Bring Your Own Language Model
TwinWorX uses a large language model for the parts of the work that call for language: drafting fault rules from a sequence of operation, explaining a fault in plain English, summarizing what happened overnight.
It does not have to be ours.
TwinWorX Copilot, our assistant and not Microsoft's, runs against any LLM endpoint, including the one your organization has already approved and already governs. It does not require Microsoft Copilot.
Four things follow
They are the four your security review will care about, and they hold when Copilot runs on your own endpoint.
There is no AI sub-processor to declare. No third-party model sits in the data path, so there is nothing to add to your data processing agreement.
Your existing approval is the approval. Many organizations have spent a year getting one specific model through governance, under specific terms. We fit inside that approval rather than asking you to assess another one.
No training on your data, structurally. Every vendor promises this. On your own endpoint, your operational data never reaches a third-party model at all, so the promise is not the thing protecting you.
No inference markup. Model cost and capacity stay in your account, so inference is not a line in our price.
What the assistant does with it
TwinWorX Copilot reasons over your model, not over a stream of context-free points.
It explains alarms and faults: why this one is active, what is likely behind it, and, where several arrive together, which is the cause and which are its symptoms.
It answers operational questions the way a colleague would. Has this happened before, and on what? What is unusual about this month's consumption? What changed on this unit since last week?
It helps with analysis: summarizing an anomaly, drafting a briefing, or interrogating a trend without anyone building a report first.
And it surfaces your procedures: the likely causes and first-line steps held against each fault rule, which your team tunes to its own operating procedures during deployment. The assistant tells an operator what your organization would tell them, not what a generic manual would.
The point is that a second-year operator on a night shift gets the answer that would otherwise need the one engineer who knows the plant.
What it is never allowed to do
A language model never acts on your building. It writes rules your engineers can read before they run, and a deterministic rule engine checks every one against safety limits and interlocks before anything executes. Life-safety systems stay connected read-only, and every writeback runs behind seven controls. Swapping the model changes the quality of the language. It does not change the boundary.
One honest caveat
Model independence is not model indifference. The quality of what comes back depends on the model you supply. We will ask which one early, rather than after go-live.
Bring your AI governance questions. Thirty minutes with one of our engineers, with your architecture or security lead in the room.
