Big companies now split AI work by model. Give each of your processes its own model too.

If you picked one model and use it for everything, large companies are now doing it differently: they choose a model for each job. If you are still deciding which work to hand to AI at all, start with what to hand to AI first. This post is about the next question: once a process runs on AI, which model should run it?

What changed

On September 27 the Financial Times reported that U.S. companies are moving more of their AI work onto open-weight models. These are models whose weights are published, so anyone can host and run them, either through a provider or on their own machines.

The numbers in the FT’s report:

Mentions are up sixfold. According to AlphaSense data cited by the FT, executives mentioned open-weight or open-source models six times as often in earnings calls and investor conferences in August and September as in the same two months a year earlier.
Most tokens on one gateway are now open. The FT reported that open-weight models handled 56% of the tokens that went through Vercel’s AI Gateway in August, up from 7% in December.
AT&T runs about 40% of its AI workloads on open models and aims to reach 70% within a year, according to Andy Markus, its chief data and AI officer. Markus told the FT the company is doing 45 billion tokens a day, and that AT&T moves traffic from an expensive closed model to an open model when it can keep the accuracy. It also tunes open models on its own data. All of these figures are company-reported.
Tinder’s annual AI spending rate went up tenfold. Its chief technology officer told the FT that the company was spending at a rate of $1 million a year in January and $10 million by July (company-reported). Tinder has started sending some queries from non-technical users to open-weight models.
Custody matters too. The FT reported that Digital Realty will not put customer data into a frontier model.

Tokens are not the same as dollars

Vercel’s own report on its gateway, published September 17, adds an important detail. In August, open-weight models handled 56% of the gateway’s tokens but only 14% of its estimated spending. Anthropic took 64% of that spending. Vercel estimates spending from the labs’ published list prices.

Open-weight models did most of the token volume and a small share of the spending. Frontier models still took most of the money. Vercel says teams reserve frontier models for the tasks that justify the premium. Nobody in these reports dropped frontier models.

Why this matters to a business of one

An owner who pays for agents by the token has the same problem as Tinder, only smaller. When an agent starts running a process every day, the bill grows with the volume. Sorting invoices, tagging support emails and filling in CRM fields are high-volume work, but each step is simple.

The other question is custody. When a model runs on your own machine, client data does not have to go to an outside lab. That will not settle every data question you have. For some clients, though, it is the easiest answer to give.

A three-question routing rule

Run each AI process you have through these three questions. Answer them for the process, not for the business as a whole.

Can a person check the output in under a minute? Examples are a category, a filled field, or a short reply that uses a template. If yes, a cheaper model’s mistakes are easy to catch.
Does it run at least weekly, in the same shape each time? Repeated work is where the cost adds up and where you can test a model against past examples.
Does it touch data a client would not want sent to an outside lab? If yes, look at an open model you host, or with a host you choose, before you think about price.

How to read the answers:

Yes to 1 and 2: try an open-weight model first. Keep the frontier model as the fallback.
No to 1: keep the frontier model. This includes judgment calls, writing that goes out under your name, and anything hard to undo. Saving money doesn’t help if nobody can tell when the answer is wrong.
Yes to 3: decide where the process runs before you decide which model runs it. This is a custody question, and your contracts may already answer it.

Test it before you switch

Don’t move a process on trust. Take 20 past cases that the current setup handled, run them through an open-weight model, and score both sets of results against the same short checklist: right answer, right format, and nothing invented. If the open model matches on the cases that matter, route new work to it and spot-check a few each week. If it doesn’t match, you have lost an afternoon, not a client.

What to do this week

List every process that an AI model runs for you now, and note which model runs it.
Run each process through the three questions above.
Pick the one process with the highest volume that got a yes on questions 1 and 2, and run the 20-case test.
For any process that got a yes on question 3, check what your client contracts say about where their data may go.
Look at last month’s AI bill by process, if your provider shows it, so you know what you are trying to cut.

If you would rather work through this with someone, Build is a dedicated team that works and builds with you and teaches you to run what it builds.