Skip to main content

AI Settings for Clinics: Doctor Control, Permissions and Fallback

Define AI controls by task, active clinic, role and encounter, and make the effective setting visible before use. Separate configuration, generation and clinical approval. Test clinic switches, stale sessions, disabled tasks and late results against a written policy. Turning off new requests does not necessarily cancel existing work, stop every capture path or remove accepted records.

By Bractis · Updated · 15 min read

Independent clinic owners, visiting doctors and small Indian clinic groups evaluating controllable AI documentation workflows across locations and staff roles.

Define the task that each setting controls

An AI switch may enable dictation, ambient recording, OCR, history summaries, translation or suggestions. These tasks use different inputs and create different outputs. Before configuring the product, list the tasks the clinic intends to evaluate and the specific boundaries for each. An organization approving one task should not accidentally enable every feature with an AI label.

Separate availability from action. A clinic administrator may make a drafting feature available to a doctor; the doctor still chooses whether to use it for an appropriate encounter, follows the patient-information process and reviews the result. Feature availability should not be mistaken for blanket recording permission or automatic clinical acceptance.

NIST’s voluntary generative-AI profile discusses governance across the AI lifecycle. The control matrix here is an original purchasing and operating aid, not a prescribed regulatory configuration or evidence that any product already implements each control.

Source: NIST AI 600-1: Generative AI risk management profile

Make clinic and clinician preferences resolve predictably

A visiting doctor may work at two clinics with different approved workflows. One may permit doctor dictation while another is still evaluating the provider’s terms. Ask which setting wins when organization, clinic and doctor preferences differ, and whether the effective value is visible in the encounter.

Write the intended precedence in plain language before asking for a demonstration. For example, a doctor preference should not silently override a clinic restriction. This is a proposed policy choice to verify, not a claim about Bractis’s current implementation. If the product uses a different model, assess whether it can express the clinic’s required boundaries.

Test switching locations in the same session and signing in on a second device. A remembered preference from Clinic A should not obscure Clinic B’s actual state. The doctor needs to see the active clinic and task permission before capture, rather than discover the inherited setting after a transcript has been produced.

On a narrow screen, scroll the table horizontally to read all columns.

A proposed AI control map to verify in the actual product
Control layerDecision it should make clearWhat it should not imply
Organization or clinic policyWhich tasks and data paths are permitted hereEvery clinician must use AI for every encounter
Staff roleWho may configure, start, edit or review each taskAdministrative access equals clinical approval
Clinician preferenceWhich permitted task the doctor chooses to useA preference overrides the clinic’s restrictions
Encounter choiceWhether the task is appropriate and explained for this visitA previous encounter choice covers all later uses
Record approvalWho accepts the clinical content and its destinationsGeneration alone creates an approved note or prescription

Test permissions with the people who will hold them

Use separate fictional role accounts in the demonstration, such as receptionist, doctor and clinic administrator. Check the actual allowed actions, not merely whether a menu item is hidden. A staff member who uploads a scan may need to organize the source without accepting an extracted diagnosis or changing the clinic’s AI processor settings.

Identify sensitive configuration actions: enabling a new capture mode, changing retention-related options, connecting a provider or expanding the source data available to summaries. Assign those decisions to an accountable owner and record the reason for change. Avoid a shared administrator login that makes later reconstruction difficult.

Clinical review also needs a distinct boundary. Test whether an unreviewed draft can reach a signed note, prescription or patient message through another route. If the application cannot enforce the intended boundary, document the limitation and decide whether the proposed use is appropriate rather than relying solely on a staff reminder.

Worked example: the same doctor has different effective permissions

K-08 is a fictional two-clinic configuration exercise. The organization makes a documentation service available, Clinic A permits doctor dictation, and Clinic B has not approved that task. A visiting doctor prefers dictation. Reception can organize uploaded sources, while a designated administrator can change the permitted configuration. Neither administrative action is clinical acceptance.

Write the expected action for each role and location before the vendor demonstrates it. In this proposed policy, organization availability, clinic permission and role permission must all permit the requested task; the doctor’s preference cannot widen those permissions. The relevant encounter explanation and choice still need their own process. This is a clinic policy model to evaluate, not a description of Bractis’s current authorization rules.

An effective setting should explain its scope and reason. “Dictation unavailable in Clinic B under its current policy” is more useful than a disabled microphone with no context. If the product cannot determine the active policy after a session or network change, the clinic needs an explicit unresolved state and a manual documentation route, not an invented affirmative setting.

During the demonstration, perform the attempted action with each permitted test role. Hiding a configuration menu is not sufficient evidence that a role cannot use another route to make the change. The vendor should demonstrate enforcement appropriate to the product, using authorized fictional accounts, and explain how stale sessions are handled. This is a purchasing test request, not authorization to bypass controls on a live system.

On a narrow screen, scroll the table horizontally to read all columns.

K-08: proposed action expectations for two fictional clinics
Role, context and requested actionRelevant boundaryExpected result under this proposed policy
Doctor in Clinic A starts permitted dictationClinic and role permit it; encounter prerequisites have been addressedDeliberate start is available; content still needs clinical review
Same doctor in Clinic B starts dictationClinic B has not approved this taskDoctor preference does not enable the task here
Reception organizes a source, then attempts to accept its clinical draftSource organization and clinical acceptance are different actionsOrganization remains available; clinical acceptance requires its authorized reviewer
Administrator enables an approved task, then attempts to sign a noteConfiguration authority does not establish clinical signing authorityEach action is evaluated against its own permission
Doctor opens a session whose current policy cannot be resolvedThe effective permission is unknownExpose the unresolved state and retain manual documentation
Doctor in Clinic A chooses manual documentation for the encounterAvailability does not require useThe visit can continue without starting the AI task

Specify what turning a feature off actually does

Disabling a feature can mean several things: preventing new requests, stopping current capture, cancelling queued work or hiding the interface. These are different behaviors. Ask the vendor to explain the supported result, then test it with a fictional task already in progress.

For example, start a fictional summary request, disable the task and observe what happens when the result returns. Does it appear as a pending draft, get discarded or write into the record? The clinic should not infer cancellation from a disabled button. Equally, an already accepted note should not vanish merely because future AI use is turned off.

For audio, inspect capture and upload separately where the product permits. A setting saved in another browser tab may not immediately change an existing recording session. Establish the actual propagation behavior and a clear stop procedure. Do not promise instant revocation across devices until that behavior has been demonstrated.

Track a late result against the policy and record it started with

K-08 also tests a history-summary task in Clinic A. At 15:20:00, a fictional job starts under policy version 6. At 15:20:05, an authorized owner saves version 7, disabling new summary requests. The doctor completes the encounter note manually at 15:21:00. The earlier job returns at 15:21:04. These are illustrative times, not a measured service latency or a cancellation guarantee.

There are two questions: whether new work is blocked under version 7, and how work already started under version 6 is handled. A saved setting answers neither by itself. The test should attempt a new request through the supported workflow and inspect the late result’s disposition. If cancellation is requested, record whether the product confirms it or merely acknowledges the request.

Under this proposed policy, the late result must not replace the manually accepted note or become approved clinical content automatically. It should remain clearly identified for the defined review/disposal process, or be discarded according to the actual supported policy. The clinic needs to know which behavior applies. A spinning request disappearing from view is not evidence that processing stopped or a returned draft was handled.

Keep the result attached to its originating clinic, patient, encounter and source version. Switching to Clinic B or opening another chart while the task runs must not make the currently visible chart the destination. The display context when a result arrives is different from the request context that should identify its record. Demonstrate that boundary separately from the basic on/off switch.

NIST’s voluntary generative-AI profile includes deactivation protocols and attention to retained data and dependencies when phasing out a system. That supports asking about more than a visible toggle. The particular K-08 rules and timings are original operational examples, not prescribed regulatory settings or evidence of a product implementation.

On a narrow screen, scroll the table horizontally to read all columns.

K-08: fictional policy change, manual completion and late summary
TimeObserved test eventEvidence the owner needs
15:20:00Summary job starts under policy version 6Originating clinic, encounter, source scope and job reference
15:20:05Version 7 disables new summary requestsCurrent effective setting and actual rejection of a new request
15:21:00Doctor completes the encounter note manuallyAccepted record version that must remain protected
15:21:04Earlier job returnsDisposition of the late draft; no unintended overwrite or approval

Source: NIST AI 600-1: Generative AI risk management profile

Re-enable a task through a bounded decision and a known draft backlog

Turning the switch back on should not automatically release every old draft. In a fictional extension of K-08, four unaccepted results remain: two from the approved test scope, one based on an obsolete source and one with an unresolved request-context reference. Record the groups separately. Their count describes outstanding work, not four notes ready for approval.

The responsible clinician can review the two current-scope drafts under the applicable policy. The obsolete-source item needs a deliberate decision about regeneration or disposal, and the unresolved-context item needs investigation before clinical use. Re-enabling future generation does not supply missing source or identity evidence for any of them. Preserve the appropriate record and administrative history for the chosen outcome.

Retest the boundary that changed before expanding use. For a permission correction, test the affected role in both an existing session and a fresh one. For a clinic-switch defect, start a fictional task in A, switch to B and verify its destination. For a late-result defect, finish manually before the response arrives. Keep the controls and fallback usable if a retest fails.

Write the handover for the next clinician and administrator. It should state what is available now, which old results remain unresolved and who owns them. A general message saying AI is working again leaves staff to infer whether historical drafts are current and whether accepted records changed. A specific scope decision is easier to operate during a busy multi-doctor OPD.

Review defaults when adding a doctor, clinic or device

New entities often inherit settings. Test a newly added clinic, a new doctor and a returning user on a replacement computer. Record which settings are copied, which require an explicit choice and whether the default is understandable before the first task begins.

A narrow initial configuration is easier to evaluate: a named clinician, a defined documentation task and a known source scope. Expansion can follow evidence and an owner’s decision. This is a practical rollout approach, not a universal legal requirement to use a particular default.

Include language and input-source defaults in the review. A previous doctor’s output language or an unexpectedly broad history source can affect the next draft even if the basic AI permission is correct. A shared workstation therefore needs clear user and encounter context as well as an organization setting.

Keep configuration changes traceable and retestable

Record who changed a setting, the affected scope, the old and new value, the reason and when it took effect. Verify what evidence the product provides rather than assuming every configuration action appears in the clinical audit log. Clinical edits and administrative changes may have different log coverage.

A change should have an acceptance check tied to its effect. Enabling a new language calls for the relevant language cases; changing the provider calls for data-path and terms review; changing a role requires permission checks. Repeating every unrelated test is less useful than preserving a small, meaningful set for each boundary.

Plan a rollback that describes both future use and existing work. Reverting a setting may stop new generation without removing drafts already created. Identify who reviews those drafts and whether any issued output requires follow-through. Configuration rollback is not a substitute for correcting a clinical record.

Make usage limits and service failures understandable

If the product has usage limits or paid add-ons, establish what happens when they are reached. The doctor should know whether a draft was never requested, is still processing or failed. An unexplained spinner can lead to repeated requests and duplicate content. Verify available controls and pricing terms rather than assuming a particular limit or charging model.

Keep the established documentation route usable when AI is unavailable or disabled. Test completing a fictional encounter manually, then receiving a late AI result. The late result should be handled deliberately rather than overwrite the clinician’s completed work. The product’s exact behavior must be observed and recorded.

The goal is a configuration the clinic can explain, operate and review. It is not the largest collection of enabled features. A specific unresolved capability, such as uncertain cancellation of in-flight audio, should remain visible in the purchase or rollout decision.

One visiting doctor, two different clinic policies

A fictional visiting physician uses approved dictation at Clinic A, then opens an encounter at Clinic B, where that task remains disabled pending review. The test checks that the effective setting is clear and that the doctor’s preference cannot silently carry the task across. A second test disables a running request and records how its late result is handled without replacing a manually completed note.

Clinic AI Configuration Review Worksheet

Use these prompts in your clinic’s own notes. Record the result, responsible person and next action for each item.

Task map
List each AI task, its data scope and the people allowed to configure and use it.
Effective permissions
Test precedence and defaults across clinics, roles, devices and encounters.
Changing state
Observe disable, pending jobs, late results and already accepted records separately.
Governance
Record changes, task-specific retests, fallback and responsibility for existing drafts.

Questions doctors ask

What should happen when a clinic’s current AI permission cannot be determined?

The product should expose the unresolved state and support the established manual route rather than infer permission from a remembered preference. Verify the clinic’s intended rule and the product’s actual enforcement before relying on it.

Does re-enabling AI make earlier drafts ready to accept?

No. Review their source version, originating context and outstanding issues separately. A restart permits future work within its scope; it does not resolve obsolete sources, unknown destinations or clinical approval of existing content.

Is one AI on/off switch enough for a clinic?

It may not express different tasks, roles and encounter choices. Establish what it controls and whether the product supports the boundaries the clinic needs.

Should a doctor’s AI preferences follow them between clinics?

Only within the permitted policy of the active clinic. Verify precedence and the visible effective setting instead of assuming preferences transfer safely.

Does disabling AI cancel a request already running?

Not necessarily. Test prevention of new requests, active capture, queued processing and returned results separately.

Does an administrator enabling AI approve the generated notes?

No. Feature configuration and acceptance of clinical content are separate decisions with different responsibilities.

What should happen if AI returns after the doctor finishes manually?

Handle the late result deliberately and protect the completed record from unintended replacement. Demonstrate the product’s actual behavior and define responsibility for remaining drafts.

Sources and scope

Updated on . Controls described here are evaluation requirements to discuss and test, not assertions that every Bractis configuration already supports them. No independent clinical or legal review is claimed.

  • NIST AI 600-1: Generative AI risk management profile

    July 2024; relevant governance/deactivation passages checked 28 October 2026, including GV-1.7-001 and GV-1.7-002. Voluntary cross-sector context; the K-08 policy, timing and controls are original examples.

  • Bractis product overview

    First-party context. Confirm supported permissions, defaults, cancellation and version behavior in the current product.

Try the workflow
with Bractis.

Switch a fictional doctor between two clinic policies, disable an in-flight task and inspect the late result, approval boundary and manual fallback.

Request a clinic demo