DATA123 Journal · practical guide · DAMA-DMBOK 2.0 Revised · updated 2026-11-10
The data and AI governance wheel
How do you turn data and AI governance into a practice you can measure? The 11 DAMA-DMBOK knowledge areas, our own 12th area for AI knowledge and agents, 130 checkable rules, a RACI matrix that follows you as you read, and concrete KPIs and rollout examples.
In short: where do you start?
Don't start with every rule at once. Pick one critical business metric or data product, name an owner, define what it means and trace its path from source to consumer. Only then extend the controls to other domains.
The minimum core: owner → definition → critical fields → source / lineage → access rights → quality tests → monitoring → incident handling.
What is DAMA-DMBOK, and what is this wheel for?
DAMA-DMBOK (the Data Management Body of Knowledge) is a framework for data management. It helps an organisation agree on the capabilities it needs, from data quality and architecture to security, metadata and BI. It is not, however, a ready-made checklist you can roll out mechanically.
This page keeps the official split into 11 knowledge areas: data governance in the centre, ten related areas around it. The outer ring for AI knowledge, models and agents is our addition, not an official 12th DAMA area. The rules are practical recommendations, not DAMA norms quoted word for word.
In 2026 the current reference is DMBOK 2.0 Revised Edition; DMBOK 3.0 is still being written. So these points are not a literal standard from a new edition.
How is a rule different from a good intention?
“Data must be of good quality” is a goal. “The revenue report is published only when its total for the same period and currency matches the approved source within the agreed tolerance” is a rule. It has an owner, a way to check it and a clear consequence when it fails.
The RACI matrix under the wheel shows the typical roles for the selected area. It changes with the area as you scroll: A accountable, R responsible, C consulted, I informed. It is a starting template; in a real organisation each process needs its own approved version.
In this register every rule has an ID, a requirement, an explanation and a check. The check line can become an SQL test, a monitoring KPI, an audit question or a task for the owner.
A worked example: why do two revenue reports disagree?
One KPI, several areas working together
- Governance (DG): a finance representative approves what “revenue” means, who is accountable for the number and how disputes are settled.
- Modelling (DM) and BI (DW): the fact grain, the invoice or order date rule, currency, VAT and adjustments are written down.
- Integration (DI) and quality (DQ): late records are included, totals are reconciled and duplicates are checked.
- Metadata (MD) and security (SEC): the data's path (lineage), refresh time and the right access rights are visible.
- AI: the agent answers only from the same approved definition and gives a source you can check.
The result: not just another table, but a metric whose value you can explain, reproduce and verify.
How do you apply it in 30 days?
- Week 1. Pick one critical KPI or data product and a business owner; define the most confusing terms.
- Week 2. Map the sources, transformations, consumers, sensitivity class and critical data elements (CDEs).
- Week 3. Add 3–5 automated tests: completeness, uniqueness, freshness and reconciled totals.
- Week 4. Agree on incidents, exceptions and escalation; measure where you stand and decide what to extend next.
Initial priority: the highest impact × likelihood of error × risk of it going unnoticed, adjusted for the cost of doing it. This is a guide for choosing, not a formal DAMA maturity score.
Below is the register of checkable rules. Pick an area with the wheel or the list above; the search at the top filters individual rules.
Our 12th area · AI knowledge governance
AI knowledge and agent governance: what changes when data starts acting?
DAMA-DMBOK explains how to manage data as an asset. But an AI system does more than read it: it picks context, draws conclusions, writes code, calls tools and sometimes starts actions. So here we propose an extra AI knowledge area of our own: not an official part of DAMA, but a practical model you can check.
The key idea: context is a managed product, and an AI result is only a hypothesis until it has been checked against the task. One of the most important control lines runs between what the system knows or reads and what it is allowed to do.
Our six AI governance sub-areas
01 / AI
Value and risk
No AI for its own sake: the decision's benefit, owner, law and impact.
02 / AI
Data and context
Origin, semantics, permissions, RAG quality, freshness and keeping synthetic data apart.
03 / AI
Models and learning
Evaluation sets, generalisation, smaller models and bringing in new data.
04 / AI
Agents and safety
Identity, tool rights, limits on actions, the rule for not acting, delegation.
05 / AI
Experiments and programming
Prompts as code, and a controlled loop: hypothesis → test → decision.
What do we borrow from Karpathy and Sutskever?
Natural-language instructions change how a program behaves, so they need versioning and tests. His autoresearch experiment loop lets an agent change code, run a time-boxed trial, measure a metric and keep only the improvements. Our practical takeaway: give agents freedom to experiment only in an isolated environment with clear brakes.
In a 2025 interview he stressed that models can look strong in evaluations yet be unreliable in new real situations; he also discussed more efficient learning and the possibility of learning after deployment. Our practical takeaway: test new cases, use independent tests, measure real benefit and manage feedback safely.
These two ideas are inspiration, not a governance standard written by Karpathy or Sutskever. The specific controls also draw on the NIST AI RMF, OWASP 2026 and the current EU AI Act timeline.
In practice: one AI agent worth checking
The task: an employee asks the agent, “Why is this month's revenue below plan?”
- Semantics: the agent must use the approved definition of revenue and plan, the period, currency and filters.
- Access: the query must inherit the user's rights; the agent cannot read another department's data.
- Evaluation: we check for a correct SQL / KPI result, a grounded explanation, how it handles not knowing, and latency.
- Actions: the agent can suggest checks, but cannot quietly change the budget or send confidential data.
- Feedback: a wrong answer is logged with its real cause (data, retrieval, semantics, model or tool) and added to the regression tests.
What we measure: task accuracy, the share of unsupported claims, permission breaches, the share of manual fixes and the cost of one correct answer.
How to start rolling it out in 4 weeks
- Week 1. One concrete scenario and an owner; agreed limits for success and for unacceptable errors.
- Week 2. Data sources, rights, context versions, a simple baseline and 20–50 typical test questions (a starting size, not a universal one).
- Week 3. 5–10 hard or unfamiliar questions, prompt injection tests, a refusal rule, limited tools.
- Week 4. A pilot, cost per correct result, an error library, and a procedure for releasing and rolling back new versions.
DG · 10 rules
Data governance
Who decides about data, how those decisions are carried out, and how you can tell whether they work.
- DG-01
Every critical data domain has a business owner and a data steward.
The owner is accountable for decisions and meaning; the steward for day-to-day care. IT may administer, but does not automatically become the business owner.
Check Share of critical domains without an owner or steward
- DG-02
Every material policy requirement has a control that can be checked.
Describe the control owner, frequency, evidence and what happens when the requirement is breached.
Check Share of material requirements without a control
- DG-03
Data decisions are recorded in one place.
Date, decision, who made it, deadline for implementation.
Check Number of decisions past their deadline and not implemented
- DG-04
Exceptions to the rules are documented, time-limited and reviewed according to risk.
State the reason, the person responsible, compensating measures and a review date; 90 days can be an internal guideline, not a universal standard.
Check Active exceptions without a review date
- DG-05
The state of governance is tracked with a small set of KPIs that inform decisions.
For example 5–10 indicators; targets, owners and review frequency depend on the organisation's maturity and risk.
Check Indicators without an owner or a decision they inform
- DG-06
New critical data products pass governance acceptance criteria.
Before publishing, check the owner, classification, purpose, quality criteria and consumer access.
Check Number of critical products published without that check
- DG-07
A rule that can be checked by code is checked by code.
Manual checks stay only for decisions that cannot be automated.
Check Share of checks that are automated
- DG-08
Governance maturity and effectiveness are assessed periodically.
A yearly comparison helps, but what matters most is real incidents, costs and business results.
Check Share of actions from the previous assessment not yet done
- DG-09
Critical data elements (CDEs) are chosen by their business impact.
Register first the metrics and fields that finance, compliance, security or customer decisions depend on.
Check Share of CDEs without an owner, definition or quality thresholds
- DG-10
Governance has a clear path for decisions and escalation.
Decide who approves definitions, access and quality exceptions; record who settles disagreements between departments.
Check Age of unresolved cross-domain disputes
DA · 9 rules
Data architecture
How data travels between systems, layers and consumers, and what drives those choices.
- DA-01
Critical business objects have defined authoritative sources.
The source may differ by attribute or lifecycle stage; record which system is responsible for which data.
Check Share of critical attributes without an authoritative source
- DA-02
The path of critical data flows from source to consumer can be traced.
Document regulated and most important flows first; collect lineage automatically where possible.
Check Share of critical flows without an up-to-date path
- DA-03
Data layers have a clear purpose and access rules.
Bronze/raw is for provenance and recovery, Silver for transformation, Gold/curated for approved use; exploration on raw data is possible with suitable controls.
Check Uncertified sources used for official KPIs
- DA-04
Architecture decisions are recorded as ADRs next to the code.
Context, decision, alternatives, consequences.
Check Number of significant changes without an ADR
- DA-05
Every extra physical copy has a justified reason.
Weigh performance, cost, security, data residency and recovery; a link or a view does not always replace a copy.
Check Duplicated datasets without a documented purpose
- DA-06
Production changes are deployed in a controlled, reversible way.
Separate environments by risk; use CI/CD, versions and logged emergency exceptions (break-glass).
Check Unrecorded or irreversible PROD changes
- DA-07
Systems exchange data through approved interfaces.
API, events, CDC, files or managed database links are chosen by purpose; uncontrolled direct changes in the source are not allowed.
Check Share of integrations without a responsible owner and a contract
- DA-08
Critical technologies have an owner and a lifecycle plan.
An exit or migration scenario is needed according to dependency, vendor and cost risk.
Check Number of critical technologies without an assessed alternative
- DA-09
Critical data products have service level objectives (SLOs).
Define availability, freshness, latency and cost limits; agree them with the consumer.
Check Share of products whose SLOs are not measured
DM · 9 rules
Data modelling and design
How business concepts become structures that can be queried correctly.
- DM-01
The business concepts of a critical data model are agreed with the domain owner.
Confirm objects, relationships and meanings before use; iterative review is fine, prototypes need not wait.
Check Number of critical models with unapproved concepts
- DM-02
Every fact table declares its grain.
“One row = …” in one sentence, verified by a uniqueness test.
Check Number of facts without a declared grain
- DM-03
In a dimensional model, keys and unknown members are defined consistently.
Surrogate keys and UNKNOWN members apply where a star schema needs them; this is not a universal rule for every database.
Check Share of facts with missing or wrong dimension links
- DM-04
Shared dimensions and their mappings are managed across areas.
A shared date, customer or organisation makes KPIs comparable; different business meanings may need separate dimensions.
Check Unaligned dimensions used in official KPIs
- DM-05
The need to keep history is set by how an attribute is used.
SCD Type 1/2, valid-from/to or event history is chosen by the questions that will need answers.
Check Share of history-sensitive attributes without a retention policy
- DM-06
Data types and constraints match the meaning.
Dates as dates, money as DECIMAL, proper time zones, sensible constraints on keys; avoid catch-all VARCHAR.
Check Number of wrong types and undefined time zones
- DM-07
The model is versioned with its DDL, and changes are backward compatible or come with a migration.
A column's meaning does not change without a new version.
Check Number of unversioned schema changes
- DM-08
There is one naming convention, checked automatically.
Patterns for objects, schemas, columns and keys.
Check Number of naming violations in new objects
- DM-09
The semantic model's contract states KPI grain and filter rules.
Revenue by invoice date can differ from revenue by order date; define dimensions and exceptions.
Check Share of critical KPIs without a semantic contract
DS · 8 rules
Data storage and operations
How data is stored, recovered, maintained and removed across its whole lifecycle.
- DS-01
Critical data products have an RPO, an RTO and a tested restore.
How often to test depends on risk; checking only that a backup exists is not enough.
Check Share of critical products without a successful restore test
- DS-02
Retention periods are set by data type and law.
Expired data is deleted or anonymised automatically.
Check Volume of data kept beyond its retention period
- DS-03
Critical data jobs have an owner, monitoring and escalation.
A 15-minute alert may suit emergency flows but not every nightly load; set SLAs/SLOs.
Check Share of critical jobs without working alerts
- DS-04
Loads are idempotent or have a safe re-run path.
Staging, transactions, merge or atomic publishing protect against partial results.
Check Number of unsafe re-runs or partial publications
- DS-05
Capacity, performance and cost are monitored against chosen thresholds.
Thresholds follow real capacity and budget; the same 70% and 90% limits do not suit everyone.
Check Capacity or budget overruns without a response
- DS-06
Infrastructure and configuration are versioned as code.
The environment can be rebuilt from the repository.
Check Number of manually configured components
- DS-07
Significant and recurring incidents get a root cause analysis.
After an incident, record impact, cause, recovery and the preventive action.
Check Number of recurring incidents without prevention
- DS-08
Unused data objects are reviewed periodically.
90 days without reads is a signal for review, not grounds for automatic deletion: check seasonality and dependencies.
Check Share of unused objects that were reviewed
SEC · 9 rules
Data security
Who may see, change or pass on data, and on what basis.
- SEC-01
Data is classified before broad access is granted.
A safe default class applies until the owner confirms restrictions and the data category.
Check Number of datasets without an approved classification
- SEC-02
Access is granted through groups and roles on least privilege.
Personal permissions only temporarily.
Check Number of direct personal permissions
- SEC-03
Access is reviewed according to data sensitivity and user risk.
Sensitive access may be reviewed quarterly, other access less often; leavers lose access immediately.
Check Number of unreviewed or no longer justified accesses
- SEC-04
Passwords and keys live only in a secrets vault.
Code is scanned before every change; a finding is rotated first, then cleaned up.
Check Number of secrets in code and history (target 0)
- SEC-05
Data is encrypted in transit and at rest.
TLS certificate validation is never switched off.
Check Number of unencrypted connections or disabled checks
- SEC-06
Non-production environments follow data minimisation.
Use synthetic, masked or pseudonymised data where possible; justify any real-data exception legally and technically.
Check Number of unnecessary sensitive copies outside PROD
- SEC-07
Access to sensitive data is audited in proportion to risk.
Log access actions without breaching employee privacy or the retention periods of audit logs.
Check Share of critical data access without the required audit
- SEC-08
Passing data outside has a legal and organisational basis.
Check purpose, contracts, vendors, data location and the use of AI tools.
Check Number of unapproved outbound flows
- SEC-09
Data access rights also apply to RAG and agent tools.
Check user and service identity before retrieval and before the result; a vector index grants no new rights.
Check Number of successful unauthorised RAG queries
DI · 9 rules
Data integration and interoperability
How data moves reliably between systems without losing its meaning.
- DI-01
Every critical integration has a data contract.
Schema, semantics, grain, freshness, owner, tolerances, SLA/SLO and the change process.
Check Share of critical integrations without an agreed contract
- DI-02
Structural changes are announced in advance and versioned.
Breaking changes get a transition period.
Check Number of unannounced breaking changes
- DI-03
Unagreed structural changes are caught before publishing.
Critical errors are blocked; tolerable extra fields can be accepted under the contract's versioning policy.
Check Breaking changes that went unnoticed
- DI-04
Incremental loads have a reliable change-detection mechanism.
CDC, watermark or source log; describe how deletes and updates are handled and the lateness window.
Check Number of lost or duplicated changes
- DI-05
Late-arriving data is handled by reloading a window.
The window is set in the contract.
Check Number of differences caused by lateness
- DI-06
Faulty records are never lost silently.
Use a quarantine / dead-letter queue, an owner and reliable reprocessing.
Check Age of unexamined quarantined records
- DI-07
Dependent steps run on the actual result, not just on the job finishing.
Retry transient errors with a limited number of attempts; idempotency stops duplication.
Check Number of jobs wrongly started after a failure
- DI-08
Every run is logged.
Start, end, rows, status, version.
Check Number of runs without a log entry
- DI-09
Source changes are coordinated with data consumers.
Before releasing a new version, check the key reports, APIs and agent queries.
Check Number of consumers broken by a source change
DCM · 8 rules
Document and content management
How unstructured data is managed: documents, files, e-mails and records.
- DCM-01
Every document collection has an owner, a classification and a retention period.
This applies to shared folder structures too.
Check Number of collections missing these fields
- DCM-02
For every document it is clear which version is valid.
Version history may stay, but users and the AI index must see which version is current.
Check Number of documents with several unmarked active versions
- DCM-03
Metadata is filled in at upload, not later.
Type, date, party, related object.
Check Share of documents without required metadata
- DCM-04
Legally significant records are kept with integrity controls.
Immutable records, version history or WORM, depending on the applicable law and document type.
Check Number of legal records changed improperly
- DCM-05
Search and document indexes keep the source's access restrictions.
A recipient must see neither the document nor a quote from it without rights to the original.
Check Permission leaks through indexes
- DCM-06
Critical Excel and SharePoint files have an owner and a managed structure.
Plan migration by risk, error cost and frequency of use, not automatically after 3 months.
Check Number of critical unmanaged files
- DCM-07
Documents are checked and classified before AI indexing.
Remove outdated versions and duplicates; assess sensitive content, legal limits and possible prompt injection text.
Check Share of outdated or disallowed sources in the AI index
- DCM-08
The index is refreshed and purges deleted or restricted documents.
Deleting a document or changing its permissions shows in the search and vector index within the agreed time.
Check Time from a permission change to the index update
RMD · 7 rules
Reference and master data
How the key business objects keep one identity across all systems.
- RMD-01
Key business objects have known authoritative sources.
Customers, suppliers, products, assets and employees; the source of truth may differ by attribute.
Check Share of critical objects without a source map
- RMD-02
Master records are linked through managed identifiers.
Do not demand one physical ID in every system; keep a canonical ID or a mapping table with provenance.
Check Share of unlinked and wrongly linked records
- RMD-03
Rules for detecting and merging duplicates are approved.
A person reviews ambiguous merges; the audit history and, where possible, an undo are kept.
Check Share of improper merges and duplicates
- RMD-04
Reference code lists have an authoritative source and a version.
Examples: ISO 3166, ISO 4217, SI / UN/CEFACT; local exceptions are documented.
Check Number of unauthorised duplicate lists
- RMD-05
New master records are created through a managed process.
The process can be federated as long as validation criteria, accountability and a duplicate check are in place.
Check Number of records created by bypassing the check
- RMD-06
Master data quality is measured monthly.
Completeness, duplicates, validity.
Check Trend of the quality score
- RMD-07
Hierarchies are versioned with validity dates.
Organisation structure, chart of accounts, product groups.
Check Number of hierarchies without validity dates
DW · 9 rules
Data warehousing and BI
How data becomes reports and metrics people can trust.
- DW-01
Official KPIs are calculated from trusted, approved data.
Usually a curated layer or a certified semantic model; a managed direct source can be a justified exception.
Check Share of official KPIs on unapproved sources
- DW-02
Critical KPIs have one agreed business definition.
It may have several technical implementations, but they must give the same result and share filters.
Check Number of KPIs with mismatched formulas or filters
- DW-03
A certified report has an owner, an SLA, KPIs in the glossary and tests.
An uncertified report is clearly marked.
Check Number of certified reports missing these fields
- DW-04
A report shows when its data was last refreshed.
The user can see how fresh the data is.
Check Number of reports without a refresh time
- DW-05
Date logic and financial periods are used consistently.
A shared calendar or documented mappings; turning off Power BI auto date/time is recommended practice, not a universal duty.
Check Number of unaligned time dimensions in official reports
- DW-06
Row- and object-level security is tested where it is needed.
RLS, OLS, roles or separate models are chosen by data sensitivity; forbidden queries are tested too.
Check Number of sensitive models with failing permission tests
- DW-07
Self-service and official content are clearly separated.
Self-service reports can be certified after quality, owner and KPI checks.
Check Number of uncertified reports used as official
- DW-08
Unused reports are reviewed with their owner periodically.
90 days is a signal; before archiving, check quarterly, annual and legally required reports.
Check Number of rarely used reports not reviewed
- DW-09
Critical KPIs are reconciled with the source accounting.
For financial figures, compare totals, periods, currencies, filters and adjustments.
Check Share of KPIs without a recent reconciliation
MD · 8 rules
Metadata management
How you know what the data is, where it came from and who uses it.
- MD-01
Critical data products have a catalogue entry.
Owner, description, classification, source, purpose, refresh frequency and consumers; widen the coverage in stages.
Check Share of critical products without a description
- MD-02
Technical metadata is collected automatically.
Schemas, lineage and usage from the systems; only business meaning is filled in by hand.
Check Share of technical fields filled in by hand
- MD-03
Business glossary terms have a contextual definition and an owner.
“Revenue” can mean different metrics; for each version give the domain, formula and linked objects.
Check Number of ambiguous critical terms without context
- MD-04
The lineage of critical KPIs can be traced to a trusted source.
Include significant transformations, filters, the model version and, where possible, record-level provenance.
Check Share of critical KPIs without sufficient lineage
- MD-05
Description coverage is measured.
Share of columns with a description in the curated layer.
Check Share of described columns
- MD-06
Updating metadata is part of every change.
A change without an updated description is not accepted.
Check Number of changes without a metadata update
- MD-07
Usage metadata is collected responsibly.
Use usage data when cleaning up objects, but do not collect individual users' actions beyond need.
Check Share of objects without usage statistics
- MD-08
AI context sources carry validity and provenance metadata.
For a document or KPI, give the version, refresh time, access rules and limits of reliability.
Check Share of AI sources without a date or provenance
DQ · 9 rules
Data quality
How to prove that data is fit for a specific purpose, and how to watch its quality through changes.
- DQ-01
Data quality thresholds depend on the business purpose.
E.g. payment IDs 100% complete, other fields by the cost of errors; no single 99% norm for everything.
Check Share of CDEs without approved quality thresholds
- DQ-02
Critical data products have automated quality tests.
Key uniqueness, completeness, validity, consistency, freshness and referential integrity, as the purpose needs.
Check Share of critical products without applicable tests
- DQ-03
Critical facts are reconciled with an authoritative source periodically.
Compare totals, counts and other control values within an agreed tolerance; a checksum alone does not prove business correctness.
Check Reconciliation differences above the agreed limit
- DQ-04
Failed quality tests stop or flag publishing according to criticality.
Critical failures block; warnings may let things continue with a visible status and an owner's decision.
Check Critical failures that reached consumers
- DQ-05
Quality errors are fixed as close to the source as possible.
If that is impossible, corrections are marked, versioned and agreed with the data owner.
Check Number of undocumented manual corrections
- DQ-06
Recurring quality incidents are added to regression tests.
High-impact errors get an automated check if one can be stated reliably.
Check Recurring critical errors
- DQ-07
Users see the quality status and its key dimensions.
Do not hide problems behind one “quality score”; show freshness, completeness and check status.
Check Share of critical products without a visible quality status
- DQ-08
The quality issue log has an owner and a deadline for every issue.
An issue without an owner does not exist.
Check Number of issues past their deadline
- DQ-09
Data quality is assessed after changes and after source drift.
Watch not only the schema but also distribution shifts, new categories and unusual totals.
Check Number of undetected critical data drift cases
AI · 35 rules
AI knowledge, model and agent governance
Our own 12th knowledge area: how AI gets meaningful data, how its work is measured, which autonomous actions are allowed and how learning is managed. Not an official DAMA-DMBOK area.
Pick the most important checks for your risk and use case. Every rule has an ID and a measurable criterion.
- AI-01
Every significant AI use case is registered and assessed.
Purpose, owner, data, vendor, model, impact, and the EU AI Act role and risk category where relevant.
Check Share of production AI use cases without a risk assessment
- AI-02
AI agents reach only the data and tools their purpose allows.
A list of allowed sources and rights; raw access is possible for controlled use, but not to get around security and semantics.
Check Number of agents with unchecked access
- AI-03
AI context is managed as a versioned product.
Context sources, metric meanings, terms and access conditions have an owner; documents and model instructions are traceable.
Check Context sources without an owner or version
- AI-04
A knowledge graph is built only when relationships really call for it.
If complex business relationships are needed, every edge has provenance and validity; a simple RAG does not need a graph.
Check Share of unfounded edges and unreliable graph conclusions
- AI-05
AI answers and actions are checked with a versioned evaluation set.
A versioned golden set of real and negative cases; measure factual correctness, retrieval, tool actions, p95 latency, cost and refusal quality separately. Keep the final holdout apart.
Check Share of model, prompt or context changes that were not evaluated
- AI-06
AI answers about facts give grounds that can be checked.
A citation must point to a real source and the right place; having a citation does not prove the claim is supported.
Check Share of unsupported claims or wrong citations
- AI-07
RAG search indexes are maintained as living data products.
Owner, version, refresh, deletions, permission changes and quality checks, across the whole ingest → retrieve → answer path.
Check Outdated, wrongly visible or deleted documents in results
- AI-08
Using data for AI rests on a lawful purpose and minimisation.
Sending confidential and personal data is assessed against the legal basis, vendor terms and suitable safeguards; masking is not always the only measure.
Check Unjustified transfers of sensitive data
- AI-09
AI agent actions are limited by their consequences and reversibility.
Reading on least privilege; write tools only with a clear scope, audit, tests and human approval for high-impact actions.
Check Number of unauthorised write or external actions
- AI-10
Auditing AI interactions keeps traceability without hoarding content.
Record the model, prompt, tool and data version, the action and the result; limit sensitive content by retention period.
Check Share of critical AI interactions without traceability
- AI-11
AI-generated code is held to the same engineering rules.
PR review, tests, licences, secret scanning and security; responsibility stays with the organisation and the developer.
Check Share of AI code changes without these controls
- AI-12
Agents resist prompt injection and untrusted external content.
Documents, websites and tool responses are treated as data, not instructions; test request redirection and data exfiltration.
Check Share of prompt injection test attacks that succeed
- AI-13
Every AI tool has its own least-privilege identity and limits.
Allowed tools, arguments, rate limits, external domains and logs; MCP connections are vetted like third-party integrations.
Check Share of agent tools with unlimited rights
- AI-14
Changes to the model, indexing and instructions are tested before release.
Include past errors, hallucination, citation accuracy, security and business cases in the evaluation set.
Check Share of changes without regression and security tests
- AI-15
Continuous monitoring measures quality, cost and user impact.
Track answer errors, refusals, latency, token cost and attempts at forbidden actions; escalate deviations.
Check Cases above the quality or security threshold that went undetected
- AI-16
AI memory and long-term context have validity and purge rules.
Guard against stale context and memory poisoning; user rights and deletion requests apply to agent memory too.
Check Outdated or unauthorised entries in agent memory
- AI-17
AI decisions have human accountability and an incident stop path.
For high-impact cases: rollback, kill switch, human review and vendor incident handling; human approval must be informed, not a blind button press.
Check Share of critical AI systems without a stop or escalation plan
- AI-18
Every AI solution is compared with a simpler alternative.
Before choosing an agent or an LLM, try SQL, rules, search or a smaller model on the same task. Complexity must bring measurable extra value.
Check Share of AI scenarios without a baseline comparison
- AI-19
The provenance, rights and limits of data used by AI are traceable to the source.
Register the versions, licences, consents, selection and removals of training, fine-tuning, evaluation and RAG corpora. Being able to read something is not the right to use it.
Check Share of key corpora without provenance and permitted purpose
- AI-20
AI is tested on new scenarios, not only familiar ones.
Test new periods, unfamiliar departments, format variants, languages and unusual questions separately. A good internal score does not prove generalisation.
Check Quality drop on new or dissimilar cases
- AI-21
Independent AI evaluation data is protected from contamination.
Keep a separate closed holdout set; do not tune instructions, the retriever or the model on the same final tests. Version the cases and their provenance.
Check Share of test sets with detected overlap or leakage
- AI-22
Models are chosen by the trade-off between task quality, cost and speed.
For each scenario compare at least two options, including a smaller or local one where suitable. Measure real task success, p95 latency and the full cost per request.
Check Share of scenarios without a repeatable cost–quality comparison
- AI-23
AI has a clear rule for not knowing and handing over to a person.
Convincing text is not enough: when sources are missing, rights are lacking or the result is contradictory, the agent asks for clarification, declines or hands over to a person.
Check Share of false confidence and critical cases not handed over
- AI-24
Every AI experiment has a goal, a baseline and a reproducible protocol.
Record the hypothesis, data / model / code versions, an equal budget, the metric set in advance, the result and the decision to keep or revert.
Check Share of experiments without a reproducible record or baseline
- AI-25
Autonomous AI experiments run isolated and with limited resources.
Karpathy's autoresearch principle: an agent may propose, try and discard changes, but not change PROD on its own. Set a budget, a time limit, allowed files and a shutdown.
Check Number of runs over budget or with disallowed changes
- AI-26
Natural-language instructions, prompts and tool schemas are versioned as code.
The practical lesson of “Software 3.0”: an instruction is configuration that changes behaviour. PRs, diffs, tests and rollback apply to prompts, agent roles, tools and system context.
Check Share of production behaviour changes without a version and regression
- AI-27
Feedback on AI answers becomes verified learning examples.
Keep user signals, expert-confirmed corrections and model-generated labels apart; mark unclear cases and keep an error library.
Check Share of corrections without a reliable label or decision history
- AI-28
Synthetic data is kept apart from real-world observations.
Every synthetic set has a generation method, a purpose and a label; assess distribution differences and model feedback loops.
Check Share of unlabelled synthetic data in AI training or evaluation sets
- AI-29
Continuous model or context learning has a safe release path.
New data does not automatically become new truth: human feedback is checked, regression is run, a shadow/canary trial is held and a rollback is ready.
Check Share of learning or re-indexing releases without regression or rollback
- AI-30
AI quality is measured by the outcome of the real task, not just a test score.
Assess whether the user got a correct solution, saved time and caused no error; do not equate a high benchmark with real value.
Check Real task success rate and the cost of corrections
- AI-31
Structured AI outputs are validated before being stored or used for a decision.
SQL, JSON, financial figures and action arguments are checked against the schema, referential integrity, business rules and permissions; do not execute just because the answer is well-formed.
Check Number of wrong arguments or results that reached execution
- AI-32
Delegation between agents does not widen authority.
Passing a task to another agent passes only the necessary context and narrower or equal rights; check the chain of accountability and loop limits.
Check Share of delegated actions without a confirmed identity and limits
- AI-33
For high-impact AI decisions, the distribution of errors is assessed.
Where relevant, check for systematic errors across user groups, periods and departments; keep a legal basis and avoid collecting sensitive attributes beyond need.
Check Number of uninvestigated significant differences in outcomes
- AI-34
Changes in external model and agent dependencies are managed as supply chain risk.
Record model / API / MCP / library versions, change notices and contractual data use terms. A vendor change triggers selective tests.
Check Share of critical changes without an impact check
- AI-35
RAG retrieval quality is assessed separately from text generation.
Separate finding the source (recall@k), how well the answer is supported and final correctness; then it is clear whether the error is in retrieval, the document or the model.
Check Share of retrieval misses and unsupported answers
Nothing found. Try another word.
What do real organisations most often lack?
Not more documents, but three links: who is accountable for the meaning, what the result rests on and what happens when it is wrong. Tools can collect lineage, a catalogue or test logs automatically, but people still have to answer for the definition of “revenue” and the financial risk they accept.
What changed with AI agents?
A data consumer used to look at a report. Now an agent can search, call tools, run SQL and start an action. A quality score or a citation is no longer enough: you need to check where the context came from, the user's rights, the limits on the agent's tools, prompt injection attacks and whether the output is correct. A RAG search index must also reflect deleted documents and changed access.
The AI controls draw on the NIST AI RMF Generative AI Profile, the OWASP Agentic Top 10 and ISO/IEC 42001. The EU AI Act applies according to a system's purpose, role and risk category; not every BI assistant is automatically a high-risk AI system.
Frequently asked questions
Does a small organisation need every rule?
No. Start with critical data, an owner, clear KPIs, access and 3–5 quality tests. Add depth to the controls according to impact and risk.
Are Bronze → Silver → Gold layers required?
No. It is a useful architecture practice, but not an official DAMA obligation. What matters is tracing the source and clearly separating unchecked data from approved data.
Does the data owner have to come from IT?
Usually not. DMBOK 2.0 Revised stresses a business person accountable for decisions in a specific data domain; IT can carry out the technical duties.
Does an AI agent need a knowledge graph?
Not always. For many tasks a maintained semantic model and document search are enough. A graph is worth it when you really need to reason about complex relationships between objects.
Is this list an official DAMA standard?
No. DAMA provides the framework of knowledge areas. The rules here and the 12th AI area are a separate, practical application of it.
How is our AI area different from MLOps or LLMOps?
MLOps and LLMOps mostly describe how models are built, released and monitored. This extra area brings together data origin, semantics, rights, agent actions, experiments, risk and the result of the real task.
Does Karpathy's autoresearch mean AI should be allowed to change production on its own?
No. In his example the agent experiments with training a model on a limited budget and improving one metric. For production systems we recommend an isolated environment, independent evaluation and an approved release.
Is a single AI accuracy percentage enough?
No. Even 95% answer accuracy can be unacceptable for critical financial decisions. Keep apart retrieval quality, how well claims are grounded, the safety of actions, new cases, p95 latency and the cost of an error.
Sources and method
The sources were used to check the structure of the knowledge areas and the current context of AI governance. The deadlines, controls, KPIs and examples are practical recommendations, not rules quoted word for word from these publications.
- DAMA-DMBOK: the official knowledge areas and the 2.0 Revised changes.
- DMBOK 3.0: status of the work (2026).
- NIST AI RMF: Generative AI Profile (2024).
- OWASP LLM Top 10 and OWASP Agentic Top 10.
- ISO/IEC 42001:2023: AI management system.
- European Commission: the AI Act and when it applies.
- Andrej Karpathy, “Software Is Changing (Again)”, 2025, and the autoresearch project, 2026.
- Ilya Sutskever interview: generalisation and learning, 2025-11-25.
- OWASP: the 2026 LLM Top 10 and the Agent Control Standard.
A practical interpretation of DAMA-DMBOK 2.0 Revised. The rule wording is our own recommendation, not official DAMA text. The AI area is our independent practical method, not an official DAMA standard. Last reviewed: 2026-11-10.
