The most common question I get about AI isn't "what can it do". It's "what will it cost us and when does it pay for itself". And there's a problem with it: while price lists for websites, e-shops or training are a few minutes of searching away, prices for putting AI into business processes are barely published anywhere. The result is that a company gets a quote for €1,600 with no reference point at all for whether that's good value or twice what it should be. This article provides that reference point: three layers of cost, my own prices, three model budgets and a way to calculate payback before you sign anything.
Costs have three layers: implementation, ongoing operation and people's review/administration time. My offer: audit €1,000, pilot from €2,400, 3-5-process program from €6,000, optional oversight from €320/month. Audit deduction follows implementation terms. Calculate payback from net benefit after recurring costs; freed time is not automatically cash saved.
Why you can't find the prices anywhere
There's no conspiracy behind it, just three practical reasons.
Jobs vary wildly in size. "Implementing AI" can mean a week of work on one form or a six-month project wired into a warehouse system. Vendors worry that a published price either scares off small clients or undercuts the large ones.
Compare equivalent scope. Separate licences from integration, verification and support. Different amounts may represent different commitments; inspect the actual proposal to find out why.
Plenty of quotes rely on vagueness. When a quote doesn't say what exactly will be working at the end, the price can be built on the impression left by a presentation rather than on scope. That's exactly how you pay for a workshop that produces a PDF instead of a working process.
I go the other way: prices are on the service page, and every quote states what specifically will be working when it's done and how we'll know.
Layer 1: the one-off deployment
Break one-off work into process analysis, implementation/integration, tests on permitted representative data and training/documentation. Their shares depend on scope; universal percentages are not justified without a work estimate.
A quote missing real-data testing and training isn't a cheaper quote. It's an unfinished one.
Layer 2: the running costs everyone forgets
Running costs are a smaller number, but they arrive every month.
- Tool licences. Include integration subscriptions and assistant accounts. Self-hosted n8n has licensing conditions; hosting it does not make every edition free or prevent connected cloud services processing data.
- Model usage. Cost depends on model, inputs, outputs, tools and retries. A single enquiry cannot universally be priced at fractions of a cent without measuring the configuration.
- Oversight and changes. Processes change: a new type of inquiry shows up, a supplier changes their invoice format, the model gets a new version. Either someone on your side watches it or there's a retainer for it.
Estimate unit consumption on representative cases, multiply by expected volume and add licences, contingency and oversight. Label this an estimate, not a guaranteed invoice.
Worked example: An AI task also has running costs
ORIGINAL EXAMPLE · FICTIONAL DATA
| Item | Unit | What to record |
|---|---|---|
| Enquiry summaries | Runs | Successful runs and retries |
| Model API | Input / output | Actual billed usage |
| Integration operations | Month | Hosting, monitoring, licences |
| Human review | Time | Corrections and approvals |
An assistant subscription is not automatically API credit.
Layer 3: your people's time
Include internal work in the budget too. Measure or estimate its volume separately. Somebody has to:
- supply samples of real cases and say what the correct result is,
- check outputs for the first two to four weeks and report what's wrong,
- change their habits, which, for people who've done the job for ten years, is the hardest part.
Budget a few hours a week from one person for the duration of the pilot. If a company isn't willing to give that time, the implementation shouldn't start; it's one of the few reasons I turn work down.
Worked example: Who contributes time to an AI pilot?
ORIGINAL EXAMPLE · FICTIONAL DATA
| Work | Owner | Concrete output |
|---|---|---|
| Inputs | Sales | Anonymised enquiries and correct categories |
| Trial | Integration owner | Summary draft without sending |
| Review | Salesperson | Incorrect details and corrected reply |
| Maintenance | Named owner | Change record and repeat test |
Record the actual time; this plan does not invent hours saved.
My prices
I separate three steps, each with a stated starting price and its own deliverable, so you can decide whether to continue:
| Step | Price | What you get |
|---|---|---|
| Operations audit | €1,000 | 2 weeks, mapping the agenda with the people who do the work, processes ranked by saving / risk / effort, a time and cost estimate per process, a 6-month plan. Deducted from the implementation if you go ahead within 3 months. |
| Pilot of one process | from €2,400 | 3-6 weeks, baseline measurement, deployment into live operations, integration with your tools, training, documentation, post-deployment measurement and 30 days of tuning. |
| Program of 3-5 processes | from €6,000 | 3-4 months, processes deployed in priority order, internal AI usage rules, a local-model variant when data must not leave the building. |
| Oversight and changes | from €320 / month | Optional. Monitoring, changes when things shift, small extensions. Not a condition for the deployed processes to keep working. |
For companies up to roughly ten people, a full audit usually isn't worth paying for, if the opening call makes it clear where to start, we go straight to the pilot and you save the €1,000. Details and what's included are on the AI automation service page.
Three model budgets
The figures below are scope estimates, not price-list items, the real price always depends on the company, the volume and the state of the systems.
Processing incoming enquiries. Test classification, extraction and draft replies. Inbox/CRM access, volume and review workload affect the budget. A pilot establishes payback assumptions.
Documents and data entry. Formats, input quality, integrations and verification of writes drive costs. Hundreds of documents alone do not establish viability.
Recurring reports and research. Cost depends on source access, calculations and fact-checking. Reporting may need complex integrations; it is not automatically the cheapest option.
An assistant that answers from internal policies, contracts and manuals and cites the source for every answer is budgeted the same way. What decides the cost there is document preparation and who may see which document; I describe the process in AI over company documents.
How to calculate payback before you sign
You don't need a twenty-row spreadsheet. Four numbers are enough:
- How many hours a week the activity takes today. Not a desk estimate, log it for a week.
- Time actually removed, after review, correction and monitoring. Establish it with a pilot, not a universal ratio.
- Your real hourly cost for that role, i.e. fully loaded, not net salary.
- Deployment plus running costs for the first year.
Annual capacity value = net hours saved per week × actual working weeks × hourly labour cost. Subtract recurring costs not already included. Simple payback = initial investment / positive annual net benefit. Compare the result with risk and company priorities; twelve months is not a universal cutoff.
What to watch for in a quote
- Percentage savings promised upfront. "You'll save 40% of the time", before anyone measured what the process costs today, is marketing. The honest version is measuring before and after with the same method.
- Missing running costs. If the quote has no estimate for licences and model usage, you'll learn it from the invoice.
- Accounts in the vendor's name. Access and accounts belong to your company. Otherwise you're also buying a dependency you can't see.
- No documentation or training. Missing access and documentation make handover harder. Have the handover verified through a concrete task.
- Vendor commissions. Ask about them and compare alternatives. Commission is a potential conflict, not proof of an unsuitable recommendation.
- Deployment without rules. Budget training and assessment of obligations based on actual use and the company's role. The AI Act guide distinguishes duties; not every AI output falls under the same rules.
When I don't recommend it
To be fair to the other side, there are situations where the answer is "not yet":
- The process changes every month and nobody can describe what a correct result looks like.
- The task takes little time and the pilot identifies no other benefit that justifies total cost.
- The data sits in five tools that don't talk to each other and nobody wants to tidy it up. Then the first step is data cleanup, not AI.
- There's nobody to watch the outputs for the first month.
- Decisions about people, such as hiring or employee assessment, need separate legal and risk assessment first; this article does not recommend them as an initial pilot.
Summary
- Costs come in three layers: one-off deployment, monthly running costs (licences + model usage) and your people's time for checking. A quote that only mentions the first layer is incomplete.
- My prices: operations audit €1,000 (deducted from implementation), pilot of one process from €2,400, program of 3-5 processes from €6,000, optional oversight from €320 a month.
- Payback uses net benefit after recurring costs; freed capacity is not automatically cash savings.
- Percentages promised before measurement are marketing. Insist on before-and-after measurement using the same method.
- Accounts, documentation and training belong to you. Without them you're buying vendor dependency.
Frequently asked questions
How much does an AI audit of business processes cost?
With me it's €1,000 for two weeks of work, with a deliverable you can act on even without involving me further: processes ranked by saving/risk/effort, a time and cost estimate for each, and a six-month plan. I deduct the audit fee from the implementation if you go ahead within three months. For companies up to roughly ten people I often recommend skipping the audit and going straight to a pilot.
What are the monthly running costs of AI automation?
Include licences, measured API use, hosting, review and maintenance. The result depends on volume and configuration; a monthly range cannot be promised without them.
How long until AI implementation pays for itself?
Divide initial investment by positive annual net benefit. Include operation, correction work and capacity actually used. This article guarantees no payback period; the company decides what fits its priorities.
Is AI worth it for a company under ten people?
It can be worthwhile if the specific process creates measurable value. Employee count alone does not determine payback. An initial discussion can establish whether a bounded pilot or broader audit is needed.
What's the difference between AI training and AI implementation?
AI training teaches your team to work with AI themselves, prompts, workflows, safety. Implementation is done-for-you: I build the process, wire it in and hand it over working. Most companies need both, which is why training the affected people is part of every deployment. If you want to try AI on a small budget first, start with training.
Do we need a custom application for this?
Usually not. Ordinary operational routine is handled by integration tools and models working over data you already have. A custom application becomes relevant once the process hits a ceiling: proprietary data, complex rules, high volumes or operations the company depends on. I describe the path from the simplest solution to custom code in the guide to your first automations.
Sources and links
- Regulation (EU) 2024/1689 (AI Act), the AI literacy duty (Art. 4) and transparency rules (Art. 50)
- OpenAI API pricing and Anthropic pricing, per-volume pricing behind model usage costs
- n8n pricing and Make pricing, licence costs for integration tools
- GDPR Regulation (EU) 2016/679, personal data processing and the processor role