Do Assign one named role as accountable for the tool across its life. Document who decides on risk and who runs day to day oversight.
Evidence Owner named in the AI inventory entry.
Sources
Do Present the tool to a group that spans clinical, safety, IT, privacy, compliance and executive roles. Record the decision and conditions.
Evidence Committee minutes or approval record.
Sources
Do Include the tool in the regular board report on AI use, outcomes and adverse events.
Evidence Board report that lists the tool.
Sources
Do Write the intended use, intended users, target population and out of scope uses. State the performance the tool is meant to deliver in this setting.
Evidence Intended use statement or model card on file.
Sources
Do List prospective risks and biases with emphasis on safety and access. Record the mitigations and the residual risk accepted.
Evidence Risk register entry.
Sources
Do Ask the vendor how the tool was validated and tested. Ask whether the vendor will tune or validate on a sample that represents your patients. Ask how bias was evaluated.
Evidence Completed vendor questionnaire and reviewer notes.
Sources
Do Test the tool in the deployment context with clinical domain experts. Tune or test on local data. Record measured performance against intended use.
Evidence Local validation report.
Sources
Do Confirm that testing covers the populations you serve. Report performance by relevant subgroups before go live.
Evidence Subgroup performance table.
Sources
Do Define permitted data uses. Prohibit re-identification. Require security compliance. Reserve audit rights with penalties.
Evidence Executed agreement with the listed terms.
Sources
Do Apply your third party AI risk policy. Write a contingency for failure or incident in the vendor tool, model or data. Document the risk controls applied.
Evidence Third party risk assessment and contingency note.
Sources
Do Protect patient data from unauthorized use or release. Where the tool directly affects care, notify patients and families, explain how their data are used, and obtain consent where relevant. Address transparency for staff.
Evidence Disclosure text, consent workflow and privacy review.
Sources
Do Define which roles may use the tool. Decide whether training is needed before rollout and whether it must recur. Provide role specific documentation.
Evidence Access rules and training record.
Sources
Do Define the roles of people and the AI configuration. Specify who can override or disengage the tool and how an override is recorded.
Evidence Oversight procedure and override log design.
Sources
Do Complete a one page model card: intended use and users, population, warnings, model and data facts, key metrics for usefulness, fairness and safety, and references. Keep it current.
Evidence Model card in the AI inventory entry, with version and date.
Sources
Do Map where the tool enters the workflow. Test usability with real users. Record the benefit, risk and cost case against the current process.
Evidence Workflow map, usability findings and the benefit risk cost summary.
Sources
Do Define fairness for this use case. List the subgroups at risk. Set how bias will be tested before and after go live. Justify any protected attribute used.
Evidence Fairness definition, subgroup list and bias monitoring strategy.
Sources
Do Run the tool in the background on live data without showing outputs. Compare against outcomes and subgroups. Approve go live only after review.
Evidence Silent run report and approval record.
Sources
Do Start with one site or team. Set success criteria for accuracy, reliability, acceptance, cost and workload. Update the risk plan from pilot findings.
Evidence Pilot plan, success criteria and the scale up decision.
Sources
Do Name the metrics, the responsible parties, the review frequency and the escalation route. Cover functionality and behavior, comparison of outputs to known parameters, use case outcomes, confidence and data currency.
Evidence Monitoring plan with named metrics, owners and frequency.
Sources
Do Monitor more often when the tool is close to a clinical decision. Administrative tools can be checked periodically.
Evidence Monitoring frequency recorded against the risk tier.
Sources
Do Track known and emergent risks using intended performance against performance in the deployed context. Document measurable improvements and declines.
Evidence Performance dashboard or periodic report.
Sources
Do Monitor and audit for bias during use as well as before deployment. Report results to the governance group.
Evidence Periodic subgroup performance report.
Sources
Do Give end users and affected communities a way to report problems and appeal outcomes. Capture user input and overrides.
Evidence Reporting channel and appeal procedure.
Sources
Do Report adverse events and ongoing errors to leadership and to the vendor. Record events in the internal incident system. Send locally found issues to the vendor.
Evidence Incident process that names the tool.
Sources
Do Set a process for confidential, blinded reporting of AI safety events to an independent organization. Designate who owns AI quality and safety.
Evidence Reporting procedure and designated owners.
Sources
Do Follow procedures to respond to and recover from a previously unknown risk. Communicate incidents and errors to relevant parties, including affected communities. Record the response.
Evidence Incident response procedure and test or tabletop record.
Sources
Do Count the tasks the tool adds for clinicians and review them with users. This item has no source clause. It is a local consideration.
Evidence Workload measure in the monitoring plan.
Sources
Do Log when users override or ignore the tool. Review the pattern for automation bias and for tool errors.
Evidence Override log and periodic review notes.
Sources
Do Limit access to the intended purpose. Keep audit trails of data access and model changes. Review them on a schedule.
Evidence Access list, audit log and review record.
Sources
Do Track input data and performance against baseline. Set the drift level that triggers review or retraining. Check drift effects by subgroup.
Evidence Drift dashboard, trigger levels and review log.
Sources
Do Have a party outside the build or buy team review the tool and its monitoring. Share the findings with the governing group.
Evidence Independent review report.
Sources
Do Require the vendor to notify you of model changes or updates that may need an unplanned performance check. Revalidate after such changes.
Evidence Contract clause and a revalidation procedure.
Sources
Do Define the performance or outcome departures that trigger suspension. Assign who acts and how the tool is disengaged.
Evidence Suspension criteria and a tested disengage procedure.
Sources
Do Plan how the tool will be phased out safely without raising risk, including what replaces it and what happens to data.
Evidence Decommissioning plan.
Sources
Do Plan periodic review of this governance plan and define the roles for it.
Evidence Scheduled review in the AI inventory.
Sources
This plan maps your answers to requirements in published governance documents. Medon is a planning aid that supports your governance committee and counsel. Each mapping to a source document carries a review status.