The question manufacturers ask about agents is less 'can it do the task?' than 'what is it allowed to do?'. On a shop floor or in a quality lab, the honest answer differs by decision.
Autonomy by decision, not by agent
The decision platform we built uses three autonomy levels — observe, recommend and act — and the alert workflow keeps a person in charge of accepting, assigning, escalating or dismissing every finding1. Our agent designs for manufacturing use the same five-level scale as our wider digital workforce, with quality, safety and recall decisions capped at suggestion.
Value-ledger stages
How agent findings are counted in the Agentic Decision Platform
| Stage | Meaning |
|---|---|
| Identified | An agent found it and quantified it |
| Approved | An owner accepted it |
| In flight | Action is under way |
| Realised | Finance confirmed the value |
Note: Four stages, as built.
Answers that show their working
For R&D and engineering questions, trust depends on provenance. The SCIKIQ R&D Knowledge Assistant shows the SQL, passages and figures behind every answer, with six types of source marker, and keeps high-security formulations behind access control1.
Where to start
Write down the decisions in one domain, set an autonomy level for each, and stand up the value ledger before the first agent goes live.