Governing the decision: evidence, authority and AI agents

Executive summary

An AI agent can compare thirty communities in moments. That does not mean the thirty communities are describing the same operating condition, and it does not mean anyone has authorized what the agent suggests.

Senior housing & care organizations are adding AI to staffing, occupancy, expense and reporting workflows. The harder question is institutional: when a consequential recommendation is made, which evidence governed it, who had authority to approve the action, and what proves the action actually happened?

Capability and authority are different questions. A more capable model improves analysis; it does not decide which conflicting record governs or hold accountability for a resident-care or workforce decision. Portfolios rarely share meaning by default. Mathematically sound comparison can sit on inconsistent inputs.

A recommendation is not permission. The agent proposes; an accountable person determines within defined limits and for a defined period. Execution is a separate fact that needs its own evidence. Preserve source evidence, surface disagreement, route it to authority, retain the determination, constrain downstream use and verify execution.

Chief executives should start with one recurring decision, not an enterprise automation promise. Boards should test boundaries, not only successful answers. Every provider claim—including ours—needs its evidence, scope, date, exclusions and acceptance gate attached.

This paper is written for readers who may never use SeniorCRE. Its questions apply whether an agent comes from an incumbent vendor, an internal team or a new platform. A provider’s brand or a model’s benchmark does not settle them.

1. The question capability cannot answer

Better intelligence makes a recommendation more persuasive. It does not make it authorized. Model confidence never creates organizational authority.

On June 2, 2026, Microsoft published a feature describing Majorana 2, its next-generation topological quantum chip. Microsoft reports a 1,000-fold improvement in qubit reliability, a mean lifetime of 20 seconds and instances lasting up to a minute. It credits Microsoft Discovery’s agentic AI with helping manage workflows, automate measurements, optimize fabrication and identify previously unnoticed flaws, including an uncalibrated temperature sensor.

These are Microsoft’s reported research claims, not independently verified findings by SeniorCRE. The feature describes agents offering guidance with scientists in the loop, and research that predates the agents. Microsoft’s expectation of a scalable quantum computer by 2029 is forward-looking, not an achieved outcome.

We cite the announcement as a prompt, not as validation of our architecture. Microsoft does not endorse or validate SeniorCRE. Our interpretation is narrow: even where an agent notices flawed evidence, the organization still decides what to do with the finding. Greater capability makes evidence, permission and accountability more—not less—important.

Ask three separate questions. Evidence: which original records support the conclusion, including contradictions and missing information? Authority: who was permitted to determine or authorize, for what purpose and within what limits? Accountability: can management reconstruct the recommendation, approval, response and later correction without filling gaps silently?

A confident recommendation is not an approval. An approval is not evidence that the action occurred. A retained record is not proof that the decision was appropriate. Keeping those distinctions visible is the discipline this paper describes.

2. Why the evidence disagrees

The portfolio is described in one language by its dashboards and in thirty by its communities. One site counts agency hours when an invoice arrives; another when attendance is verified. One measures occupancy at midnight; another reports the day’s average. A recommendation can be mathematically coherent while the operating inputs remain institutionally inconsistent.

These differences are not necessarily errors. Each can be legitimate for payroll, billing, clinical staffing or owner reporting. The governance problem is that nobody has said which definition and source govern a given decision. Disagreement is then resolved silently by whichever system someone opens.

Consider one synthetic shift: the approved staffing plan calls for twelve caregivers, scheduling shows twelve assigned, timekeeping shows ten clocked in, and an agency invoice records one contracted caregiver. The illustration assumes attendance is separately verified: eleven physically present against twelve required, one short. An invoice alone would not prove presence in a real case.

Scheduling records assignment, timekeeping records payroll attendance, the invoice records billing and the plan records requirements. Preserve all four assertions. A governed answer requires purpose-specific source authority, a designated person to resolve the conflict and evidence of what happened next.

Portfolio oversight needs common meaning without erasing local context. Standardize the figure that crosses the boundary between community and corporate, while preserving resident needs, qualifications, local requirements and immediate clinical judgment. A common dashboard must not override care decisions at the bedside.

3. A recommendation is not permission

Suppose an agent recommends transferring personnel among communities to reduce overtime. Its proposal may depend on outdated acuity or unverified coverage. The recommendation itself should not authorize the transfer. An accountable manager reviews care and workforce conditions; any authorization has a defined scope, effective period and retained rationale.

Permission has two halves: what the agent may read and for what purpose, and what it may initiate, recommend or merely flag. Both require explicit grants. If a required grant is absent, the agent should refuse or escalate rather than proceed on a reasonable-sounding assumption.

No silent merging: do not average or pick a favorable value to clear a conflict. No algorithmic judgment on authority: a designated person applies the written rule. Human accountability stays human: software does not assume professional or legal responsibility.

Keep decision states distinct: unresolved, approved, executed, verified and superseded. Each label must correspond to an actual record. A completed-workflow indicator must not disguise a missing approval, and an approval must not disguise an action that never occurred. Show a missing piece as a gap with its owner and next review.

4. A governance sequence from evidence to execution

The six-step executive operating sequence is Evidence → Disagreement → Authority → Governing Record → Intelligence → Execution. It is not a replacement for SeniorCRE’s separate seven-layer technical architecture or its proof sequence. Each step has a different job, and skipping a step does not create authority.

Evidence preserves originating records and context. Disagreement identifies conflicts without overwriting them. Authority routes consequential exceptions to the accountable person. The Governing Record retains the determination, supporting evidence, purpose, scope, rationale, effective period and supersession. Intelligence reads under explicit grants. Execution requires authorization and separate evidence of destination effects.

Coherence before Intelligence. Operating facts should be governed and reconciled under accountable rules before analytics or AI consume them. A consistent-looking dashboard is not sufficient if its source definitions and permissions remain unsettled.

The supplied paper uses “sealed” to describe intended decision-history retention. That is a design requirement, not a claim of implemented controls. Corrections and superseding decisions must remain traceable rather than erasing the prior determination.

A retained determination showing a known shortfall is only half a control if nothing shows a response. Keep the records; pair the determination with execution evidence. Where evidence is missing, record the gap and accountable owner. Do not estimate, backfill or let software average a missing number. Counsel should review any policy: record retention does not guarantee legal protection or admissibility.

5. A synthetic portfolio case study

This invented case shows method, not product capability or a staffing instruction. A thirty-community operator asks a portfolio agent where overtime is rising and whether people could move between sites. Community A is two short on nights; Community B has two above plan with clock-ins matching its schedule; Community C appears above plan but has three call-outs, two agency shifts and two recent admissions that increase acuity.

The agent proposes moving four caregivers from C to A and B for two weeks, relying on the schedule alone. This is deliberately an unsupported recommendation: the apparent surplus does not establish four transferable people, and the other records contradict the premise. Persuasive presentation is not adequate evidence.

A governed process lays the sources side by side. C’s apparent surplus is an artifact of assignments that do not establish present coverage or clinical appropriateness. The clinical coverage conflict goes to the Director of Nursing; the cross-community transfer goes to the chief operating officer. Neither the agent nor a convenience rule resolves the disagreement.

The accountable people decline the move from C, approve two from B to A for one week only after professional review of coverage and qualifications, and refuse the remaining proposal. Each determination records sources, approver, role, scope, end date and rationale.

Execution remains separate: sign-in and timekeeping show arrival and work; the clinical record shows the relevant reassessment. Missing evidence means a gap, not a completed transfer. Approval expires at the stated end date. Later evidence can supersede it while leaving the earlier determination reconstructable.

The board sees a declined C transfer, a verified B-to-A transfer supported by separate evidence, and an unresolved overtime trend with an owner and review date. Refusal and partial approval are useful results. The case establishes no measured saving, clinical outcome, risk reduction or implemented SeniorCRE capability.

6. For the chief executive

Begin with one recurring decision: a staffing exception, occupancy figure or expense variance. Define the field, period, community, intended use and accountable owner. Preserve original assertions and name unresolved disputes rather than selecting the convenient value.

The first governance register asks six questions: what may the agent read and why; which source and definition govern each input; what triggers refusal or escalation; who may approve within what limits; what confirms execution and who checks it; and how corrections and superseding decisions are retained.

The chief executive need not approve every transaction. The responsibility is a clear delegation structure that sends consequential exceptions to the right owner. Community leaders manage immediate adjustments within scope; operations and clinical roles handle exceptions within their grants; executive escalation addresses material exceptions above those limits.

Recorded decisions and verified actions are different measures. Judge the process by whether it preserves disagreement, applies the real authority boundary and distinguishes approval from occurrence. Failed tests belong in the findings. Do not remove them to improve the report.

7. For the board

An elegant AI-assisted report can contain accurate calculations while omitting the source period, reconciliation, permission or verification behind the action. Board oversight should distinguish analytical output from authorized institutional action.

A useful packet includes the decision category, responsible executive, source period, unresolved disagreement, applied authority boundary, response status, closure evidence and any gap with an owner and next review. Boards generally need proportionate control evidence, not unrestricted resident or employee records. Visibility does not grant broad sharing permission.

Test the boundary, not just the successful answer. Exceed the grant: present conflicting records and an out-of-scope instruction, then inspect refusal or escalation. Revoke permission: verify that the agent stops while earlier authorization remains reconstructable. Correct evidence: check that the old determination and its superseding record remain distinguishable and linked.

These are proposed exercises, not tested SeniorCRE production controls. Human review must be meaningful: relevant evidence and expertise, the ability to refuse, escalation paths and retained overrides. A rubber stamp does not cure unsupported evidence or relieve professionals of responsibility.

8. Evaluating claims with their limits attached

Ask every provider to distinguish Designed, Built, Validation-exercised, Operator acceptance and Measured outcomes. Request the evidence, scope, date, exclusions and acceptance gate for every material claim.

Designed supports a specified architecture and an evaluation method, not operation or safety. Built establishes components exist, not end-to-end real-world behavior. Validation-exercised supports behavior on controlled or synthetic cases, not portfolio readiness. Operator acceptance supports one scoped deployment against agreed criteria, not general outcomes. Measured outcomes require a stated baseline, period, limits and confounders; they do not guarantee another organization’s result.

As of October 21, 2026, the Operator-Controlled Operating Record is designed and not yet implemented in any community. The higher evidence levels require actual retained proof. A page, diagram, interface or completed demonstration cannot substitute for it.

9. Objections executives raise

“We already have dashboards.” Keep them. The governance question is which source and rule produced the value, who approved an exception and what happened next. A dashboard can consume governed determinations without becoming the authority owner.

“Human approval will slow us down.” Routine agreement can follow operator-written rules; consequential disputes need the right review. The burden and turnaround must be measured rather than assumed away. “Approvals become rubber stamps.” They can; test reviewers’ access to contradictions, ability to refuse and retained overrides.

“Our vendors say their agents are safe.” Ask for evidence of each authority boundary, not a brand or benchmark. “This is more paperwork.” Treat capture as part of the decision design, and test its burden. “Records create liability.” Counsel should shape policy; fewer records are not a substitute for a documented response.

“We cannot standardize across communities.” Standardize what crosses the reporting boundary while preserving legitimate local meaning and clinical judgment. “Why not wait?” Evidence and permission questions apply to the agents already in use, independently of adopting a new platform.

10. Where SeniorCRE fits, and where proof remains open

SeniorCRE® is developing operator-controlled operating infrastructure for senior housing & care. Its thesis is operator-held authority over evidence, determinations and downstream use—not vendor ownership of organizational decisions. The operator-held control plane defines what agents may read, act on and escalate.

Deployment is the operator’s choice. Road 1 retains existing systems and governs what they produce. Road 2 selects SeniorCRE itself for defined system-of-record functions, including its intended EHR/eMAR configuration after applicable gates clear. Replacement is neither required nor prohibited; neither road implies current production deployment.

The Operating Record is designed and not yet implemented in any community. Native clinical configuration is unavailable for production; medication administration and eMAR are roadmap, not built or deployed. Reliance requires scoped implementation, function-specific safety review, end-to-end validation and operator acceptance. This paper claims no operational, clinical, legal or financial benefit, compliance assurance or endorsement.

Begin with or without SeniorCRE: choose one decision, write the six-question register, reconstruct one recent recommendation and record what you cannot produce, then run the three boundary exercises. Ask each provider to place claims on the evidence scale. An Executive Briefing starts from your existing controls, not assumed product fit.

Appendix · Board checklist

For each question, record what management has shown, what is absent, the gap owner and the next review date. Do not infer completion. The downloadable public edition includes the meeting worksheet, chief executive questions and glossary.

1. Which consequential decisions now involve an AI agent or automated recommendation?
2. For each decision, who is the accountable executive?
3. What may each agent read, and where is that grant written?
4. Which source and definition govern each input, and who wrote the rule?
5. What makes an agent refuse or escalate, and to whom?
6. Who may approve each action, within what limits and for how long?
7. Does the packet separate unresolved, approved, executed, verified and superseded?
8. Where a status says complete, is there evidence of approval and execution?
9. Can management reconstruct one recent decision end to end today?
10. Have boundary exercises been run, and what failed?
11. Can reviewers see conflicting records, refuse and retain overrides?
12. Where does each provider claim sit on the evidence scale, with date, scope and exclusions?
13. What privacy, professional and legal limits govern supporting records?
14. What remains an assumption, and who owns each evidence gap?