Open Models or ChatGPT? What Manufacturers Should Know Before They Pick an AI
For the repeat back-office work most plants start with, open-weight models cost less to run, keep data in an account you own, and can be pinned, tested and audited. Here is how to decide.
Short answer: for repeat back-office work such as entering customer purchase orders, chasing supplier confirmations or matching certificates to lots, a manufacturer is usually better served by an open-weight model running in an account it owns than by per-seat chat licenses. It costs less per task, the data does not go into a vendor's chat app, the model version can be pinned, and the model can be tested and swapped. Closed frontier models still have a place, for the occasional step that truly needs them.
What Is an Open-Weight Model?
An open-weight model is an AI model whose trained weights are published, so anyone can run it on their own infrastructure or on a cloud provider of their choice. Current examples include OpenAI's gpt-oss, Meta's Llama, Mistral and Qwen. People often call these "open-source models"; strictly, "open-weight" is more accurate, because the training data and code are not always published. What matters to a business is the same either way: you can run the exact model you tested, where you choose, for as long as you choose.
A closed model, such as the ones behind ChatGPT, Copilot, Claude or Gemini, is only available as a service. You send it your data and get an answer back, on the vendor's terms and the vendor's schedule.
How Do They Compare on Cost?
Closed AI usually reaches a company in one of two ways: per-seat licenses for chat assistants, or per-token prices for API use. Per-seat licenses charge for every employee who might use the tool, whether or not it does any work. Frontier API prices are set by the vendor and are typically several times higher per task than an open-weight model of the size most back-office work needs.
An open-weight model running on a platform such as Cloudflare is paid for by the work it does. For a workflow that reads a few hundred order emails a week, that cost is small and, more importantly, it is measurable. Before committing to a build, measure the cost per order on your own work, beside the hours it saves.
Where Does Our Data Go?
This is the question plant owners and controllers ask first, and they are right to. When staff paste a customer PO, a drawing note or a price into a chat tool, that data has left the company's systems, and nobody wrote down where it went.
An open-weight model deployed in an account the company owns keeps the data path short and written down: the email arrives, the model reads it inside your account, the draft lands in your ERP for approval, and the run is logged. No vendor trains on it. If an auditor or a customer asks where their data goes, there is a one-page answer.
Who Controls Changes?
Closed models update on the vendor's schedule. A model that read your purchase orders correctly last month can behave differently after an update you did not ask for. In manufacturing terms, that is an uncontrolled process change.
With an open-weight model you pin the version. It changes only when you test the new one on your own examples and sign off, the same way you would validate a change to a production process. Add a named person who approves every action that sends, pays, files or changes a record, and a log of every run, and you have an AI step that fits a quality system instead of fighting it.
Can We See How It Works?
The weights of an open model are public, so the model can be evaluated by anyone, and you can see exactly which model ran each step of a workflow. If a better or cheaper model appears, you can swap it without rebuilding the workflow around a different vendor's tools. That is the difference between owning a capability and renting one.
Are Open Models Good Enough?
For reading orders, emails and PDFs, extracting fields, matching part numbers and customers, drafting entries and routing exceptions, current open-weight models do the job. Do not take that on faith: run the workflow on real work beside your team, measure time per item, rework and exceptions before and after, and decide on the numbers.
Where a single step genuinely needs a frontier model, use one for that step only, with the reason and the cost written down and the owner's sign-off. Most workflows never need it.
A Simple Way to Decide
- Use per-seat chat licenses for open-ended help: drafting a letter, summarizing a meeting, brainstorming. Write a short policy on what data may go into them.
- Use an open-weight model in your own account for repeat work that touches the ERP, customers or suppliers, where cost per task, data path and change control matter.
- Escalate to a frontier model only for a named step, with approval.
Start with one workflow that already eats hours every week, such as customer PO entry, measure it, and let the numbers decide what comes next.