The AI bill, and the router that shrinks it

The pilot phase is over. And now the bill arrives.

More and more companies are running into the same thing: the AI pilots have finally become real rollouts, with real effect on the business. Which makes sense, because by now it is clear that AI can genuinely add value, and on a far larger scale than earlier waves of automation.

And precisely for that reason, something that was there all along but too small to notice is now becoming visible: AI at scale can get quite expensive. Handelsblatt wrote about it this week.

With agentic AI in particular it adds up fast. A single question from a user triggers dozens to hundreds of model calls in the background. An agent working on its own for an hour keeps the meter running hard.

Fortunately, developments in the AI world are moving extremely fast, and the hottest thing in town happens to be the answer to this problem. Not every call needs the most expensive model. Within a single model family there is easily a factor of 18 to 100 in price difference between the smallest and the largest model. And most of what an agent does is routine work: summarising, formatting, calling a tool. You do not need a frontier model for that.

Enter: the model router.

A model router is a small AI model that acts as a traffic controller, deciding per call which model is good enough. Simple task? Small, cheap model. Difficult analysis? Bring out the big one. ChatGPT has been doing this internally since GPT-5, Microsoft now has a product for it, and with open source tools such as LiteLLM you can build such a layer between your application and the models yourself.

Vendors claim savings of 40 to 70%.

Caveats (yes, those too):

• Those percentages come from the people selling routers. Take a generous pinch of salt.
• The router is a model itself, so it can guess wrong too. Without measuring quality, you do not know what your saving is costing you.
• Cheaper tokens do not release you from the question: what is this AI workflow actually bringing in? Cutting costs without knowing the return is bookkeeping with one column.

That last point is, as far as we at BrA1n are concerned, the real signal. Companies that tackle this systematically, cost and value per workflow, are the companies where AI is growing up. This is what we developed the BrA1nScan for.

Does your organisation know what its AI workflows cost? And what they bring in?