Knowledge
The subscription is not what gets expensive. The dependency is.
AI is not a question of processor, memory or model. AI is a toolbox - with more and more tools, each more specialised than the last. Like a set of screwdrivers: none fits everywhere, and whoever buys only the biggest strips the small screws. Which tool goes where is not written on the box. It is written in your workflow.
Updated: 30 September 2026
What you will not find here
No table of which model runs on which graphics card. No comparison of tools that will be called something else next month. That exists elsewhere, and that is where it belongs. Here is what comes before it: understanding what you actually need.
A toolbox, not a computer
For a short while, AI was a single tool: one large model that could do everything. Today it is a box. Small models that do one thing well. Medium ones that run on your own hardware. Large ones you pay per request. And tools that are not a model at all, but a rule with a calendar behind it.
Whoever thinks of AI in terms of processing power, memory and model size buys the biggest screwdriver and wonders about the bill. Whoever thinks of it as a box reaches for the right one, screw by screw.
Which tool goes where - your workflow tells you
It is not the model that decides, it is the point in the workflow. What happens there? What is being worked with - text, numbers, images, appointments? Which data comes along, and who owns it? Does the answer have to be verifiable, or is a good draft enough?
Whoever has answered that for every point has already sorted the tools. The hardware question comes afterwards - and usually at one single point.
The cost trap
The large models can do everything, and right now they are cheap. By all public accounts, they do not cover their costs at today's prices. That is an introductory phase, not a price list.
Meanwhile AI grows into the workflows: first the summary, then the enquiries, then the documentation. Every point saves time. Every point makes you more dependent - not on a provider, but on a price that will not stay this way. Whoever then pays for the biggest screwdriver on every screw has a problem that was not one before.
We expect the moment of cost truth. When, nobody knows. That it comes, every flat rate has shown that did not add up under sustained load.
Everything local? Almost never
The answer to the cost trap is not to bring everything in-house. A model in-house is a tool like any other - expensive to buy, limited in what it can do, and for most points in the workflow simply the wrong one.
A medical practice, four tasks:
- Point in the workflowSummarising findingsToolModelWherein-house
- Point in the workflowSuggesting appointmentsToolRule + calendarWhereno model
- Point in the workflowBilling questionsToolLookup with expiry dateWhereno model
- Point in the workflowWebsite copyToolModelWhereoutside, anywhere
Whoever starts without this sorting buys the most expensive solution for all four - or the cheapest, and sends findings where they do not belong.
Which model? Depends on the workflow
There is no right answer without the workflow. What happens, what you work with, which data is involved - the tool follows from that, not the other way round. For doctors, lawyers and tax advisors, data protection decides before anyone talks about quality: what must not leave the building needs a tool in the building. Everything else may run outside, as long as it is clear where its answer comes from.
Sort first, build second. That is not caution. That is efficiency.
Common questions
Do I need my own hardware for local AI?
For the point where data must not leave the building: yes - in-house or on a dedicated server with a provider of your choice. For everything else, no. That is why hardware comes at the end of the sorting, not at the beginning.Is a small model worse?
On a narrowly defined task usually not worse, just smaller - like a small screwdriver on a small screw. It gets worse when you use it for everything.Why talk about cost when AI is so cheap right now?
Because cheap is a phase and dependency is a condition. What you build into your workflows today will still be running in three years - at a price nobody knows today. Sort first, and the adjustment is one line. Do not, and it is a project.
Sort first, build second
Bring your workflow. We go through it point by point and tell you where which tool belongs - and where none does. Building comes after that, and only if it pays off.
Go through the tasks