SEC-04 / Data Foundations

Industrial data readiness for AI: a 40-point checklist with scoring

Before you fund an industrial AI use case, score your data against 40 checks in eight dimensions. The lowest-scoring dimension, not the average, tells you what will break first.

6 min read Updated Plant managersCIOsOT engineersCISOsOperations leaders

Key takeaways

  • Data readiness for AI is about governed data: known, connected, in context, good enough, owned, secured, on a platform you can run, with people who can use it.
  • Score each check 0, 1 or 2. Look at the lowest dimension first; one weak dimension can sink a use case.
  • Score per use case or per site, not for the whole company at once.
  • Governance and ownership checks are usually cheaper to fix than platform checks, and often matter more.
  • Repeat the scoring every six to twelve months and after major changes.

Most industrial AI use cases are approved on the strength of the idea and the demo. Few are approved on the strength of the data. That is backwards. The model is usually the easy part; the data it depends on decides whether the use case survives on a real line with real shifts, changeovers and maintenance windows.

This checklist gives you 40 checks across eight dimensions, with simple scoring. Use it before you fund a use case, before you sign with an AI vendor, and again before you scale a pilot. If you prefer a guided version, take the industrial data readiness assessment; the industrial data readiness scorecard is the spreadsheet version for workshops.

How to use the checklist

  1. Pick a scope. One site and one or two AI use cases from your AI use case inventory. Scoring "the company" produces an average that hides everything.
  2. Get the right people in the room. OT or controls, IT or IT security, the process owner and the use case owner. Add quality or maintenance as relevant.
  3. Score each check 0, 1 or 2 using the guidance below. Ask for evidence, not opinions.
  4. Total each dimension (maximum 10 per dimension) and the whole checklist (maximum 80).
  5. Act on the lowest dimension first.

Scoring each check

Score Meaning Evidence test
0 Not in place, or nobody knows No document, no owner
1 Partly in place, informal, or only for some assets Exists but incomplete or out of date
2 In place, documented and maintained for the scope Current document, named owner, used in practice

The 40 checks

1. Inventory

  1. The data sources for the use case (PLCs, historian, MES, LIMS, CMMS, cameras) are listed.
  2. Relevant tags or data points are identified, with descriptions and engineering units.
  3. Data-producing devices, including edge gateways and smart sensors, appear in the OT asset inventory.
  4. Existing data flows leaving OT, including vendor connections, are documented.
  5. The AI use case is recorded in a register with its data dependencies.

2. Connectivity

  1. Data can be collected from all required sources without manual export.
  2. Collection paths go through a defined DMZ or equivalent boundary, not direct from Level 2 to the internet.
  3. Local buffering prevents data loss during network or cloud outages.
  4. Sample rates are sufficient for the use case (and documented).
  5. Legacy or OEM-locked equipment has a supported, approved collection method.

3. Context

  1. Data maps to a consistent asset hierarchy (for example ISA-95 levels: site, area, line, cell, unit).
  2. Product or SKU and recipe are available and time-aligned with process data.
  3. Batch, lot or order identifiers are linked to process data.
  4. Machine state (running, idle, changeover, fault, planned stop) is captured.
  5. Quality outcomes (lab results, defects, rejects) can be linked back to the production run.

4. Quality

  1. Clocks across sources are synchronized and timestamps are stored in UTC with a known source.
  2. Historian compression and deadband settings are documented and suitable for the use case.
  3. Calibration records for critical sensors are current and linkable to tags.
  4. Bad, frozen and out-of-range values are detected and flagged, and quality flags are respected.
  5. For labeled data (failures, defects), definitions are written down and consistently applied. See industrial data quality.

5. Governance and ownership

  1. Each data domain in scope has a named data owner.
  2. A data steward maintains definitions, mappings and quality rules.
  3. Rules exist for who may use the data and for what purposes, including vendors.
  4. Retention periods are set for raw and derived data.
  5. Changes to tags, topic structures, historian settings or labels go through a change process (MOC or equivalent).

6. Security

  1. Data paths out of OT are documented as conduits in the zone model.
  2. Paths are outbound-only; nothing outside can initiate connections into control zones.
  3. Credentials are unique per system, least-privilege and rotated; read and write rights are separate.
  4. No AI output can write to control systems without a separate, approved risk assessment.
  5. Vendor access to data and devices is contractually defined, logged and reviewable.

7. Platform

  1. There is a defined system of record for time-series data (historian or equivalent) for the scope.
  2. The data platform (on-prem, cloud or hybrid) has an owner, a support model and a budget.
  3. Data can be extracted for model training without disrupting operations.
  4. Lineage from source tag to model input can be traced.
  5. Costs for storage, compute and egress are monitored.

8. Skills and operating model

  1. Someone in-house understands both the process and the data well enough to judge model outputs.
  2. OT and IT have an agreed way of working on data projects (for example, a joint IT/OT data review).
  3. The people who will act on AI outputs (operators, planners, inspectors) are involved in the design.
  4. There is capacity to maintain the data pipelines and models after the project team leaves.
  5. Lessons from previous pilots are written down and used.

Note: Checks 26 to 30 overlap with OT security work. If you have already scored your OT security maturity, reuse the evidence.

Reading your score

By dimension (0 to 10)

Dimension score Interpretation Typical action
0 to 3 Foundation missing Do not scale use cases that depend on this dimension. Fix first.
4 to 6 Partial, informal Pilot only, with gaps recorded as risks and owners assigned.
7 to 8 Workable Proceed, with targeted fixes.
9 to 10 Strong Maintain, and reuse as a template for other sites.

Overall (0 to 80)

The total is useful for tracking progress over time and comparing sites. It is not useful for go/no-go decisions on its own. A site scoring 60 with a 2 in security is less ready than a site scoring 50 with no dimension below 5.

Plant-floor reality: The most common pattern in mid-market plants is reasonable connectivity and platform scores, with low context and governance scores. The data is flowing; nobody can say what it means or who owns it.

Which dimensions matter most for which use case

Use case Critical dimensions
Predictive maintenance Quality, context, governance (failure history)
Vision quality inspection Quality (labels), context, skills
Process optimization Context, quality (timing), security (if write-back)
Energy analytics Connectivity, context, platform
Generative AI on plant documents Governance, security, inventory

For example, a predictive maintenance use case that scores 3 on quality because failure history is incomplete should fix the CMMS before buying more sensors.

Turning scores into actions

  1. List every check scored 0 in the critical dimensions for your use case. These are blockers.
  2. Assign an owner and a date to each. Many governance checks (owners, retention, change process) can be fixed in weeks without spending capital.
  3. Record remaining gaps as risks in the AI risk triage worksheet for the use case.
  4. Re-score after fixes and before scaling.
  5. Feed results into your plan. The 90-day AI governance plan has a natural slot for data readiness actions.

Common mistakes

  • Scoring the whole company at once and getting a meaningless average.
  • Scoring from memory instead of evidence.
  • Treating platform purchase as the fix for low context or governance scores.
  • Leaving OT out of the scoring, or leaving IT out.
  • Scoring once and never again.

Monday-morning checklist

  1. Choose one site and the AI use case closest to funding.
  2. Book a two-hour session with OT, IT, the process owner and the use case owner.
  3. Score the 40 checks using the scorecard.
  4. Circle the lowest dimension and every 0 in the critical dimensions.
  5. Assign owners and dates. Recheck in 60 to 90 days.

Effort: the first scoring session takes a few hours. Evidence gathering and fixes for one use case typically take weeks of part-time effort, mostly from people you already have.

What good looks like

Every funded AI use case has a current readiness score with evidence. No critical dimension sits in the lowest band. Gaps have owners and dates. Scores improve over time and are reused across sites. And when a vendor promises results "with your existing data," you can answer with a scored, evidenced view of what that data can actually support. The scoring approach is explained on the assessment methodology page.

Sources and further reading

  1. NIST AI Risk Management Framework
  2. ISA-95 Standard (ISA)
  3. NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

What does industrial data readiness for AI mean?

It means the plant data a given AI use case depends on is inventoried, reachable through a secure path, carries production context, is of known and adequate quality, has named owners, is protected appropriately, sits on a platform you can operate, and can be used by people with the right skills. It is specific to a use case; a plant can be ready for one and not another.

How long does a data readiness assessment take?

For one site and one or two use cases, a working session of a few hours with OT, IT, quality and maintenance usually produces a first score. Gathering evidence for each check, such as tag lists, firewall rules or calibration records, takes longer, typically spread over a couple of weeks of part-time effort.

What score do we need before starting an AI project?

There is no universal threshold. As a rule of thumb, avoid scaling a use case while any dimension it depends on scores in the lowest band. A pilot can proceed with gaps if they are known, recorded in the risk assessment and have an owner and a fix date.

Who should take part in the scoring?

At minimum, a controls or OT engineer, someone from IT or IT security, the process or area owner, and the business owner of the AI use case. Add quality for vision or quality use cases and maintenance for reliability use cases. Scoring alone in IT or alone in OT misses half the picture.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.