A customer books a repair. The agent checks the Vehicle Identification Number (VIN), approves the warranty claim, schedules the technician, with no human in the loop. Then it does the same for a vehicle under an open safety recall, because nothing in its reach flagged the VIN. The claim is technically valid. It’s also the wrong action.
On the plant floor the failure repeats with different nouns: an agent reorders a part from a supplier whose last batch is under a quality hold, because the hold lives in a system nobody connected. The context simply wasn’t there.
The sector that’s supposed to prove this out
Manufacturing and mobility are betting hardest on agentic automation and are among the first to admit it isn’t working yet. Gartner predicts that by 2029 only 5% of automakers will sustain heavy AI investment growth, down from over 95% today, as “AI euphoria” collides with weak data foundations. Industry-wide, it expects more than 40% of agentic AI projects to be canceled by 2027.
The upside is real too: McKinsey’s 2026 state-of-AI survey found advanced manufacturing scaling agents fastest in supply chain and in the manufacturing process itself - one of the few functions where organizations consistently report AI-driven cost reductions.
The failure mode has a name, and it isn’t the model
Ask why a pilot that worked on one line falls apart at scale, and the answer is almost never “the reasoning was wrong.” The agent reasoned over a fraction of the truth: a parts extract instead of the live MES, a service history missing the open recall, an approval chain in a supplier portal nobody wired in. Gartner expects organizations to abandon 60% of AI projects not backed by AI-ready data through 2026. More autonomy with less context isn’t a smarter system - it’s a faster way to be confidently wrong.
What “context” has to mean on the floor and in the field
Enterprise context means the VIN’s full service and recall history, the supplier quality status behind a given batch, who can approve a warranty payout or a line stoppage, and what happened the last time this failure appeared. That takes connectors mirroring the access controls of the source MES, DMS or PLM; a search layer fast enough for real time; a knowledge graph linking vehicles, parts, suppliers and service events; and a memory of how past exceptions were handled. One large automotive marketplace operator has built exactly that on Amazon Bedrock AgentCore - a single layer serving agents across shopping, fleet service, auctions and dealership operations.
Standardize the wiring, then add brakes
Before an agent can act on an MES or a DMS, it needs a consistent way to reach them. The Model Context Protocol, now under the Linux Foundation’s Agentic AI Foundation, does that without a bespoke integration per pairing. But reach isn’t permission. Under the EU AI Act, an AI system acting as a safety component of a type-approved vehicle - ADAS, automated driving, driver-monitoring systems is high-risk by default, regulated through vehicle type-approval, with compliance due by August 2028, as extended by the 2026 Digital Omnibus. The Machinery Regulation extends that to agents touching robotic control or safety interlocks; ISO/PAS 8800 gives teams a way to argue AI safety cases alongside ISO 26262 and SOTIF.
Let the workflow drive, not the model
The organizations getting this right aren’t letting a model run recall management or line scheduling end to end - a workflow that drifts silently is worse than one that fails loudly. A deterministic engine owns the sequence and the rollback; the agent handles the step that needs judgment and hands back a structured result. When a step fails, the workflow compensates the earlier ones instead of leaving the ERP, the DMS and the supplier portal in three different states.
None of this is a reason to slow down. It’s a reason to build in the right order: context before autonomy, standard connections before custom ones, a human on the loop before a production line pays for a decision nobody can explain.
If you’re piloting agentic AI on the floor or in the field, the question isn’t “did it work?” It’s “what did it know when it acted, and would it have caught the recall, the hold, the exception that mattered?”
Ferfier’s approach to agentic automation in manufacturing and mobility starts with exactly this enterprise-context question, not the automation layer. If you’re evaluating where your own organization stands, talk to us.
