Five cost blocks that belong in every AI project
An AI project that takes over a workflow in your business always consists of the same building blocks, no matter who builds it. When you compare offers, check whether all five appear or whether one quietly ends up with you:
- Analysis: which workflows pay off, what the build costs, which data flows where.
- Build: define the rules, build in your systems, trial run, parallel operation, handover.
- Operation: monitoring, adapting to new models, system updates and changed rules.
- Model costs: the model provider's fees per volume of text processed, or the computing power for a model on your own servers.
- Internal effort: your employees' time for conversations, rules, checks and approvals.
The first three are services, the fourth is consumption, the fifth is your own working time. The last two in particular are missing from many offers, and they often decide whether a project stays within budget.
What we charge for analysis, build and ongoing operation
We publish our prices because you should know what you are committing to before the first call. All amounts are net, plus VAT:
| Stage | Price | What is included, what it depends on |
|---|---|---|
| AI workflow analysis | €4,900 fixed price | Three days on site or remotely. A list of your workflows with today's effort, tomorrow's saving and a fixed price per workflow, a "does not pay off" column, a data flow sketch. |
| Build per workflow | €12,000 to €40,000 fixed price | Stated as a binding figure in the analysis. Rulebook, build in your systems, trial run, parallel operation, introduction, documentation. Four to six weeks per workflow, longer for extensive workflows. |
| Custom tool | Fixed price after the analysis, no range | A dedicated user interface with roles and permissions is the biggest price driver, so we only give the price once the scope is clear. |
| Ongoing operation | from €2,900 a month | Monitoring, updates when models, systems and rules change, response on the next working day, allowance for extensions. Rises with the number and scope of workflows. Cancellable monthly. |
Why the range for the build is so wide: a workflow between a mailbox and a CRM is simpler than one that connects ERP, PIM and shop. Legal texts governed by a directive need more checks than category texts. Delivering results as a file is cheaper than writing them back into a system that must not fall over in the process. Volume itself, on the other hand, hardly drives the build: whether 1,000 or 100,000 records, the workflow is the same. What rises with volume are checking, run time and model costs. All price factors in detail are on the Approach and prices page.
A typical first year with us, with an analysis, two rebuilt workflows and ongoing operation, comes to between €60,000 and €100,000 net. Code, rules, prompts and documentation belong to you after payment. There is no licence and no platform fee that keeps running afterwards.
Model costs: what a language model costs per run
Language models from providers such as Anthropic, OpenAI, Google or Mistral are billed by token, separately for input and output, at a price per million tokens. A token is part of a word; for German text you can reckon on roughly one and a half to two tokens per word, for English text somewhat fewer. Input includes everything the model has to read: the document, the rules, the glossary, examples. Output is what it writes. Output tokens are considerably more expensive than input tokens.
Prices change frequently, usually downwards, and vary greatly between model classes. As an order of magnitude, and expressly as our estimate, as of October 2026: large models cost roughly €1 to €5 per million input tokens and €5 to €25 per million output tokens. Small models cost a tenth of that or less. Many providers give a substantial discount for jobs that do not have to be finished immediately. Check the provider's current price list before any decision.
Worked example: 5,000 product texts from data sheets
Suppose a workflow produces 5,000 product texts from manufacturers' data sheets and checks each text against rules in a second step. The calculation is an estimate based on assumptions, as of October 2026, not an offer:
| Item | Assumption | Volume |
|---|---|---|
| Generating, input | Data sheet extract, rules, glossary, example: 4,000 tokens per text | 20 million tokens |
| Generating, output | Text of around 450 words: 700 tokens per text | 3.5 million tokens |
| Checking, input | Text and check rules: 2,500 tokens per text | 12.5 million tokens |
| Checking, output | Verdict with reasons: 150 tokens per text | 0.75 million tokens |
| Reruns | 15 percent of texts run a second time | Added to all items |
Together that is around 37 million input tokens and around 4.9 million output tokens. We assume a large model at €3 per million input tokens and €15 per million output tokens. That gives about €112 plus €73, a total of around €190 for the entire run, so under 4 cents per text. If the check runs on a smaller model, the amount falls further. Workflows with web research or very long documents use many times that. For texts at scale, in our experience the model costs per run are usually in the range of tens to hundreds of euros.
A second example for a continuously running workflow: reading 60 incoming enquiries a day, classifying them and assigning them to the right team. At around 1,300 enquiries a month, 3,000 tokens of input and 200 tokens of output per enquiry, the same assumptions give about €16 a month, again an estimate.
Model costs are therefore almost always the smallest item. We do not pass them on with a mark-up: they run through your own account with the provider or are billed transparently by usage. If an open model runs on your own servers, the per-token fees disappear; instead you pay for set-up and computing power. That pays off above all when the data must not leave the building, less often for cost reasons.
The internal effort that no offer mentions
Your employees are involved in every project, and rightly so: they know the rules the workflow has to follow. Plan for this time:
- Analysis: the people who do the work, one to two hours each for a conversation and for letting us watch. IT half a day for a look at the systems.
- Rulebook and access: a contact from the business department defines with us what the workflow may do and what it puts forward for approval. IT grants access.
- Trial run: the business department checks a sample point by point. This is the most demanding part for you and the most important, because every deviation becomes a rule.
- Parallel operation and handover: comparing results, briefing on documentation and rules.
You do not need a project team, but you do need a contact with decision-making authority. Reckon on a few hours a week for that person during the build.
Two further costs sometimes only appear during the project. Some systems release interfaces only in higher licence tiers or limit the number of requests. At a spare parts dealer, for instance, Salesforce limits the number of requests per month; rebuilding the caches reduced them by 71 percent so that usage stays within the allowance. We clarify such limits in the analysis so that they do not turn into a change request later.
Does it pay? A worked example
The analysis answers the question of whether a project pays with figures from your business. This example with assumed figures shows how the calculation is built. It is not a promise but a method:
- Workflow A, classifying enquiries: 60 enquiries a day at 5 minutes each, on 230 working days, is around 1,150 hours a year.
- Workflow B, maintaining product data: 200 records a week at 6 minutes each, over 46 weeks, is around 920 hours a year.
- Remaining effort: a quarter of the work stays with people, for checks and unclear cases. That leaves around 1,550 hours saved.
- Valuation: at €45 full cost per hour, that is around €70,000 a year.
- One-off costs: analysis €4,900, two builds at an assumed €22,000 and €30,000, together €56,900.
- Running costs: ongoing operation at an assumed €2,900 a month, around €34,800 a year, plus model costs of a few hundred euros.
In ongoing operation, this example leaves around €35,000 a year, and the one-off costs are recovered after about 20 months. Ongoing operation is fully included; anyone who runs the workflow with their own IT after handover saves that. Not included is what the freed-up time is worth when skilled staff do work that is currently left undone, or when enquiries are answered the same day instead of on Monday.
For some workflows the calculation is even clearer, because the alternative is a one-off major task. At an energy technology manufacturer with around 400 employees, checking several thousand pages and data sheets for environmental claims by hand would have cost three to four person-months. How the workflow solved this is described in the case study on the energy technology manufacturer. And which workflows qualify for such a calculation in the first place is covered in the article Which workflows pay off with AI.
Eight questions to ask of any offer
These questions make different offers comparable, regardless of who they come from:
- Is the price for the build a fixed price or an estimate based on effort? What happens if it takes longer?
- What does operation cost after the project, and when can it be cancelled?
- Are model costs passed on with a mark-up, and do they run through your own account?
- Are there fees per user, per transaction or for a platform that keep running after the project?
- Who owns the code, rules, prompts and documentation after payment?
- What does it cost to switch the language model when a better or cheaper one appears?
- Are the trial run and parallel operation included in the price, or does billing start again afterwards?
- How much internal time is expected from your employees, and in which weeks?
An offer that gives a clear answer to all eight questions is comparable. One that dodges three of them will end up costing more than expected.
What this means for you
An AI project at a mid-sized company does not have to be an open-ended budget. The service can be set at a fixed price, the model costs for texts at scale are usually tens to hundreds of euros per run, and the biggest unknown item is your own time, which can be planned. Whether a workflow pays is decided by setting volume times time from your own business against the costs. If you currently pay per head for standard software for a workflow and still work alongside it in Excel, that side work belongs in the calculation.
Your next step: in the cost calculator, set your licence costs and the side work in Excel against a custom tool, over five years, with your own figures.