Table of Contents
Data governance for an industrial company means deciding who is accountable for each kind of data, from historian tags and lab results to maintenance records and drawings, which rules apply to its quality, access, retention and sharing, and how those rules are checked. Start with an inventory, named owners and a few written policies, not a software platform.
Key takeaways
- Plant data needs governing as much as business data. Joint guidance led by CISA calls for "strict data governance policies" to protect sensitive OT data used by AI models (CISA et al., 2025).
- Every data domain needs a named owner, who is accountable, and a steward, who keeps definitions and quality rules up to date. IT runs the systems but should not own the data.
- Begin with an inventory of data and its metadata, which NIST's Cybersecurity Framework 2.0 lists as an outcome (NIST CSF 2.0), then classify it.
- Measure quality against a few defined dimensions, and judge it against a purpose: data is "unlikely to be equally fit for all purposes" (UK Government Data Quality Framework).
- Personal data in plant records falls under Law 18-07 as amended by Law 25-11; confirm your obligations with counsel.
What is data governance in an industrial company?
Governance is the decision layer above data management: who decides, not who does the work. A NIST glossary definition draws the line: "Data stewards enforce policies, and functional data managers implement the policies and manage day-to-day quality" (NIST CSRC glossary).
Industrial data has its own difficulties. Much of it is born in operational technology: sensors, the DCS or SCADA, the historian. Assets outlive the people who configured them, and engineering documents arrive from EPC contractors and vendors. In Algeria, one pump can be described in French in the CMMS, in English in the vendor manual and in Arabic in a shift log.
As the UK government's data quality framework puts it, "Fit for purpose data depends on effective data management and governance practices" (UK Government Data Quality Framework).
Why does data governance matter more once AI projects start?
Models expose hidden data problems. In interviews with 53 AI practitioners, Google researchers found that 92% had experienced at least one "data cascade", a compounding downstream effect of a data issue (Sambasivan et al., 2021). One interviewee working with power plant data had to assume timestamps "because metadata was missing", and the plant had been mapped incorrectly.
Ownership matters as much as quality. Google engineers warned that when an input signal and the model using it have different owners, the signal can change at any time. Their example: a model fitted to a mis-calibrated signal is suddenly affected when someone silently corrects the calibration (Sculley et al., 2015). In a plant, the equivalent is a recalibrated transmitter, a new range or a tag renamed after a revamp.
Then there is security. The CISA-led guidance notes that engineering configuration data, such as network diagrams, asset inventories and logic diagrams, has "enduring value" and is "highly valuable to cyber adversaries". It also warns that process measurements used to train a model "may become accessible or stored (statistically) for a longer period of time in the model" (CISA et al., 2025). Someone has to decide which data may go into which model.
Which industrial data needs governing?
Start by naming the domains. The owners below are a starting point, not a rule.
| Data domain | Typical system | Sensible owner | Main governance questions |
|---|---|---|---|
| Process time series | Historian, DCS, SCADA | Production or process manager | Tag naming, units, instrument changes, time zones, retention |
| Lab and quality results | LIMS | Quality or laboratory manager | Sampling time versus entry time, method changes, corrections |
| Maintenance and reliability | CMMS | Maintenance manager | Equipment hierarchy, failure codes, free-text reports |
| Engineering documents and configuration | Document management, engineering tools | Technical or engineering manager | Revisions, as-built versus design, restricted access |
| Master data: equipment, materials, suppliers | ERP | Supply chain or finance | Duplicates, codes consistent with the CMMS |
| Personal data in shift logs, HR files, access records | Various | HR, with the data protection officer | Legal basis, retention, transfers abroad |
Formulations, recipes and R&D results deserve their own domain, owned by R&D, as they are often the company's most valuable know-how.
Who owns the data, and who does the work?
| Role | Who, in a plant | Responsible for |
|---|---|---|
| Data owner | The manager of the business domain | Accountable for the data: classification, who may access it, approval to share it outside |
| Data steward | An engineer or analyst who knows the data | Names and definitions, quality rules, follow-up of errors |
| Data custodian | IT and OT administrators | Running the systems, backups, implementing access rights |
| Data users | Engineers, data team, contractors, vendors | Using data within the rules, reporting problems |
| Sponsor | A senior manager over operations and IT | Settling conflicts, funding, holding owners to account |
The CISA-led guidance asks for clear roles, naming "data owners, model developers, and end users", to avoid "liability and confusion" over responsibilities when an incident occurs (CISA et al., 2025). A new data team often supplies the first stewards; see how to build your first industrial data team.
What should the first policies cover?
Keep the first set short enough to be checked.
- Inventory. List each data source with its system, owner, location, retention and whether it holds personal data. NIST CSF 2.0 asks that "inventories of data and corresponding metadata for designated data types are maintained" (NIST CSF 2.0).
- Classification. Use three or four levels, such as internal, confidential and restricted, with formulations and control system configurations at the top. The same framework asks that assets be prioritised by classification and criticality.
- Access. Decide who may read and who may change each domain. Analytics and AI should read through read-only paths; see where AI fits in the Purdue model and one-way transfer with data diodes.
- Names and definitions. Naming rules for tags and equipment, units, and a data dictionary for the tags used in reports and models.
- Change management. Record instrument replacements, recalibrations, range changes and tag renames as events analysts can see.
- Retention and life cycle. Set how long each kind of data is kept and when it is archived or deleted; NIST CSF 2.0 asks that data be "managed throughout their life cycles".
- External sharing and vendors. The CISA-led guidance recommends a data usage policy "with clear reference to data residency, communications paths, encryption, and storage". Our list of questions about vendors training on your data helps turn this into contract terms.
- Quality rules. Define what each dimension below means for each domain, and who checks it.
How do you measure data quality?
The UK government's framework uses six dimensions from DAMA UK (UK Government Data Quality Framework). Each has a plant equivalent.
| Dimension | Definition in the framework | Plant example | Simple check |
|---|---|---|---|
| Completeness | "the degree to which records are present" | Historian gaps during a communication loss | Share of expected values present |
| Uniqueness | "no duplication in records" | The same pump created twice in the CMMS | Count of duplicate equipment records |
| Consistency | Values "do not contradict other values representing the same entity" | Different equipment codes in the CMMS and the ERP | Cross-system reconciliation |
| Timeliness | Data reflects the period it represents and is up to date | Lab results stamped with entry time, not sampling time | Delay between sampling and entry |
| Validity | Data "in the range and format expected" | Values outside the instrument range, frozen readings | Range and rate-of-change checks |
| Accuracy | "the degree to which data matches reality" | A drifting transmitter | Comparison with calibration or lab results |
Set thresholds per use: a monthly production report tolerates gaps that a soft sensor cannot. See common industrial data quality problems for causes and fixes, and process historian data for machine learning for the historian.
Which industry standards can you reuse?
- ISA-95, also published as IEC 62264, gives "an abstract model for information exchange" between business and control systems, with shared terminology and levels from the physical process (Level 0) to business planning (Level 4) (ISA).
- ISO 14224 provides a standardised format for reliability and maintenance data in the petroleum, petrochemical and natural gas industries, covering equipment data, failure data such as cause and consequence, and maintenance data such as actions and downtime (ISO 14224:2016 abstract). It is a ready-made reference for equipment hierarchies and failure codes in the CMMS.
- CFIHOS, an IOGP joint industry programme, "defines a standardized language for equipment and engineering deliverables" (IOGP). It is worth citing in EPC contracts for new units, so handover data arrives in a known structure.
What does Algerian law add?
This summarises a law-firm guide and is not legal advice. Confirm your situation with counsel against the official texts.
According to the CMS guide, Law 18-07 on personal data was amended by Law 25-11 of 24 July 2025. A controller may transfer personal data abroad only with the authorisation of the national authority. The amendment introduces a data protection officer, impact assessments for high-risk processing, and notification of a personal data breach "no later than five (5) days" after becoming aware of it. The same guide reports an ARPT decision requiring cloud hosting providers to keep customer data on national territory, and a national information systems security framework set by Presidential Decree 20-05.
In practice, the inventory must flag personal data, and the data protection officer should work with the owners of shift logs, HR files and access records. Know-how that is not personal data is protected mainly by contracts. Our guide to sovereign, on-premise AI covers the hosting side.
How does data governance connect to AI governance?
The NIST AI Risk Management Framework, released in January 2023 for voluntary use, organises AI risk work into Govern, Map, Measure and Manage. Questions such as what a model was trained on can only be answered if the data already has owners, classifications and a known origin. Oil and gas teams can test this with our AI data readiness checklist.
Which mistakes slow data governance down?
- Buying a tool first. A catalogue without owners and rules becomes another unmaintained system.
- Making IT the owner of everything. IT cannot judge whether a lab value or a failure code is right.
- Governing only ERP data and leaving out the historian, the LIMS and engineering documents.
- Writing policies nobody checks, or treating governance as a one-off project. Instruments, tags and people keep changing.
What we have seen at KABAS AI
Our two published projects are in formulation R&D, but both case studies describe data governance problems. At ADFC, decades of formulation experience existed "in various formats and locations", and the Virtual Lab project structured it into a central repository, with sovereign deployment keeping data under full control within Sonatrach's IT environment and complete traceability for audit. At ENAD Shymeca, consistent documentation and traceability across formulation projects was one of the stated challenges, and the work started with a discovery phase on the R&D team's workflows and data structures before historical data was migrated.
Software can enforce decisions once they exist. The Virtual Lab product page lists client-owned data, role-based access control, action logging, backups and data export at any time, and DALILE runs on the client's own infrastructure with encryption, access controls and audit logging. Deciding who owns each dataset, which version is the reference and what quality is acceptable stays with the client, and should come before any model.
Where should you start?
Pick the data domain behind your next AI project. Name its owner and steward, inventory its sources, apply the eight policies above to that domain only and measure two or three quality dimensions. Then extend. An AI readiness assessment helps choose the domain.
Our consulting team runs readiness assessments and helps set up AI governance, and our data science team handles data collection, preparation and structuring. Contact us to discuss your data.
Sources
- CISA, ASD's ACSC et al., Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- NIST, The NIST Cybersecurity Framework (CSF) 2.0, CSWP 29, 26 February 2024
- NIST CSRC glossary: data steward
- UK Government, The Government Data Quality Framework, 2020
- Sambasivan et al., "Everyone wants to do the model work, not the data work": Data Cascades in High-Stakes AI, CHI 2021
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015
- ISA, ISA-95 Standard
- ISO 14224:2016, Collection and exchange of reliability and maintenance data for equipment (abstract)
- IOGP, Joint Industry Programmes (CFIHOS)
- CMS Expert Guide to data protection and cyber security laws: Algeria
- NIST AI Risk Management Framework
Frequently asked questions
What is the difference between data governance and data management?
Governance decides who is accountable for each kind of data, which rules apply to it and how compliance is checked. Data management is the daily work of storing, moving, cleaning and protecting data under those rules. A NIST glossary definition puts it simply: data stewards enforce policies, and functional data managers implement them and manage day-to-day quality.
Should IT own data governance in a plant?
IT should run the systems and implement access rules, but it should not own the data. The owner of process data is usually a production or process manager, the owner of maintenance data the maintenance manager, and so on. These people know what the data means and who should see it. IT and OT administrators act as custodians.
Do we need a data catalogue or governance software before we start?
No. A first inventory can be a shared spreadsheet listing each data source, its system, owner, classification, retention and whether it contains personal data. Buy software once you know which rules you want to enforce and where manual work becomes the bottleneck.
Does data governance cover historian and SCADA data, or only business data?
It covers both. Joint guidance led by CISA treats engineering configuration data and process measurements as data to protect, and calls for strict data governance policies for sensitive OT data used by AI models. Historian tags, units, instrument changes and timestamps need owners and rules like any ERP table.
What does Algerian law require for industrial data?
Law 18-07, as amended by Law 25-11 of 24 July 2025, governs personal data, which plants hold in shift logs, HR files and access records. According to the CMS legal guide, it requires authorisation for transfers abroad, a data protection officer, impact assessments for high-risk processing and breach notification within five days. Non-personal data such as formulations is protected mainly by contracts. Confirm your case with counsel.



