Picture this. It is a Monday, and your VP has just asked how far along your team is with AI. You have a few people using it informally, a couple of experiments, and no number you could put on a slide. You are wondering whether that makes you late or just ordinary.
SimScale has a data point for you. It polled the audience at its Agentic Unlocks event, and the result was not flattering to its own market. Just 1 in 20 respondents said they are moving AI from pilot into standard workflows. Seven in ten teams have not moved past informal experiments, and 37% have not started at all. Only 16% have AI embedded in daily work or are scaling it.
If you have started, the picture is not rosy either. Among teams that have begun, more than twice as many are behind plan as ahead, 20% against 9%. And 22% have defined no AI KPIs, so they could not tell you whether they are behind even if you asked.
Before you quote any of this upward, read the caveats the post itself insists on. More than 1,000 people registered for a vendor's own AI event, and the post gives no response count for any of the four questions. An audience that signs up for an agentic engineering event is not a random sample of engineering. That bias should push adoption numbers up, which makes the low figures more interesting than the high ones. The 95% who named measurable value deserve more suspicion, especially beside a 37% who have not started. The customer results are vendor-told too, such as Convion compressing a design loop from months to under one hour, within 5% of high-fidelity CFD. What survives is the direction of the finding. The post backs it with Capgemini's remark that fewer than 10% of its clients think they have scaled AI.
Then comes the advice, and it is the part you can act on. "Do not start at the tool. Start with a record that you already own." NVIDIA's Todd McDevitt argued for determinism in the agent harness, through typed tool calls, explicit permissions and physics checks. His line: "It's important to judge the system on its traces." If you build or buy APIs, you will notice that is a statement about interfaces.
So the post turns to SimScale's own record, and this is where it gets uncomfortable. SimScale has real interfaces: 7 API pages, including the Geometry API, Mesh Operations API, Simulations API and Simulation Runs API. The catalog description records the pipeline its customers rely on, from geometry upload and mesh generation through setup, execution and results. Dexory's repetitive setup, 198 contacts and 192 bolted connectors per tower section, is exactly the kind of job that sequence could carry. But the post credits the SimScale agent with that work and does not say whether it runs through these APIs.
The catalog's measurement of that public surface is modest. SimScale's Kin Score is 34.7, rated thin. Contract quality is 57.3 and discoverability 66.1, but operational transparency is 10.5, contract governance 13.6 and access clarity 18.4. Agent Readiness is 17.9, agent-aware. Spec presence, agentic access, auth clarity and rate limit signals are lit. The agentic access profile records 21 operations, 10 of them acting. Error semantics, idempotency, dry-run mode and reversibility are unlit, and those are the properties that make a long simulation safe for an outside agent to retry or to judge afterward.
The post is careful here. SimScale's own agent may have all of that internally. The public record does not show it.
What you take away is a way to read your own Monday. Being in the 70% is ordinary, so do not panic. Define a KPI, because 22% of started teams have none. And when you judge a vendor's agent story, including a vendor whose poll you just read, ask the harness question: can the system be judged on its traces, and does the public record show retries, errors and reversibility, or only the demo?
Read the original: https://apis.io/2026/11/02/simscale-polled-engineers-on-ai-and-one-in-twenty-are-scaling/