable mix of assignments, rather than inferred from one favourable or unfavourable small sample.

Bound the trial and respond to deviations

Set a workload and a budget you can monitor before the pilot begins. In our example, 40 cases are the planned volume; €100 usage is a budgeting assumption, not a claimed automatic product spending stop. Determine who monitors actual consumption during the trial and when new assignments pause. You can then investigate a deviation before handing over more work under the same unverified pattern and increasing effort without learning what happened.

Prepare a specific response to errors. If a brief mixes up the customer reference, it remains rejected and does not pass unchecked to the next colleague. The responsible person checks the input and assignment description, corrects the identified cause and decides whether to try again. Record the additional time involved. A repeated error is a reason to narrow the scope or pause the trial rather than simply increase the number of further runs.

agent lets you stop assignments, deactivate automations and change rules. That control suits a pilot in which decisions deliberately remain with you. Describe in advance which routine work may continue and which finding requires a question. An economically useful agent needs an understandable assignment and usable sources. Raising the budget alone does not replace those foundations or resolve a missing rule about how the intended work should be completed.

Bring the decision template to your start

After the measurement window, you should be able to explain the decision in complete sentences. One possible finding from our example is: the 32 usable outcomes, excluding setup, cost less than the assumed manual comparison valuation; the first round has not recovered setup. Eight cases need separate clarification. The next action is therefore to address their causes and plan another comparable round before substantially increasing scope. This is a conditional decision under stated assumptions, not a purchase recommendation based on invented performance data.

Transfer the concrete assignment, acceptance criteria, one-off setup, actual usage and separately recorded human times to your own template. Add usable outcomes, open cases and the appropriate manual comparison. You do not need an apparently precise annual forecast from one small first trial. A traceable pilot finding explains more clearly which next step is justified and which uncertainty still needs resolution before you commit more work or funding to the process.

Explore the right agent use case and create your account through the module page. Begin with work whose intended outcome and boundaries you can name. That gives you a sound basis for your own budget: the effort required to produce work your team can actually use, alongside the costs incurred by the technical attempts that make up the process.