Plants have spent a decade buying visibility and are discovering that dashboards do not fix machines. The decision layer is what sits between knowing and doing, and it is where manufacturing software goes next.
Quick answer: A decision layer is the software layer that turns production data into executed, verified action: it detects a loss, recommends the response, routes it through human approval into the systems that do work (maintenance, scheduling, procedures), and then measures whether the promised impact arrived. It differs from analytics in one word: analytics ends at insight, a decision layer ends at verified outcome.
The modern plant does not suffer from a data shortage. It suffers from an action gap: the distance between a loss being visible on a dashboard and the fix being scheduled, executed and confirmed. In practice that distance is meetings, memory and goodwill.
The stop appears in the Monday review; someone promises to look; the work order gets created Thursday, or never; nobody checks whether the fix worked. Every step is human glue, and human glue is what breaks first under daily pressure.
Detects and explains: losses are captured at the source and classified with evidence, video included, so the input to every decision is trusted.
Recommends: episodic responses (create this work order, resequence these two orders) and systemic ones (this failure repeats: change the PM plan; this changeover always overruns: update the standard). The systemic recommendations matter most, because they rewrite the rules that generate tomorrow's losses.
Routes through people: recommendations carry expected impact and require approval, with reasons logged either way; the layer earns autonomy gradually by being right, not by decree. On the floor, this arrives not as another dashboard to monitor but as clear, prioritized work with the why attached, and operator feedback on a bad recommendation is an input the layer learns from, not an inconvenience.
Executes in the systems of work: the output is a scheduled work order, an updated plan, a revised procedure, not a notification.
Verifies: expected versus realised impact is tracked per decision, building the ledger that tells you which recommendations to trust.
Not a dashboard with alerts: alerts still end at a human's attention. Not a black-box autopilot: plants do not and should not hand control to unexplained models; the approval gate and the reason log are features, not training wheels. And not another integration project: a decision layer needs the operational systems (OEE, maintenance, planning) to share one data spine, which is why it works best where those are one platform rather than a federation of exports.
Three curves crossed. Measurement got honest enough to act on, with machine-verified capture and vision closing the classification gap. The connective systems (CMMS, planning) moved into the same platforms as the measurement, so an action has somewhere to go without a middleware program.
And AI became good enough to draft the recommendation and its expected impact, which changes the economics of the layer from consulting engagement to software. The plants adopting it first are not the most digital; they are the ones most tired of watching the same losses survive every dashboard.
How is a decision layer different from an MES?
MES executes production orders and enforces routings; it is a system of record for what is being made. The decision layer sits across MES, OEE and maintenance, deciding what should change and verifying that the change paid off.
Is a decision layer the same as AI agents in manufacturing?
Agents are one implementation mechanism. The defining property is the closed loop from detection through approved action to verified impact; whether the recommendation is drafted by an agent or a rules engine matters less than whether the loop closes.
Does a decision layer remove people from decisions?
No; it removes friction, not judgment. Recommendations pass a human approval gate with logged reasons, and autonomy expands only where the expected-versus-realised track record earns it.
Does a decision layer replace the continuous improvement team?
It industrializes their pipeline. The layer surfaces and drafts what CI teams today assemble by hand from exports and observation, freeing them from data janitor work; the systemic recommendations (change the PM plan, update the standard) are precisely the decisions CI leads should own and approve. Plants where CI embraces the layer get compounding improvement; plants where it is imposed around CI get a turf war.
What do we need before a decision layer makes sense?
Honest measurement and connected execution: machine-verified loss data and a maintenance/planning system the decisions can land in. A decision layer on top of manual logs automates fiction.
Fabrico is built as this loop: sense, decide, standardize, act, measure. The long-form argument is our From Data to Decisions series.
Login
Blog