In the span of one week, Anthropic, OpenAI, and ServiceNow all announced they were sending engineers to sit inside manufacturers.
Three firms came to the same answer around the same time. Before we get into what that problem is, quick question:
How does your plant currently bring in AI implementation support?
👋🏻 I'm Leonardo Ubbiali. This week we're looking at why the AI handoff model is breaking, and why the labs building the models decided they needed to fix it themselves.
A manufacturer I spoke to recently had been running a predictive maintenance deployment for 14 months.
Their SI contract ended in Q4 2025. In March the underlying model updated and the recommendations stopped making sense.
There was nobody to call.
For a decade, the standard AI deployment playbook has been the same.
That model was built for software that holds still.
Foundation models like Claude update on a monthly or weekly basis. So it’s natural that six months after the handoff, the model updates, the output shifts, and the SI is gone.
All three announcements this week are a response to that specific failure.
The manufacturer in context needed someone to call when the model was updated. That is what embedded engineers provide.
Anthropic's firm puts engineers inside the company and they stay, maintaining the deployment as the underlying models improve.
When the model updates, the engineer is already in the building.
OpenAI built the same structure. Tomoro was acquired specifically to start with 150 working engineers rather than hire from scratch.
Their engineers have shipped production AI in environments where the system cannot go down.
ServiceNow and Accenture announced the same model four days earlier, with FDE teams building in production before enterprise rollout begins.
The Blackstone and Hellman and Friedman portfolios span hundreds of mid-market manufacturers.
So they become the first customers, followed by the independent manufacturers.
The manufacturers moving fastest are reaching out now and writing protective clauses around data residency and IP ownership of fine-tuned model weights before signing anything.
Five things you can do this quarter
The problem: Your AI system is in production and the underlying model has updated. You have no process for managing what changes.
What you need: A description of the system, how many times the model has updated since go-live, and what process you currently have in place.
The Prompt (copy this):
I'm a [YOUR ROLE] at a [FACILITY TYPE] manufacturing plant. We have an AI system in production [DESCRIBE THE SYSTEM] deployed by [INTERNAL TEAM / SYSTEMS INTEGRATOR]. The underlying model has updated [NUMBER] times since we went live and we have [NO PROCESS / A PARTIAL PROCESS] for managing those changes.
Tell me: What does a model governance process look like for a production AI system in a manufacturing environment? What should we monitor after each model update to catch output drift before it affects operations? What are the three questions I should ask any vendor or integrator about ongoing model stewardship before signing a new contract? If I wanted to engage one of the new lab-affiliated services firms, what should I have ready before that first meeting?
What you'll get back:
A model governance framework for your specific deployment, the three metrics to track after each update, a vendor contract checklist, and a preparation guide for engaging forward deployed engineers.
OpenAI Deployment Company launch announcement
The section on how forward deployed engineers work through an engagement, diagnostic, workflow selection, production deployment — is worth reading before your next vendor conversation.
Time to value: 10 minutes
The manufacturer I spoke to had spent three months looking for someone to call before they found us.
If that sounds familiar, hit reply below!
Leo