{"scale":[{"name":"Not in place","level":0,"meaning":"Nothing documented or practiced."},{"name":"Informal","level":1,"meaning":"Practiced by some people some of the time; nothing written or assigned."},{"name":"Defined","level":2,"meaning":"Written, assigned to an owner, practiced consistently, with records."},{"name":"Managed","level":3,"meaning":"Measured, reviewed on a schedule, improved, and able to show evidence to a third party on request."}],"scope":"Organizations that use AI systems (vendor products, embedded features, agents, internal builds) in their operations, whether or not they also build AI.","status":"comment_period","license":"CC BY 4.0","version":"0.1","controls":[{"id":"B1","help":"Someone with authority owns AI governance, there is a regular forum where AI decisions are made, and leadership is kept informed. Example: the COO owns it, a monthly 30-minute AI review with IT, legal, and a business lead, and a quarterly note to the board.","refs":{"eu":"Art. 4; Art. 26","us":"TRAIGA safe harbor","nist":"Govern","other":"Thai draft deployer duties","iso42001":"Clause 5; A.2; A.3"},"title":"Accountability and governance structure","intent":"A named executive owns AI governance; roles and a decision forum exist; the board or equivalent is briefed.","statements":["No one is responsible for AI governance and there is no forum where AI decisions are made.","A person or team handles AI questions as they come up, but nothing is written down or formally assigned.","A named executive owns AI governance under a written charter, a decision forum meets, and minutes exist.","Ownership, the forum, and board reporting are reviewed on a schedule, measured, and can be evidenced to a third party."]},{"id":"B2","help":"A list of every AI system: vendor products (Copilot, ChatGPT), AI features inside other software (the CRM's assistant), internal builds, agents, and tools staff adopted on their own, each with an owner, purpose, data, people affected, and a risk tier. Example: the Workspace inventory.","refs":{"eu":"Art. 26; Art. 50","us":"Colorado SB 26-189","nist":"Map","other":"Korea operator duties","iso42001":"A.4; A.6.2.7"},"title":"AI inventory","intent":"Every AI system in use is listed with owner, purpose, data, people affected, and risk tier, including vendor, embedded, and unapproved tools.","statements":["We do not keep a list of the AI systems in use.","Some teams keep partial lists; vendor and embedded AI tools are mostly unknown.","A maintained register lists every AI system with owner, purpose, data, people affected, and risk tier, updated at least quarterly.","The register is kept current through a discovery process, reconciled against procurement and IT records, and audited."]},{"id":"B3","help":"People know the rules for AI at work and have had training that fits their role; a salesperson and a developer need different depth. Example: an AI acceptable use policy everyone has read, plus the AI literacy course for all staff.","refs":{"eu":"Art. 4","us":"","nist":"Govern","other":"Thai draft oversight-staff duty","iso42001":"Clause 7.2, 7.3; A.4.6"},"title":"Acceptable use and AI literacy","intent":"People know what they may and may not do with AI, and have training suited to their role.","statements":["There are no rules about using AI at work and no training.","Informal guidance exists; some people have had ad hoc training.","A published acceptable use policy and role-based training exist, with completion records.","Training is refreshed on a schedule, tested for effectiveness, and tailored to high-risk roles; completion is tracked and reported."]},{"id":"B4","help":"Before a new AI use goes live, someone checks it for prohibited uses, decides its risk tier, and lists the laws that apply. Example: a one-page screening form completed before a tool is adopted or a chatbot launched.","refs":{"eu":"Art. 5; Art. 6; Annex III","us":"TRAIGA prohibited uses; Colorado high-risk definition; NYC LL144","nist":"Map; Govern","other":"","iso42001":"Clause 6.1; A.5"},"title":"Use-case screening and risk classification","intent":"New and changed AI uses are screened before deployment for prohibited uses, risk tier, and applicable law.","statements":["New AI uses go live without any review.","Some new uses are discussed informally before launch; prohibited uses are not checked systematically.","Every new or changed AI use passes a written screening step that checks prohibited uses, assigns a risk tier, and records applicable law.","Screening outcomes are tracked, sampled for quality, and re-run when systems, uses, or laws change."]},{"id":"B5","help":"Vendor AI is reviewed before purchase and while in use: where the data goes, whether it trains the vendor's models, which subcontractors see it, and what the contract says. Example: the AI Tool Register plus a signed DPA for each vendor that handles personal data.","refs":{"eu":"Art. 25; Art. 26","us":"","nist":"Govern 6; Map 4","other":"HAIP 2.0 deployer questions","iso42001":"A.10"},"title":"Third-party and vendor AI management","intent":"Vendor AI is reviewed for data handling, model changes, subprocessors, and contract terms before and during use.","statements":["Vendor AI tools are bought and used without an AI-specific review.","Procurement asks some AI questions, but there is no checklist or contract language.","A vendor AI review checklist, contract clauses, and a review record exist for each vendor AI product.","Vendor reviews are refreshed on a schedule, model and subprocessor changes are tracked, and findings feed the inventory and risk tiers."]},{"id":"B6","help":"You know what data flows into AI systems (personal, confidential, customer) and control it with settings, policy, and training, with retention defined. Example: training turned off in your AI tools and customer data limited to the approved tool with a DPA.","refs":{"eu":"GDPR; Art. 26","us":"CPPA ADMT rules","nist":"Map; Measure","other":"Korea","iso42001":"A.7"},"title":"Data governance for AI use","intent":"What data goes into AI systems is known and controlled, including personal data, confidential data, and retention.","statements":["There are no rules on what data may be put into AI systems.","People are told to be careful, but rules are unwritten and uneven.","Written data rules for AI use exist, a DPIA trigger is defined, and approvals are recorded.","Data flows into AI systems are monitored, violations are detected and handled, and rules are reviewed against new uses and laws."]},{"id":"B7","help":"Where AI informs decisions about jobs, credit, housing, insurance, health, or education, a competent person reviews, can override, and owns the outcome, and affected people can appeal. Example: a named reviewer for every AI-assisted hiring step, with a documented override.","refs":{"eu":"Art. 14; Art. 26(2)","us":"Colorado; Illinois HB 3773; NYC LL144; CPPA ADMT; TRAIGA","nist":"Manage","other":"","iso42001":"A.9; A.6.2.6"},"title":"Human oversight of consequential decisions","intent":"Where AI informs decisions about people's jobs, credit, housing, insurance, health, or education, a competent human can understand, override, and is accountable.","statements":["AI outputs can drive decisions about people with no human check.","Humans are involved but have no defined authority, training, or record of overrides.","Named reviewers with documented override authority oversee each consequential use, and decisions are sampled periodically.","Oversight is measured (override and error rates), reviewers are trained and assessed, and results drive system changes."]},{"id":"B8","help":"People are told when they interact with AI, when AI materially affects a decision about them, and when content is AI-generated. Example: a chatbot disclosure at the start of each conversation, labels on AI images, a notice in adverse-decision letters.","refs":{"eu":"Art. 50","us":"California bot and transparency laws; TRAIGA government disclosure; Colorado","nist":"Govern; Map","other":"Korea labelling","iso42001":"A.8"},"title":"Transparency to people","intent":"People are told when they interact with AI, when AI materially affects a decision about them, and when content is AI-generated.","statements":["People are not told when they interact with AI, when it affects decisions about them, or when content is AI-generated.","Some disclosures exist but are inconsistent across channels.","Disclosure notices are in place for chatbots and decisions, and a labelling practice exists for synthetic content.","Disclosures are tested for clarity, kept current with each jurisdiction's rules, and audited across channels."]},{"id":"B9","help":"AI systems in use are watched for quality, drift (performance changing over time), misuse, and complaints, and logs are kept for a set period. Example: a monthly review of chatbot transcripts and complaints, with logs kept for twelve months.","refs":{"eu":"Art. 26(5), (6); Art. 12","us":"","nist":"Measure; Manage","other":"Thai draft six-month logs","iso42001":"A.6.2.6; A.6.2.8"},"title":"Monitoring, logging, and performance review","intent":"AI systems in use are monitored for quality, drift, misuse, and complaints, with logs kept for a defined period.","statements":["AI systems are not monitored after launch and logs are not kept.","Some systems are watched informally; logs are inconsistent and retention is undefined.","Each system has a monitoring owner, a review cadence, and logs retained to the longest applicable rule.","Monitoring is automated where possible, drift and complaints are tracked as metrics, and reviews lead to documented actions."]},{"id":"B10","help":"AI incidents (harmful output, data leakage, misuse) are defined, reported internally, fixed, and reported externally where the law requires. Example: an incident procedure that includes AI events, with the Incident Desk duties for your jurisdictions.","refs":{"eu":"Art. 73; Art. 26(5)","us":"TRAIGA cure process","nist":"Manage","other":"Insurer remediation record","iso42001":"A.8.4"},"title":"Incident management and reporting","intent":"AI incidents are defined, reported internally, remediated, and reported externally where the law requires.","statements":["There is no definition of an AI incident and no way to report one.","Incidents are handled ad hoc through general IT or legal channels.","An incident procedure covers AI, a log exists, and the external reporting map (regulators, affected people, vendors) is written.","Incidents are drilled, root causes are analyzed, external reporting is tested against deadlines, and lessons change controls."]},{"id":"B11","help":"AI systems and their data are protected: access control, prompt injection testing for customer-facing AI, prevention of data leakage, and misuse monitoring. Example: the Security Essentials safeguard S16 in place and tested.","refs":{"eu":"Art. 15; Art. 26","us":"","nist":"Cyber AI Profile","other":"OWASP LLM Top 10; CSA AICM","iso42001":"A.6.2.4; ISO 27001 controls"},"title":"Security of AI systems","intent":"AI systems and their data are protected against unauthorized access, prompt injection, data leakage, and misuse.","statements":["AI systems are not covered by the security program.","General security controls apply, but AI-specific threats such as prompt injection and data leakage are not addressed.","AI-specific security requirements are in the security program and customer-facing systems are tested.","AI systems are red-teamed on a schedule, findings are tracked to closure, and controls are mapped to a recognized AI security reference."]},{"id":"B12","help":"Higher-risk uses get an impact assessment, including bias testing, before deployment and periodically after. Example: an assessment of the resume-screening tool covering purpose, data, people affected, bias tests, and mitigations, repeated yearly.","refs":{"eu":"Art. 27; Annex III","us":"NYC LL144; Colorado; Illinois","nist":"Map 5; Measure 2","other":"","iso42001":"A.5; ISO 42005"},"title":"Fairness and impact assessment for higher-risk uses","intent":"Higher-risk uses get an impact assessment including bias testing before deployment and periodically after.","statements":["No impact or bias assessment is done for any AI use.","Assessments are done occasionally, without a method or records.","Completed impact assessments, including bias testing, are on file for each higher-risk system with findings and actions.","Assessments are repeated periodically, use defined metrics and thresholds, and results are reviewed by someone independent of the system owner."]},{"id":"B13","help":"You can show what you did, when, and why: policies, inventories, assessments, training records, decisions. Example: the evidence locker in the Workspace with dated items linked to each control.","refs":{"eu":"Art. 26; Art. 12","us":"AB 1405 record expectations","nist":"Govern","other":"Insurer documentation requests","iso42001":"Clause 7.5; A.6.2.7"},"title":"Records and evidence","intent":"The organization can show what it did, when, and why, to an auditor, regulator, customer, or insurer.","statements":["We could not show an auditor, regulator, customer, or insurer what we did or when.","Evidence exists in scattered emails and documents without owners or retention.","An evidence practice with owners and retention keeps documentation for each system, control, and obligation.","Evidence is reviewed for completeness on a schedule and has been produced successfully for an external party."]},{"id":"B14","help":"At least once a year, governance is reviewed against new laws, new uses, and incidents, and improved. Example: an annual review meeting with minutes and a short list of changes.","refs":{"eu":"","us":"","nist":"Govern 1.5; Manage","other":"HAIP annual update","iso42001":"Clauses 9, 10"},"title":"Periodic review and improvement","intent":"Governance is reviewed at least annually against changes in law, use, and incidents, and improved.","statements":["Governance is never reviewed.","Reviews happen when something goes wrong.","A management review of AI governance happens at least annually with a record and an improvement log.","Reviews use metrics, track actions to closure, and incorporate changes in law, use, and incidents."]},{"id":"B15","help":"If you publish AI-generated or AI-edited media, you label it, keep provenance (Content Credentials) where you can, and verify inbound media when it matters (invoices, identity, press). Example: labels on AI images and a procedure for checking suspicious documents.","refs":{"eu":"Art. 50(4); transparency Code","us":"California AI Transparency Act (providers); platform duties","nist":"","other":"Korea and China labelling; C2PA; India IT Rules","iso42001":""},"title":"Synthetic content provenance","intent":"If the organization publishes AI-generated or AI-edited media, it labels it, preserves provenance where feasible, and verifies inbound media where it matters.","statements":["AI-generated media is published without labels or provenance.","Some content is labelled, inconsistently, and provenance data is lost in publishing.","A labelling and provenance procedure exists, Content Credentials are kept in the publishing chain, and high-stakes inbound media is verified.","Labelling and provenance are checked automatically at publish time, verified against each jurisdiction's rules, and audited."]}],"overallLevels":[{"max":0.99,"name":"Initial"},{"max":1.99,"name":"Developing"},{"max":2.69,"name":"Defined"},{"max":3,"name":"Managed"}]}