Sign in or create a free account to save this Card to your Deck.

Readiness Card · Evidence of progress

Apply risk-tiered AI approval requirements

APPROVAL MUST MATCH THE CONSEQUENCES.

Readiness for an AI deployment is a decision about a specific use case, the information it touches, the actions it can take, and the consequences of error. Leadership and AI calls for risk-tiered approvals, explicit rules for permitted AI use, and human oversight where decisions matter. A low-consequence internal drafting aid may require only routine controls; a system that communicates with customers, influences employment, affects money, or changes sensitive records warrants a more demanding review. This Card states the governing questions, but the supplied manuscripts do not establish a complete company-specific approval matrix, named signatories, or universal numerical thresholds. Those must be approved by the organization before this Card can certify deployment readiness.

Expanded teaching

Expanded teaching

The boundary this Card is meant to mark

The relevant transition is from experimenting with a capability to authorizing it for an operating role. A convincing demonstration is evidence that a workflow can function in a test setting. It says little about whether the system can handle bad inputs, confidential data, unusual users, or consequences that cannot be undone. Approval must address those differences. A business may authorize a narrow internal use while withholding permission for customer-facing actions using the same model.

What changes the level of scrutiny

Consider the severity of an error; whether an action is reversible; whose information is processed; whether confidential or regulated information is involved; whether the system speaks or acts externally; how much autonomy it receives; how failures will be observed; and who is accountable for the outcome. Leadership and AI explicitly contrasts lower-risk activities such as internal research and drafting with higher-consequence uses such as employment, credit, and customer-facing decisions. The examples guide proportional review; they do not establish a universal legal classification.

The minimum information for a release decision

Require a use-case description that states the business purpose and success measure. List the permitted inputs, data classes, and destinations; the exact outputs or actions allowed; the human reviewer or escalation owner; the situations that require stopping; the testing evidence; monitoring responsibility; and a rollback or disable procedure. Specify whether the model only drafts, recommends for approval, or can take bounded actions. The more consequential the use, the less acceptable it is for ownership to remain implicit.

Evidence before authorization

Test normal cases, ambiguous instructions, incorrect records, access attempts, and exceptions relevant to the workflow. Check whether outputs are accurate enough for their purpose and whether a person can recognize uncertainty before action. Verify that approvals cannot be bypassed by a prompt and that activity is logged appropriately. Use permitted synthetic or anonymized cases for initial testing where feasible. A pilot outcome measured only in time saved is incomplete if quality, privacy, or harmful errors have not been examined.

Who decides and what remains open

Executive leadership owns the organization's risk posture and policy. The designated business owner is responsible for the workflow's purpose and outcomes. Technical and appropriate legal, privacy, security, or HR reviewers examine the concerns within their remit. The precise approval roles must come from an approved company policy, not from this manuscript. The source books recommend risk-tiered control but do not assign a fixed executive sign-off rule to each category. Until that policy exists, label proposed thresholds and reviewers as pending, and do not convert this Card into an automatic 'ready' badge.

The first governance artifact

Draft a Risk-Tiered AI Use Case Record for one proposed deployment. State purpose, impact, data, autonomy, human oversight, tests, stop conditions, and decision owner. Classify its risk provisionally and ask accountable leadership to approve the categories, reviewer roles, and evidence requirements. The release decision should record who reviewed what, when, and under which version of the policy. Reevaluate whenever scope, data, autonomy, or users change.

Diagnostic inquiry

Questions to ask, not conclusions

Has an authorized person reviewed this specific AI use case, its data, autonomous actions, foreseeable errors, required human oversight, and evidence of safe operation under an approved policy?

Questions that distinguish possible explanations

  • What could the AI system change, send, recommend, or disclose without an immediate human review?
  • Which errors would be difficult to reverse or harmful to a customer, employee, or the company?
  • Who has authority to approve, suspend, and reevaluate this use under the company's published policy?

Practical application

Practical application

Choose one planned AI deployment. Complete a use-case record covering purpose, inputs, scope of actions, risk, testing, oversight, exceptions, and shutdown. Obtain a company-specific governance decision before any consequential live use; leave unresolved criteria marked pending.

Ask Atria

Talk through this Card

Talk it through with us

If this Card describes your week, a free conversation is the fastest way to see what it means for your company and what to do next.

Book a free conversation