What does an AI integration really cost? An honest budget range
"What will this cost us?" is the legitimate first question of any decision-maker, and the most honest first answer is unfortunately "it depends". That is true, because an AI integration is not a product with a list price but a project with many levers. It just helps no one with budget planning.
This guide makes the answer concrete. It breaks the cost into five blocks, names a solid ballpark figure for each from our projects, shows the items that tend to be missing from quotes, and closes with the one question that really decides the budget: what is the value worth to you, and how do you measure it.
The five cost blocks at a glance
Anyone comparing quotes is often comparing apples with oranges, because one quote contains only the development and the other also includes operation and maintenance. A shared grid helps. An AI integration almost always consists of the same five blocks, either one-off or ongoing.
| Cost block | What is behind it | Type |
|---|---|---|
| Project / development | Concept, integration, interfaces, testing, acceptance. The actual integration work. | one-off |
| Data work | Preparing, structuring, cleaning and making data searchable. Often the biggest hidden item. | one-off, partly ongoing |
| Inference / model usage | The cost per request to the AI model, usually billed by token or usage. | ongoing |
| Maintenance / development | Adjusting prompts, switching models, fixing errors, adapting to new requirements. | ongoing |
| Operation / infrastructure | Hosting, monitoring, security, data-protection documentation, licences where applicable. | ongoing |
The two blocks on the right, maintenance and operation, are regularly underestimated in the first enthusiasm over a pilot. An AI integration is not a piece of furniture that is set up once. It is more like an employee who has to be trained, supported and kept up to date.
Blocks 1 and 2: project and data work, the expensive start
The largest one-off costs arise at the beginning, and the larger part of them often lies not in the AI itself. Connecting a language model to an application is comparatively well-understood craft today. How such a model actually works is explained in our knowledge article on large language models.
The data work is almost always the more expensive part. If the AI is to work with your company knowledge, meaning product data, documentation or support knowledge, that data first has to be prepared: structured, cleaned, made searchable. This step is often called RAG in technical jargon and is regularly the real project behind the AI project. A drive full of scanned PDFs is not a usable data foundation but a preliminary undertaking with its own budget.
As a ballpark from our projects: a first, clearly defined use case, such as an AI-supported search across your own knowledge or an assistant for a recurring task, sits as a one-off investment usually in the low to mid five-figure range. If the data is already clean and structured, the figure drops. If it first has to be put in order, the data work alone can make up the larger part.
Blocks 3 to 5: running costs, often smaller than expected
For the running costs there is good news: the pure model usage, the inference, has become surprisingly cheap for most mid-sized use cases. The price per request has fallen sharply over the last few years. For an internal assistant with a manageable request volume, the model costs often sit in the two- to low three-figure range per month, not infrequently less.
The running costs that do weigh in are human, not machine. Someone has to maintain the solution: sharpen prompts, check with new model versions whether a switch is worthwhile, fix errors, keep the data foundation current. Plan for a fixed, if small, share of a position for this, not zero. An AI solution that no one maintains loses quality within months, because models, data and requirements keep moving.
For operation, the range depends heavily on whether you rent models as a service through an interface or run them yourself. For almost all mid-sized companies, renting is the economically right choice. When running them yourself is worthwhile after all, and how to decide this fundamental question cleanly, is covered in our article on make or buy for AI solutions.
The hidden items that rarely appear in the quote
Three kinds of cost rarely appear in quotes but decide both success and total budget. We name them openly, because concealing them helps no one.
Internal adoption. An AI solution only takes effect once people use it. Training, support and the time during which the team is slower because it is learning something new appear in no development quote, but they cost real money. This time is well invested, but it has to be planned in.
Ongoing data maintenance. Data ages. Products change, prices too, documentation grows. Without continuous maintenance, the AI will at some point answer on the basis of yesterday. This item is ongoing, not one-off.
Governance. Data-protection documentation, a rule for what the AI is explicitly not allowed to do, and some form of quality control belong to a serious integration. The effort is manageable, but it is not zero, and it turns out more expensive if it comes after the fact.
The only question that really decides: what is it worth to you?
In the end the budget question can only be answered sensibly from the other side. Not "what does the AI cost" but "what is the task worth that it is meant to take over". An integration that saves two hours of qualified work per day pays off at almost any realistic project price. One that only delivers a good feeling pays off at none.
That is why measurement comes before the quote. For the chosen use case, establish a baseline: how much time, how many errors, how much throughput time does the task cost today? Only this figure turns every quote into a decision instead of a matter of taste. How to select such a use case, pilot it with an owner and a baseline measurement and assess it honestly after 90 days is described in our article on AI strategy for mid-sized companies.
Is there a rule of thumb for the total budget?
A rough orientation from our projects: alongside the one-off investment, expect running costs of around 15 to 25 percent of the project amount per year for maintenance, operation and data work combined. That is a ballpark, not a fixed price, and it varies strongly with the use case. More important than the exact figure is to plan for these running costs from the start, instead of looking only at the project price.
Why is the data work so expensive?
Because it makes up for what has been left lying for years. For an AI to work with your company knowledge, the data has to be structured, consistent and searchable. In many companies, however, it sits scattered, in different formats and partly several times over with different truths. This clean-up work is a prerequisite, not an accessory, and it pays off beyond the AI project, because clean data is useful elsewhere too.
Can we start small to limit the budget?
Yes, and that is exactly what we recommend. A single, clearly defined use case with measurable value limits the risk and provides a solid basis for deciding on further steps. The common mistake is not starting too small but starting too broad: ten ideas half-heartedly instead of one properly. With a first success in hand, the next budget is far easier to justify.
Are expensive providers automatically better?
No. The price says something above all about the scope and the included support, less about the quality of the AI itself, since the underlying models are the same for most providers. Pay less attention to the daily rate and more to whether the quote names the hidden items honestly: data work, adoption, maintenance. A cheap quote that conceals these blocks becomes more expensive in operation than a dearer one that plans for them.
The first step to a solid figure
Before you request quotes, create the one figure that makes every quote assessable. Choose a single recurring process that costs a lot of time today, and measure for one week how much time it actually takes up. This ballpark is your benchmark for every cost estimate.
If you would like to make this assessment together without starting a project right away, talk to us. In a Future Check we sort out your use case, the realistic range and the expected value, in clear figures rather than promises. You can arrange the appointment through our contact form.
Go deeper in our knowledge base
Want to know what these topics mean for your company? The Future Check shows you the biggest levers within 2–4 weeks.
