Key takeaways
- Readiness belongs to a specific workflow, not to the company as a whole.
- A baseline, representative test set and failure conditions matter before model selection.
- The production owner and safe fallback must be named during discovery.
- A technical proof of value is not the same as an operating pilot.
What industrial AI readiness actually means
Industrial AI readiness is the ability to evaluate one proposed AI-supported decision using real operating evidence and responsible ownership. A plant can have a modern data platform and still be unready if nobody owns the workflow. A smaller factory can be ready with modest infrastructure if the problem is bounded, the records are usable and the team can review errors safely.
Do not begin with a model catalogue. Begin with a sentence: When this event happens, this person currently decides this action using this evidence. That sentence exposes the user, decision, inputs, action and accountability.
The six readiness gates
1. A decision worth improving
Choose a repeated maintenance, quality, planning, energy or knowledge decision. Define the present delay, error, workload or exposure without assuming that AI is the answer.
2. A named process owner
The owner defines acceptable behaviour, supplies domain reviewers and decides whether the system can enter production. An IT sponsor alone is rarely enough for an operating workflow.
3. A reviewable baseline
Record how the workflow performs now. Depending on the use case, the baseline may include missed events, false alarms, inspection escapes, review time, downtime, energy intensity or accepted recommendations. Use definitions the plant already understands.
4. Representative evidence
Check whether the available records cover shifts, products, asset states, failures, acceptable variation and known edge cases. More rows do not fix missing operating conditions or inconsistent labels.
5. A permitted architecture
Document where data may move, which systems can be read, whether write-back is allowed, who can access output and how models or knowledge will be updated. Offline operation still needs identity, logs, backup, patching and support.
6. An acceptance and fallback plan
Define the held-out evaluation set, thresholds by scenario, human-review capacity and what happens when confidence is low or the system is unavailable. Safety, quality and regulatory owners retain authority.
Evidence to collect during readiness
- A one-page workflow map with user, decision, evidence, action and escalation
- A baseline with definitions, period, exclusions and owner
- Sample records from normal, adverse and unusual conditions
- Source-system owners, access route and data restrictions
- Error costs for missed events and false alerts
- A preliminary production boundary and operating owner
- Acceptance questions that can produce a go, revise or stop decision
Proof of value versus production pilot
A technical proof of value asks whether the data contains a useful signal. A production pilot asks whether the complete system works inside the real workflow with identity, interfaces, monitoring, review, fallback and operating support. Treating the first as the second is a common source of failed scale-up.
Applying Industry 5.0
The European Commission frames Industry 5.0 around human-centricity, sustainability and resilience. For an SME, that means asking three additional questions: Does the system help people perform safer and more meaningful work? Does it improve resource use without shifting harm elsewhere? Can the operation continue when the model, network or supplier is unavailable?
A responsible next step
Run readiness on one workflow and one representative dataset. The output should be a decision record, not a sales score: proceed to evaluation, close a named gap, choose a different use case or stop.
