What does AI cost a company? A realistic overview for SMEs
5 min read
Short answer
The cost of AI has three parts: licences for ready-made tools, usually per user and month, one-off costs for setup and connection to your systems, and ongoing costs for usage, hosting and maintenance. A prototype for a clearly defined use case is often possible for a mid four-figure amount in euros.
What does this mean for your company? Book an intro call
Ask about the cost of AI and you often hear one of two extremes: “It costs almost nothing, ChatGPT already exists” or a six-figure strategy project. For most small and mid-sized companies the truth lies in between, and it can be narrowed down fairly precisely.
What costs make up an AI project?
- Licences for ready-made tools: AI assistants such as ChatGPT, Microsoft Copilot or European alternatives are usually billed per user and month. They suit general tasks such as writing, summarising and research.
- One-off implementation: if AI is to be built into your processes, for example reading invoices and passing them to accounting, development is needed. That includes connecting your systems, testing with real data and onboarding.
- Ongoing operation: usage costs for the AI model, hosting, monitoring and maintenance. Models evolve, interfaces change, and occasionally things need fine-tuning.
What does using an AI model cost?
Most providers bill by usage, measured in so-called tokens, small pieces of text. For typical SME tasks such as reading a document or classifying an email, costs are often in the range of a fraction of a cent to a few cents per case. With a few thousand cases a month, that usually adds up to double-digit or low three-digit amounts.
These costs are almost never the problem. Implementation and maintenance weigh far more.
Which costs are often forgotten?
- Data preparation: documents are scattered, master data is outdated. This needs sorting out before the start.
- Your team's time: for conversations, tests and feedback. Plan for it, it is time well invested.
- Review effort: if a person checks results, that time belongs in the calculation.
- Data protection: reviewing the data processing agreement, updating records of processing, possibly an impact assessment.
- Model changes: older models are switched off. A well-built solution can be moved over with manageable effort.
How do you spot inflated offers?
- A large strategy project is sold before the first working prototype.
- A custom model is to be trained although ready-made models are sufficient. More in Do we need our own AI model?
- The offer says nothing about ongoing costs or ties you to one provider for a long time.
- Nobody works out when the investment pays for itself.
How to keep costs under control
Start with what you already have: existing licences and the features of your current software often cover more than expected. Whether additional licences pay off is discussed in Microsoft Copilot for everyone? For an integrated use case, a prototype with a fixed budget and a clear success criterion comes next. Which applications suit this is covered in Which AI projects pay off?
Julius Launhardt works out AI projects for small and mid-sized companies in advance, including ongoing costs, and then builds them. More on the AI consulting page.
Frequently asked questions
Is ChatGPT for business not cheaper than a dedicated project?
For general tasks such as drafting texts, often yes. As soon as AI needs to be built into your processes, for example to transfer data from documents into your software, you need an integration that a chat tool alone cannot provide.
What does an AI prototype cost?
It depends on the scope. For a clearly defined use case, a working prototype is often possible within a few weeks and for a mid four-figure amount.
Are AI models getting cheaper?
Prices per request have dropped considerably in recent years. Still, plan with current prices and build the solution so that the model can be swapped later.
Are open-source models free?
The models themselves often are, running them is not. You need powerful servers or a hosting provider. For small volumes, a European provider that bills by usage is usually cheaper than your own hardware.