Should we use one AI model for everything, or a cheaper model for simpler work?
The short answer
In my view, most companies will end up using more than one model. Keep the work that is sensitive, or that needs careful reasoning, on a model you trust with your data. Give the simple, repetitive work to a cheaper model. Price alone should not decide it: where your data goes, and what the law says about it, decide first.
The race is changing
For a few years, AI companies competed on who had the strongest model. In my view, that race is turning into a race on price. As I see it, some providers are pushing AI to become a commodity, so that the contest is settled by price. For many everyday tasks, the gap between the strongest models and the cheaper ones matters less than it used to.
For a company, this is good news. You no longer have to pay the highest price for every task.
Not every task needs the strongest model
Look at what your agents actually do in a day. Much of it is simple work:
- Sorting incoming emails by subject.
- Pulling the date, the amount and the supplier from an invoice.
- Turning a long message into three lines.
In my experience, a cheaper model often does this well enough. Paying for the strongest model here is like asking your most experienced manager to sort the paper mail.
Other work is different. Reviewing a contract clause, answering a client about money, or deciding which supplier to pay first needs careful reasoning. A mistake there can cost far more than the model does.
Mixing models is already happening
Some companies already run hybrid systems. They send sensitive work to a model they trust, and send the simple work to a cheaper one. A simple rule, or the agent itself, picks the model for each task.
Why price alone is not enough
A cheaper model may be run by a provider in another country, under other laws. Before you send it anything, ask three questions:
- What data will it see? If the task touches personal data, client records or salaries, the cheaper option may not be allowed at all.
- Where will the data be processed? Many countries limit sending personal data abroad. The pages under What do the rules say in my country? explain the rules in 14 countries.
- How will you know if it gets worse? A cheap model that is wrong one time in ten can cost more than an expensive one. Measure its answers on your own work before you rely on it.
In my view, governments will decide this for their own agencies on grounds other than price, such as data protection laws and state policy. Private companies have more room to choose, and I expect many of them to move their simple work to cheaper models.
What this means for you
- List the tasks your agents do, and mark each one simple or sensitive.
- Keep sensitive work, and work with personal data, on a model you trust and that the law allows.
- Try a cheaper model on the simple work, and compare its answers with your current model on real examples.
- Build your agent so that changing the model for one task does not mean rebuilding everything.
The question When we use an AI agent, where does my company’s data go, and can it stay in my country? explains what to check before any data leaves your company.