Best fit
An institution deciding about many people at once, where appeals exist but do not change the rule. Public benefits is the pattern behind two of the five scored cases.
Public benefits and debt recovery →The standards bind a decision, not a model, so they apply in any sector. They do the most in six, and less in six others.
On this page
An institution deciding about many people at once, where appeals exist but do not change the rule. Public benefits is the pattern behind two of the five scored cases.
Public benefits and debt recovery →A team shipping an agent or an automated decision whose code it controls. The record format is specified, and an exported record stream can be graded today.
AI agents that take actions →An agency or a health plan buying one of these systems from a vendor. It can cite clause numbers in the contract and require the vendor to show, at each release, the evidence that the clauses are met.
Citing the standards in contracts →The framework earns its cost where four things are true.
When all four hold, the casebook pattern can recur: challenges are upheld one at a time while the rule that produced them keeps running. When one is missing, the framework does less. The boundaries below say what.
Each names what the system decides, the scored cases that came from it, the standards to adopt in policy or a contract, and the check to run first.
Whether a person gets a benefit, how much, and whether they owe it back.
Two of the five scored cases are here. In both, people challenged their decisions for years while the rule that produced the errors kept running. The burden of disproof often falls on the claimant, and a wrong debt stands while they challenge it.
The usual defense: "Each wrong decision was reversed on appeal." A reversed decision is not a corrected rule. A claimant who cannot opt out is owed more correction, not less.
Whether a person gets credit, at what limit, and whether they can use the money or the account they already have.
These decisions run in real time, and many take effect before anyone could review them. The framework does not ask them to wait. It asks that each hold come with a receipt and a review deadline, and that access be restored when no fraud is confirmed.
The usual defense: "The fraud model is accurate." Accuracy is evidence for a hold. It is not permission to keep one open without a receipt or a deadline.
Whether to send, refund, write, deploy, or approve, under authority someone delegated to it.
The tooling is furthest along here. Agent decisions are cheap and numerous, and a logging layer may not record them as decisions at all. The standards specify the record each action leaves, treat the agent's authority as a lease that expires, and treat a chain of agents handing work to each other as one delegation.
The usual defense: "The agent passed its evals." Passing evals shows what the agent can do. It still needs a grant that names its scope and expires.
Whether a treatment is authorized or a claim is paid, and how a patient is triaged.
Coverage decisions are frequent, each one is consequential, and a delay is itself a harm. Patients and clinicians already appeal them, so the casebook's question applies directly: does an upheld challenge change the rule?
The usual defense: "Patients can appeal." Appeals that win one case at a time, while the rule stays, show the rule is wrong. They do not show the system works.
Whether a person's post stays up, their account stays open, or their listing stays visible.
Enforcement runs at a volume no team could review case by case, and a wrong suspension can cut off a person's income or audience. Removal often has to be fast. The framework asks that each removal come with a receipt, that exceptions have limits, and that content be restored if the takedown is overturned.
The usual defense: "Users can appeal a takedown." An appeal leads to a remedy only if an overturned takedown restores the content by a deadline.
What grade a student receives, or whether a worker is accused, disciplined, or let go on the system's figures.
Two scored cases are here, and neither system was a machine-learning model. Ofqual's 2020 model was a statistical standardization. Horizon was branch accounting software. The standards apply anyway, because they bind the decision, not the component that made it.
The usual defense: "The system is robust." The Post Office told each defendant that of Horizon. Its supplier kept a log of known defects, which the defendants never saw.
"Law in force" names rules that already ask for the same evidence. It is not legal advice, and meeting a standard here does not show that a system complies with any of them.
Kinds of system where one of the conditions fails, and what still applies.
No single recommendation changes a person's status, access, money, or risk, so there is no decision to challenge. The framework applies where personalization sets a price, a credit offer, or who is shown a job.
The standards evaluate how a model is put to work deciding about people, not the model itself. A model can pass every benchmark and still be deployed under an approval nobody renews or can revoke. Until someone deploys it, there is no one the standards can hold to a clause.
When a named person decides each case, has the time and the information to disagree, and has their disagreements recorded, most of the framework's work is done. What remains is checking that those three stay true as volume grows.
A board that approves ten grants a year reads each one. Receipts, review deadlines, and graded records add cost there without catching what ordinary review would miss.
A build pipeline or an internal dashboard changes no one's status. The exception is an agent with write access to customer accounts or production records. That agent acts on authority someone delegated to it, and the agents context applies.
The person who grants the agent its authority is usually the person its mistakes land on, and can stop it at any time. No institution operates the delegation, so no one answers to a clause for it. What still applies is the delegation: a scope that ends, a stop that works, a record the person can check without the agent, and a way for the people the agent writes to or pays to reach the person. When a vendor runs the same agent for many people, the agents context applies to the vendor.
AI agents that take actions · An AI agent acting for one person
It does not judge an institution's goals. Policy, commercial, and political aims are the institution's to choose.
Whatever the goals, the framework asks for the same three things: approval that expires unless renewed, a reason with each decision, and a named person who can halt the system without penalty.
A scenario missing any of these describes a decision. It does not show one that can be enforced or challenged.
Start with the automated account lock. It runs every step from receipt to remedy, and the other scenarios vary it. The sector comparisons set standard governance beside governance built in.
A tool that checks evidence and a tool that scores answers are different instruments. So is a standard nobody has adopted yet.
The record conformance checker reads the decision records a system exports, checks them against the published schema, recomputes the hashes that link them, and reports the conformance level they earn. The receipt and record formats are published as JSON Schemas that any validator can run in CI.
The delegation audit, the corrective capacity self-assessment, the workload modeler, the three validators, and the 60-second self-test score what a team enters about its system. None of them verifies it.
Every standard is a draft, apart from one that is deprecated. None has force until an organization adopts it or a contract cites it. The case scores are the Institute's reading of what courts, inquiries, and regulators found.
If your system fits, start with one workflow. The delegation audit takes it through six questions and lists the permissions nobody on your team can justify.