AI agents
AI agents that work on your development data, with a full trail
The large PLM vendors shipped copilots in 2026. Siemens, PTC, Dassault and Contact all answer questions over your data now. An agent is a different thing. It prepares the variant, runs the campaign, drafts the change request. That work is worth nothing if the agent used a stale model, saw data it should not have, or left no record of what it did.
The conditions
Three things that have to be true first
Complete and current data in one place. An agent that reads a requirement from Polarion, a model from a laptop and a result from a share works from three different dates. Permissions that apply to the agent exactly as they apply to the person who started it. A trail: what the agent read, what it ran, what it produced, and that it was the agent and not a person.
PTC's own guidance on engineering agents says much the same about grounding. The difference between vendors is whether the platform enforces the three conditions or the team promises them.
In Ref
How agents work here
Ref was designed for agents from the start. An agent works inside the same environment as the team, under the same rules as the person who started it. It reads what that person may read. What it produces is marked as machine-authored and lands in the same record as human work, with the trail. It drafts; a person decides what is kept.
That last sentence is our decision. The models could do more. We would defend the decision in front of an assessor, and we expect to have to. Virginia Tech researchers found in 2025 that standard text metrics cannot tell a language-model-written systems engineering artefact from an expert’s, while a close reading finds premature requirements, unsubstantiated numbers and overspecification. Those were the models of early 2025, and every generation since does better. The reason to keep a person in the loop survives that progress: a failure a metric cannot see stays hard to see as the drafts improve. The draft that reads right is the one that needs a person.
The work
What agents do, and what they do not
They summarise a campaign, draft a change request from a failed test, prepare a variant from an existing one, run a Python analysis over results and file it with its inputs. They do not release or authorize product changes.
Throughput
When output multiplies
Agents multiply output. More variants, more runs, more drafts than a team can read. The platform's job is to keep decisions grounded in evidence when that happens: every draft points at what it was made from, every comparison at its runs. The review meeting gets shorter because the question of where a number came from is already answered.
Sovereignty
Models and data
Agents run on models hosted on Microsoft Azure in the EU. Your data is not used to train anything. That a result was produced by an agent is recorded with the result. Which models we use, we say in the demo.
Limits
What it costs
Agent work consumes tokens, metered per tenant and per topic; administrators and users see the consumption in the application and decide on it. An agent's draft for a campaign needs a person's review, and that time is where the value is decided; budget for it. Agents act only when a person starts them. Grunetal holds no ISO 27001 certificate as of September 2026.
Questions
Asked before every pilot
Can the agent see everything?
It sees what the person who started it sees. Nothing more.
Is agent output labelled?
Yes. Every artefact an agent produced carries that mark in the record, permanently.
Which models?
Models hosted on Azure in the EU. We name them in the demo, and we change them when a better one is available.
Can we switch agents off?
They act only when started by a person. A team that never starts one never sees one.
Related
If you want to see an agent run a campaign and file the result with its trail, ask for a demo. Tell us about your project.