Europe’s 2019 AI ethics checklist put procurement and QA in the governance table

The EU expert group’s April 2019 AI guidelines came with a table saying which functions should use their assessment list. For organizations that buy AI, the table mattered more than the principles, and the AI Act later made deployers’ duties law.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

The EU expert group’s April 2019 AI guidelines came with a table saying which functions should use their assessment list. For organizations that buy AI, the table mattered more than the principles, and the AI Act later made deployers’ duties law.

  1. The week On April 8, 2019, the European Commission’s High-Level Expert Group on Artificial Intelligence presented its Ethics Guidelines for Trustworthy AI.
  2. Decision and escalation: management and the board The table gives top management a role that is easy to describe and hard to do.
  3. Rules and their upkeep: compliance and legal The table gives the compliance, legal or corporate responsibility function a maintenance role.
  4. Evaluation and checking: development, quality assurance and operations Three roles in the table do the hands-on work.
  5. People and purchases: human resources and procurement The last two rows supply the inputs.
  6. Objections: European, non-binding and a box to tick For a US bank, health system or agency, the guidelines had no legal force in 2019.
  7. What a buyer can do with the table today The table can be turned into a short set of changes to existing processes.
  8. How the table lands in each sector The table is generic by design.
  9. The seven requirements, read by a buyer The seven requirements were written for anyone involved with AI.
  10. Contracts that carry the table to the supplier For a buyer, several rows of the table depend on information the supplier holds.
  11. Evidence that the structure works A governance structure is easy to draw and harder to prove.
  12. With hindsight The pilot went ahead.

The week

On April 8, 2019, the European Commission’s High-Level Expert Group on Artificial Intelligence presented its Ethics Guidelines for Trustworthy AI. The document states that it was made public on that date. The group had published a first draft in December 2018 and taken comments through an open consultation. The Commission’s press release the same day said the group consisted of 52 independent experts representing academia, industry and civil society.

The guidelines name three components of trustworthy AI: it should be lawful, it should be ethical, and the third concerns soundness from both a technical and a social perspective, since, the document says, even with good intentions AI systems can cause unintentional harm. They set out seven key requirements: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability. The document says it does not intend to substitute for any form of current or future policymaking or regulation, and describes itself as a living document to be reviewed and updated over time. It says that while it offers a horizontal foundation for AI applications in general, different situations raise different challenges, and that it should be explored whether a sectoral approach is needed in addition, given the context-specificity of AI systems.

The third chapter is the practical one. It sets out a non-exhaustive assessment list, marked as a pilot version, to put the requirements into operation. The chapter says the list particularly applies to AI systems that directly interact with users, and is primarily addressed to developers and deployers of AI systems, whether self-developed or acquired from third parties. It describes two parallel piloting tracks: a qualitative one, with a small selection of organizations from different sectors and of different sizes testing the list in depth, and a quantitative one open to all interested stakeholders. The chapter says compliance with the list is not evidence of legal compliance, and that the list must be tailored to the specific use case and context in which the system operates. It also recommends a governance structure involving both the operational level and top management, with a table of relevant roles: management and board, compliance and legal, product and service development, quality assurance, human resources, procurement and day-to-day operations.

The Commission’s release set out a three-step approach: setting out the key requirements for trustworthy AI, launching a large-scale pilot phase for feedback from stakeholders, and working on international consensus building for human-centric AI. It said members of the expert group would present their work in detail at the third Digital Day in Brussels on April 9, and that the pilot would start in summer 2019, and that companies, public administrations and organizations could already sign up to the European AI Alliance to be notified when the pilot started. It said that following the pilot, in early 2020, the expert group would review the assessment lists, after which the Commission would evaluate the outcome and propose any next steps. The release named healthcare, financial risk management and fraud detection among the areas where AI can bring benefits, and said the Commission wanted to work with partners such as Japan, Canada and Singapore and in the G7 and G20.

The Commission’s press release on the April 2019 guidelines led with the seven requirements, under the heading “Seven essentials for achieving trustworthy AI”. They were quotable, they sounded like principles, and they set Europe’s position in a debate about AI ethics. For a bank, a health system or a public agency, we think the more useful page was later in the document: a short table that said which parts of an organization should use the assessment list and what each should do.

The table matters because of who it addressed. The guidelines say the assessment list is primarily addressed to developers and deployers of AI systems, whether self-developed or acquired from third parties. A bank, health system or agency may buy much of its AI from vendors, embedded in products it already uses. For a buyer, principles about how AI should be built are someone else’s job. What the buyer controls is how its own functions approve, purchase, test, staff and run those systems. The table spoke to exactly that.

Our argument is that the guidelines placed AI oversight inside functions that already exist rather than in a new ethics body, and that this was the part of the document most transferable to organizations outside Europe. We group the table’s seven roles into four parts, and add a fifth section on objections:

  • Decision and escalation: management and the board.
  • Rules and their upkeep: compliance, legal and corporate responsibility.
  • Evaluation and checking: product and service development, quality assurance and day-to-day operations.
  • People and purchases: human resources and procurement.
  • Objections: the guidelines were European, non-binding and in danger of becoming a box-ticking exercise.

The four groups do not overlap: each covers a distinct kind of work, from deciding to rule-setting to checking to supplying people and products. Together they cover every role in the table. The guidelines say the roles depend on the organization, and that the assessment process can be built into existing governance mechanisms or set up as new processes, depending on the organization’s structure, size and resources. We take the first path.

Decision and escalation: management and the board

The table gives top management a role that is easy to describe and hard to do. It says management discusses and evaluates the development, deployment or procurement of AI systems, and serves as an escalation board for evaluating AI innovations and uses when critical concerns are detected. It adds that management involves those affected by the introduction of AI systems, such as workers, and their representatives throughout the process.

Procurement is named at board level

The word procurement appears in management’s own row. The guidelines did not treat buying AI as a lesser case than building it. For a health system that licenses a clinical decision support tool, or an agency that buys a case-triage system, that is the right framing. The decision to deploy is the organization’s, wherever the model was trained.

We think the practical step is to make AI deployments that affect people a category that management reviews, with a short standard paper: what the system does, who it affects, what the assessment found and what remains uncertain. The paper does not need to be long. It needs to exist, so that management has seen the decision before it is made.

A register of systems

Management cannot review what it cannot see. The table assumes the organization knows which AI systems it develops, deploys or procures. For a buyer, that knowledge is harder to assemble than it sounds, because AI arrives through several routes: a direct purchase, a feature in a licensed product, a pilot run by one department, a tool staff adopt on their own. We think the first concrete step for management is to ask for a register: each system, its owner, its purpose, who it affects and when it was last assessed. The register is also what compliance needs to map legal obligations and what quality assurance needs to plan its checks, so one list serves three rows of the table.

Escalation as a defined path

An escalation board is only useful if things reach it. The table connects escalation to the operational roles: quality assurance escalates when results are not satisfactory or unforeseen results are detected, and product development’s logged results are discussed at management level. That gives escalation two sources, a planned review and an unplanned finding. Both need a route, a named recipient and a record of what was decided.

Affected people in the room

The guidelines ask management to involve those affected, giving workers as the example, through information, consultation and participation procedures. For a bank deploying a tool that monitors employee communications, or a hospital deploying one that schedules nurses, the affected people are staff. For an agency deploying a benefits triage tool, they include the public. We think the useful question for management is who is affected and how their view reached the decision, and the useful record is a short note of the answer.

The table gives the compliance, legal or corporate responsibility function a maintenance role. It monitors the use of the assessment list and its evolution to meet technological or regulatory changes, updates internal standards or policies on AI systems, and ensures that the use of such systems complies with the current legal and regulatory framework and with the organization’s values.

The list is not the law

The guidelines are careful about the line between ethics and law. They say the assessment list does not address the first component of trustworthy AI, lawful AI, and that compliance with the list is not evidence of legal compliance. They also note that existing law, such as data protection law, already mandates particular processes, and that internal procedures for data protection compliance may help with ethical data handling too.

For a US organization, the legal layer is its own: fair lending and consumer protection law for a bank, HIPAA and medical device rules for a health system, civil rights law and agency-specific statutes for a public body. The assessment list does not replace any of that. We think compliance’s task is to keep a map showing, for each AI system, which legal obligations apply and which parts of the assessment list overlap with them, so that the same evidence is not gathered twice under different names.

Keeping the standard current

The guidelines called themselves a living document and the assessment list a pilot version. Anyone adopting them in 2019 was adopting something that would change. The table’s instruction to monitor the list’s evolution is the right one: an internal AI standard needs an owner, a review date and a log of what changed and why. Without those, the standard ages while the systems it governs keep changing.

Values as a stated input

The table asks compliance to check use against the organization’s values as well as the law. We think that is unusual language for a compliance function, and an organization may give that part to a different team. Whichever team holds it, the values need to be written down specifically enough to decide a case, or the check cannot be done.

Keeping the assessment records

Compliance also owns the question of how long assessment records are kept and where. An assessment made at purchase may be needed years later, when a decision the system made is challenged or a regulator asks how the system was approved. We think the records should be kept for at least as long as the system is in use plus the period during which its decisions can be challenged, and stored where they survive a change of supplier. The assessment, the quality assurance check, the management decision and the supplier’s documentation belong together, so that a later reader can follow the decision without reconstructing it.

Evaluation and checking: development, quality assurance and operations

Three roles in the table do the hands-on work. Product and service development uses the assessment list to evaluate AI-based products and services and logs all the results, which are discussed at management level. Quality assurance ensures and checks the results of the assessment list and escalates if a result is not satisfactory or if unforeseen results are detected. Day-to-day operations, meaning developers and project managers, include the assessment list in their daily work and document the results and outcomes.

Logging as the common thread

All three rows mention logging or documenting results. The guidelines say elsewhere that it is strongly recommended to log all results in both technical and management terms, so that problem solving can be understood at all levels of the governance structure. For a buyer, this is where vendor dependence shows. An organization that buys an AI system may not have access to the information it would need to answer parts of the assessment list. The log should record those gaps as gaps, with the request made to the vendor and the vendor’s answer.

Quality assurance as a second pair of eyes

The table separates doing the assessment from checking it. Product development runs the list. Quality assurance checks the result. That is a familiar control pattern, the same one that separates a developer from a code reviewer or a preparer from an approver. It is worth applying to AI because the assessment involves judgment. A team that wants to ship a system has a reason to answer the list generously. We think quality assurance should be able to stop a deployment it cannot verify, and the authority to stop should be written down before the first disagreement.

Operations after go-live

The day-to-day row asks developers and project managers to include the list in their daily work. For a deployed system, that means monitoring after launch: checking that the system behaves as assessed, that its inputs have not drifted and that complaints or errors are routed to someone who reads them. The guidelines’ key guidance says that ensuring trustworthy AI is about continuously identifying and implementing requirements, evaluating solutions and ensuring improved outcomes throughout the AI system’s lifecycle. A one-time assessment at purchase does not do that. A monitoring routine with an owner does.

Complaints, errors and incidents

Monitoring after launch needs inputs. Some come from the system: logs, error rates, drift in the data it receives. Others come from people: a customer who disputes a decision, a clinician who overrides a recommendation, a caseworker who notices a pattern. We think an organization should decide where those human signals go for each AI system, and make sure they reach the operations owner as well as the general complaints process. A complaint about a decision is evidence about the system that made it, and the guidelines’ escalation route, from quality assurance to management, works only if that evidence arrives.

Tailoring to the use case

The guidelines say the assessment must be tailored to the specific use case in a proportionate way. A chatbot that answers branch opening hours and a model that flags patients for sepsis risk should not get the same depth of review. We think evaluation teams need a short triage step that sets the depth of assessment by the system’s effect on people, and records the reason. That keeps effort where it matters and gives quality assurance a basis to challenge the depth chosen.

Testing on the buyer’s own data

A supplier’s test results describe the supplier’s test population. A buyer’s customers, patients or citizens may differ in ways that matter. Where it is lawful and practical, we think a buyer should test a system on its own data before relying on it, or run it in parallel with the existing process for a period and compare outcomes. The comparison is evidence that quality assurance can check and management can read. Where local testing is not possible, the assessment should say so, and the monitoring plan after launch should be tighter to make up for it.

People and purchases: human resources and procurement

The last two rows supply the inputs. Human resources ensures the right mix of competences and diversity of profiles for developers of AI systems, and ensures that the appropriate level of training on trustworthy AI is delivered inside the organization. Procurement ensures that the process to procure AI-based products or services includes a check of trustworthy AI.

The procurement row

The procurement row is one sentence long, and for buyers it is the most consequential in the table. We think procurement is the point at which an organization has the most bargaining power over a vendor and the least information about the system. Once the contract is signed, that power falls. If the organization has not asked for documentation, logs, testing results and notice of changes by then, it may not get them later.

We think a procurement check for AI can be short and specific. It asks the supplier for the intended use and known limits of the system; the data it was trained or configured on, at the level the supplier can disclose; the testing the supplier has done and its results; what logs the system produces and how the buyer can reach them; how the supplier will notify the buyer of material changes; and how the buyer can turn the system off. The answers go into the evaluation log, and gaps are recorded as gaps.

Embedded AI

AI can also arrive as a feature inside software the organization already licenses. A records system adds a summarization feature, or a fraud platform adds a new model. Procurement may never see a new purchase. We think the check should also run at renewal and on notice of a material product change, and contracts should require that notice, so that embedded AI gets the same review as a standalone purchase.

The human resources row

The HR row has two parts. Competence and diversity of profiles apply to teams building AI, which in a buying organization may be small. Training on trustworthy AI applies to everyone involved, including the people who will oversee and use the system. We think the second part matters more for buyers. A clinician using a decision support tool, or a caseworker using a triage score, needs to know what the tool is for, where it is weak and what to do when it seems wrong. That training is part of deployment, not an optional extra.

Overrides as a signal

Training also shapes what happens when a user disagrees with a system. A clinician who overrides a recommendation, or a fraud analyst who clears an alert the system flagged, is exercising the human oversight the first requirement describes. Those overrides are also data. A rising override rate can mean the system is drifting, or that users have stopped trusting it, or that it was never suited to the local population. We think training should tell users that overriding is expected when their judgment differs, and operations should count overrides and read a sample of them. An organization that discourages overrides loses its oversight. One that ignores them loses its signal.

Objections: European, non-binding and a box to tick

It was a European document

For a US bank, health system or agency, the guidelines had no legal force in 2019. They were addressed to a European audience and framed around European values and law. Our answer is that the governance table is not specific to Europe. It describes how an organization can spread oversight across functions it already has. That design problem is the same in New York as in Brussels. US organizations with European operations also had a more direct reason to pay attention, as the hindsight section shows.

Non-binding guidance gets ignored

The guidelines said they did not intend to substitute for current or future regulation. Voluntary guidance competes for attention with mandatory work, and may lose. We think the answer lies in where the work is placed. A separate ethics process is easy to drop when budgets tighten. A step inside procurement, quality assurance and operations is harder to drop because those functions keep running regardless.

Checklists become box-ticking

The guidelines anticipated this objection themselves. Their key guidance says ensuring trustworthy AI is not about ticking boxes, and that an assessment list will never be exhaustive. The risk is real. An organization can answer every question on a list and learn nothing. The protection is the separation the table builds in: one function answers, another checks, and management sees what was found. We think a list that someone else will check gets answered more carefully than one that goes into a drawer.

We already have model risk and privacy programs

Some organizations already run model validation, privacy impact assessments and vendor risk reviews. They may see the assessment list as duplication. The guidelines agree that the list can be incorporated into existing practices rather than carried out as a standalone exercise. We think the right move for such an organization is to map the seven requirements to its existing reviews, find the requirements no current review covers, and add only those.

It was too early

In April 2019 the assessment list was a pilot version and the guidelines promised revision. An organization might reasonably have waited for the final list before acting. We think waiting made sense for the detailed questions and not for the governance structure. Deciding which functions own which parts of AI oversight did not depend on the final wording of any question, and the structure could absorb a revised list when it arrived.

An ethics board still has a place

Our argument is not that ethics boards are useless. Some questions are hard enough, or new enough, that a group with broad experience and some distance from the business is the right forum. The table’s management row already allows for that: it describes management as an escalation board for critical concerns. An organization can staff that escalation role with an ethics board if it wishes. What the table adds is everything below the escalation: the routine assessment, checking, purchasing, training and monitoring that decide whether hard questions are noticed in the first place. A board without that structure beneath it sees only what someone chooses to bring it.

What a buyer can do with the table today

The table can be turned into a short set of changes to existing processes. None needs a new committee. The order matters less than the coverage: an organization can start with whichever function is easiest to change, as long as every row eventually has a named task and a record. We would start with procurement and the register, because they determine what the other rows see. A system that enters the organization without passing procurement, and never appears on the register, escapes every other row of the table. Once those two are in place, quality assurance and operations have a defined set of systems to check and monitor, and management has a list to review.

  • Management: add AI deployments that affect people to the decisions management reviews, with a standard short paper.
  • Compliance and legal: keep a register of AI systems with the legal obligations that apply to each, and own the internal AI standard with a review date.
  • Development and operations: run a proportionate assessment before deployment, log the results and gaps, and monitor after launch.
  • Quality assurance: check each assessment and hold written authority to stop a deployment it cannot verify.
  • Human resources: train the people who will oversee and use each system on its purpose and limits.
  • Procurement: add an AI check at purchase, renewal and material product change, with contract terms for documentation, logs, change notice and shutdown.

Each of these changes sits with a function that already has a budget, a manager and a reason to keep running. That is the design advantage the table offered in 2019, and we think it remains the most practical part of the document.

How the table lands in each sector

The table is generic by design. The guidelines themselves say it should be explored whether a sectoral approach is needed in addition to their horizontal framework. We sketch how the four groups of roles map onto the three sectors this library covers. These are our observations, not requirements, and each organization should check the specific obligations that apply to it with counsel.

Banks and other financial firms

A bank may already have much of the table in place for other reasons. Model governance, where a bank has it, can assign development, independent review and approval to separate people, which mirrors the split between product development, quality assurance and management. Vendor risk management may already review third-party products at purchase and renewal. Compliance may already track the consumer protection and fair lending rules that apply to credit decisions. The gap, as we see it, is coverage. A model governance program may cover models that produce numbers used in risk and finance, and miss AI features embedded in customer service, fraud operations or employee tools. The table’s value for a bank is as a check that every AI system, including those that do not look like traditional models, passes through the same four groups.

Health systems

A health system may buy AI inside clinical and administrative software, and its users are clinicians whose judgment the system is meant to support. The human resources row matters more here than elsewhere: the people overseeing the system are the people using it, at the bedside or in the scheduling office, and their training is the oversight. The procurement row matters too, because a health system may have little visibility into how a vendor’s model was built or tested. We think the useful step is to treat clinical AI features like other clinical changes: reviewed by a committee that includes clinicians, introduced with training, and monitored through the same channels that catch other safety problems.

Public agencies

An agency deploying AI affects members of the public who may have no alternative provider and may not know a system was involved. The management row’s instruction to involve affected people takes a public form: notice, explanation and a way to challenge a decision. Procurement in an agency may run through formal acquisition rules, which gives the procurement row a natural home: requirements written into the solicitation and the contract. We think an agency can apply the table most directly by making the AI check part of acquisition planning, so that documentation, logs and change notice are contract terms rather than requests made after award.

What the three have in common

In all three sectors, the organization may be a buyer more than a builder, the affected people are outside the team that chose the system, and the functions the table names already exist. The work is to give each function a specific AI task and a record to keep. That is less dramatic than an ethics board, and we think more durable.

The seven requirements, read by a buyer

The seven requirements were written for anyone involved with AI. A buyer can read each one as a short list of things to ask for or check. We give our reading below. It is not the guidelines’ own assessment list, which is longer and more detailed; it is a buyer’s summary.

  • Human agency and oversight: name who will oversee the system, confirm they can override or stop it, and train them to do so.
  • Technical robustness and safety: ask the supplier for testing results and known failure modes, and plan a fallback for when the system is unavailable or wrong.
  • Privacy and data governance: establish what data the system uses, where it goes and who can see it, and map that to existing privacy obligations.
  • Transparency: confirm the system’s outputs can be traced and explained to the degree its use requires, and that affected people are told when they are dealing with it.
  • Diversity, non-discrimination and fairness: ask how the supplier tested for unfair bias, and test on the organization’s own population where possible.
  • Societal and environmental well-being: record the system’s broader effects that the organization can reasonably assess, including on its own workforce.
  • Accountability: name the owner, keep the logs, and make sure affected people have a route to redress.

Read this way, the requirements sort themselves onto the table. Oversight and training go to human resources and operations. Testing and data go to procurement and quality assurance. Transparency and redress go to compliance and management. The principles and the table were two views of the same work.

Contracts that carry the table to the supplier

For a buyer, several rows of the table depend on information the supplier holds. Development cannot assess what it cannot see, quality assurance cannot check what it cannot test, and operations cannot monitor without logs. The contract is where the buyer secures that information. We describe the terms we think matter, as our view of sound practice rather than as a legal requirement; counsel will shape the actual language.

Documentation

The supplier provides a description of the system’s intended use, its known limits, the conditions under which it was tested and the results. The description is updated when the system changes. Without it, the buyer’s assessment rests on marketing material.

Logs and access

The supplier states what logs the system produces, how long they are kept and how the buyer can reach them, including after the contract ends. Where the supplier hosts the system, the logs may sit on the supplier’s infrastructure. The buyer needs a right to obtain them for monitoring, for complaints and for any review by a regulator or court.

Change notice

The supplier notifies the buyer before a material change to the model, its training data or its intended use takes effect, with enough time for the buyer to reassess. This term is what lets the operations row work after go-live. Without it, the system the buyer assessed and the system it is running can drift apart without anyone noticing.

Incidents and cooperation

The supplier notifies the buyer of incidents and malfunctions affecting the system, and cooperates with the buyer’s investigation of complaints. A buyer that receives a complaint about a decision may need the supplier’s help to understand why the system produced it.

Turning it off

The buyer can suspend use of the system without penalty if it finds a risk it cannot accept, and the supplier supports an orderly exit. The management row gives management the role of escalation board. An escalation board that cannot stop a system has limited authority.

Evidence that the structure works

A governance structure is easy to draw and harder to prove. We think a short set of measures, reported to management on a regular cycle, shows whether the table’s roles are being performed.

  • Coverage: the number of AI systems in the register, and the number found outside it since the last report.
  • Assessment: the number of systems assessed before deployment, and the number deployed without an assessment and why.
  • Checking: the number of assessments quality assurance reviewed, and the number it sent back.
  • Escalation: the issues that reached management, and what was decided.
  • People: whether the staff overseeing each system completed training on it.
  • Suppliers: the number of contracts with documentation, log access and change notice terms, and the number without.

The measures are counts and records, not scores. They tell management whether each row of the table is doing its job without asking management to judge the technical content of each assessment. They also give an internal auditor something to test, which is the last line of separation the table implies but does not name.

We would add one qualitative item: a short note, each cycle, of the hardest case the structure handled and how it was resolved. Counts show that the process runs. The hard case shows whether it works when the answer is not obvious.

With hindsight

The pilot went ahead. The Commission’s library page for the guidelines says that, starting from June 26 (the page gives no year), the assessment list underwent a piloting process open to all stakeholders, with an open survey, in-depth interviews with representative organizations and a channel for feedback through the European AI Alliance. The piloting phase closed on December 1, 2019. Based on the feedback, the expert group presented the final Assessment List for Trustworthy AI, known as ALTAI, in July 2020, which the page describes as a self-assessment checklist for developers and deployers of AI.

The principles then entered the recitals of binding law. Recital 27 of the AI Act, Regulation (EU) 2024/1689, recalls the 2019 Ethics Guidelines for Trustworthy AI developed by the independent expert group, lists the seven principles and says that, without prejudice to the regulation’s legally binding requirements, those guidelines contribute to the design of coherent, trustworthy and human-centric AI. It says the principles should in any case serve as a basis for the drafting of codes of conduct under the regulation.

More important for buyers, the Act gave deployers of high-risk AI systems duties of their own. Article 26, as shown on the Commission’s AI Act Service Desk, requires deployers to take appropriate technical and organizational measures to use such systems in accordance with the instructions for use, and to assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support. Deployers must monitor operation, inform the provider and authorities of risks and serious incidents, and keep automatically generated logs under their control for a period appropriate to the system’s purpose, of at least six months unless other law provides otherwise. Employers must inform workers’ representatives and affected workers before using a high-risk system at the workplace, and public authorities must comply with registration obligations. For financial institutions subject to internal governance requirements under Union financial services law, the Article says the monitoring obligation is deemed fulfilled by complying with those governance rules.

The timing moved. The Commission’s AI Act page, as checked on November 15, 2026, says the Act entered into force on August 1, 2024 and became applicable on August 2, 2026, with some exceptions; that prohibited practices and AI literacy obligations applied from February 2, 2025; and that governance rules and obligations for general-purpose AI models applied from August 2, 2025. It says the rules for high-risk use cases in certain sensitive areas listed in Annex III were extended to December 2, 2027, and those for high-risk AI embedded in regulated products to August 2, 2028, as a result of the AI Omnibus. The page says this proposal to simplify the AI Act was adopted on November 19, 2025, that a political agreement was reached on May 7, 2026, and that it entered into force on July 27, 2026. The page names credit scoring that can deny citizens a loan among the high-risk use cases.

In the United States, NIST released its AI Risk Management Framework on January 26, 2023, intended for voluntary use. NIST’s page says the framework is being revised as part of the White House AI Action Plan, and that on April 7, 2026 NIST released a concept note for a profile on trustworthy AI in critical infrastructure. What held up, as we read it, is the table. The seven principles became a recital. The duty to assign competent human oversight, keep logs and monitor a bought system became law in Europe for high-risk uses. Those duties fall on the functions the 2019 table named.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • AI governance
  • EU AI Act
  • trustworthy AI
  • procurement
  • NIST AI RMF