An AI proposal becomes reviewable when you can explain what the complete system does, who relies on it, and which decisions its outputs influence. Begin with the deployed workflow; then assign use categories and regulatory questions.
What you will learn
Distinguish predictive models, generative AI, general-purpose models, and deployed systems: AI use-case definition and regulatory triage, system and workflow boundaries.
Map inputs, outputs, interfaces, suppliers, users, affected people, and downstream decisions: AI use-case definition and regulatory triage, workflow mapping.
Separate medical-product functions, drug-development evidence tools, and internal operational tools while recognizing mixed workflows: AI use-case definition and regulatory triage, initial classification.
Identify behavior-relevant changes and assign an initial review owner: AI use-case definition and regulatory triage, change identification and escalation.
Estimated time
20-25 minutes.
Blueprint mapping
AI use-case definition and regulatory triage → system and workflow boundaries and initial classification questions → domain weight 12%; no separate topic weight supplied. Source anchors: bp.d1.t1 and bp.d1.t3.
Learning frame
Course source snapshot dated November 5, 2026. This lesson teaches foundational scoping and triage, not a definitive legal classification. Specific applicability and commencement questions require the controlled sources addressed in the Regulatory Sources and Applicability Register topic.
Prerequisites
None.
Key outputs
A bounded workflow map, a provisional use-category record, a behavior-relevant change list, and a named initial regulatory review owner.
WHY
Why it matters
The same model can support different decisions in different deployments. Its technical label cannot establish the regulatory status of every use.
A boundary drawn around only the model omits controls that may determine whether its output is appropriate: retrieval sources, interfaces, reviewer actions, and downstream permissions.
An internal tool may influence regulated evidence, manufacturing records, or quality decisions. Internal location alone does not settle applicability.
A system may change behavior without new model weights. Prompts, retrieved content, configuration, and operating procedures belong in the initial map.
Where things stand: The supplied November 5, 2026 blueprint separates drug-development credibility, device quality and change controls, EU AI Act decisions, and GxP computerized-system controls. It identifies February 2, 2026 as the QMSR application date and treats EU GMP Annex 22 as draft readiness material. These distinctions explain why triage must identify the workflow before selecting a framework. This lesson does not independently establish later amendments, provision-specific EU dates, or the legal status of a particular deployment.
KEY TERMS
Key Terms & Must-Know Facts
Table 1. Boundary-mapping reference terms
Term
Working meaning
Question to document
Predictive model
A model used to estimate a class, score, value, or outcome from inputs.
Which estimate influences which decision?
Generative AI
AI used to produce content such as text or other generated outputs.
How is generated content checked before use?
General-purpose model
A model capable of supporting a range of tasks rather than one narrowly bounded deployment.
How is its broad capability constrained in this use?
Deployed system
The operational combination of model, software, data connections, people, and procedures.
What actually reaches a user or another system?
System boundary
The stated scope of components and interfaces included in the assessment.
What is inside, outside, and crossing the boundary?
Workflow boundary
The start, end, handoffs, and decisions of the activity supported by the system.
Where does AI influence begin and end?
Input and output
Information received and results produced at a specified stage.
Which transformations occur between them?
Interface
A point where information or instructions pass between components, people, or services.
Who can read, edit, export, or act?
User and affected person
The operator interacts with the tool; an affected person may experience consequences without using it.
Whose actions and interests must be represented?
Supplier dependency
An externally supplied component or service on which the workflow relies.
Who controls its versions, access, and changes?
Context of use
The specific purpose and circumstances in which an output is used.
What decision, population, setting, and reliance are intended?
Change surface
A component or operating condition whose modification could alter behavior.
What evidence and review does the change require?
Must-know facts:
Predictive and generative describe technical functions; neither is a complete regulatory conclusion.
A general-purpose model and its deployed application occupy different levels of description.
Human review is a workflow feature whose actual tasks and authority must be specified.
One workflow can cross several use categories and require coordinated review.
Record unresolved classification questions explicitly rather than converting uncertainty into an assumption.
A provisional review owner coordinates the next assessment; that assignment does not itself authorize deployment.
CORE
Core content
1.1.1.1 1. Separate the model from the deployed system
A Explanation: choose the operational unit of assessment
WHY THIS IS TESTED: A technically accurate model description can still leave the consequential workflow undefined. The decision conflict is whether to assess a model capability alone or the complete arrangement that turns its output into action. Choose the complete arrangement when defining deployment boundaries.
Three descriptions can all be true: a model has broad capabilities, a particular function generates text, and the deployed system drafts a quality document for review. These are complementary descriptions, not competing use categories. For a predictive function, document what the score means operationally. For a generative function, document what happens to the generated content. Do not assume that every generated narrative is decision support or that every numerical score is a medical product.
Include the model service, application, prompt or task instructions, connected sources, output processing, users, and relevant procedures. A third-party component may sit outside organizational control while remaining a dependency in the assessed workflow. See Fig 1.
Describe model capability
Identify the deployed function
Add software and connected sources
Add people and procedures
State the decision influenced
Fig 1. Capability versus deployed use Exam takeaway: Model capability is only one part of the system that requires assessment.
flowchart TD; A[Model capability] --> B[Configured application]; C[Connected information] --> B; B --> D[Human and procedural controls]; D --> E[Output used in a decision];
A model inside an operational system Exam takeaway: Include the surrounding components that determine how the model is used.
{"prompt":"Square educational illustration on a white background, cool blue and gray with teal accent. Show an AI model icon inside a larger operational boundary containing prompt document, retrieval library, software interface, human reviewer, and output record. Outside the boundary show a supplier connected to the model service and affected people connected to the downstream outcome. Thin crisp outlines, generous margins, no figure numbers, no regulatory badges. Use short labels only.","caption":"A model inside an operational system","size":"1024x1024"}
B Worked example: a drafting assistant
Scenario: A development team evaluates a general-purpose model that drafts summaries for a study involving adults. One source record concerns a 67-year-old participant with diabetes and chronic kidney disease; a mildly elevated glucose result appears in the record. The vendor emphasizes that the model is a general assistant and that a scientist signs every summary.
Identify the task: Define the assessed system, not the participant's treatment. Extract key facts: The application selects source information, generates a draft, and presents it for scientist review. Apply the rule: Include source selection, instructions, model version, draft interface, review procedure, and destination. Eliminate: The general-assistant label does not describe the actual deployment; the glucose result does not determine the governance boundary. Verify: Ask whether the approved summary enters a development evidence package and what the scientist must check. Until that destination is known, the map remains incomplete.
C Common trap: treating review as a boundary shortcut
Wrong answer: Assess only the model because a scientist signs the result. Why tempting: A signature appears to place all responsibility outside the AI system. Single eliminating clue: The application determines which source information the scientist receives; that upstream selection remains part of the workflow.
D Checkpoint: identify the missing scope
A team documents a text-generation model and its accuracy evaluation but omits the source-selection interface. What is the most appropriate next step?
Expand the boundary to include source selection and its connection to the output.
Keep the boundary unchanged and assign the omission to user training.
Replace the model description with a general-purpose-model label.
Limit the assessment to the signed final document.
Answer: 1. Option 1 is right because the interface affects what the system can generate. Option 2 is tempting because training matters, but training does not describe the omitted component. Option 3 is tempting because it clarifies capability, but it does not map the deployment. Option 4 is tempting because the final record matters, but it loses the path by which that record was produced.
If you missed this: Review Separate the model from the deployed system - focus on components that influence the output before approval.
1.1.1.2 2. Map information, people, and decision handoffs
A Explanation: choose consequence tracing over screen tracing
WHY THIS IS TESTED: A screen-level diagram can look complete while omitting the decision that creates risk. The decision conflict is whether to stop at the displayed output or follow its permitted downstream use. Choose a trace that starts with information entering the workflow and ends with the action, record, or decision it supports.
For every handoff, record the source, transformation, recipient, permitted action, and responsible role. Show external interfaces even when their internals are unavailable. Distinguish a user from an affected person: a quality reviewer may operate the application while a patient, participant, or product recipient is affected by its consequences.
Specify what human oversight means. Can the reviewer compare the output with sources, reject it, request clarification, and stop downstream use? Those are questions to resolve, not evidence that oversight is effective merely because a reviewer exists. Show exceptions such as missing input, failed retrieval, or unreviewed export. See Fig 2.
Locate the input source
Trace transformations and interfaces
Name the user and reviewer
Follow the downstream decision
Identify affected people and exceptions
Fig 2. Trace beyond the output screen Exam takeaway: The downstream decision and affected people belong on the workflow map.
flowchart TD; A[Source information] --> B[AI transformation]; S[Supplier service] --> B; B --> C[Review interface]; C --> D[Authorized decision or record]; D --> E[Affected people]; C --> F[Reject or clarify];
Operator and affected person Exam takeaway: The person using an AI tool may differ from the person affected by its output.
{"prompt":"Square professional life-science educational illustration, white background and cool blue gray palette. Left show a quality reviewer at a workstation with a draft output. Center show a controlled handoff to a product decision record. Right show an adult patient icon receiving a product. Separate short labels User and Affected person. Make the human roles visually distinct; no treatment advice, no clinical values, no figure number. Thin outlines and generous margins.","caption":"Operator and affected person","size":"1024x1024"}
A review label versus an actionable review gate Exam takeaway: Document reviewer access and authority rather than relying on the phrase human oversight.
flowchart TD; A[Claimed human review] --> B{Tasks and authority described?}; B -->|Yes| C[Map source checks and disposition]; B -->|No| D[Record unresolved control question]; C --> E[Evaluate adequacy in later assurance review]; D --> E;
B Worked example: the hidden export
Scenario: A hospital research office uses an AI summary tool for records that include a 59-year-old participant with hypertension and asthma. A normal oxygen saturation appears in the record. The screen marks every result as draft, but a spreadsheet export automatically feeds a study eligibility worksheet. The vendor highlights that the tool has no prescribing function.
Identify the task: Find the consequential handoff. Extract key facts: The exported result influences an eligibility workflow despite the draft label. Apply the rule: Extend the map through export, worksheet transformation, investigator review, and the resulting decision. Eliminate: Absence of prescribing does not establish that the workflow is inconsequential; the oxygen saturation does not answer the mapping question. Verify: Confirm whether the worksheet is advisory, who checks source records, and whether an unreviewed value can influence a decision. Escalate the unresolved clinical-investigation protection question to the appropriate owner without deciding eligibility yourself.
C Common trap: assuming a label controls the workflow
Wrong answer: End the boundary at the draft screen. Why tempting: The interface appears to prohibit reliance. Single eliminating clue: An automatic export feeds the eligibility worksheet, creating a downstream path that the label alone does not explain.
D Checkpoint: identify the affected role
A quality specialist operates an AI tool that prioritizes manufacturing deviations for review. Which addition most improves an incomplete map containing only the specialist and the application?
The supplier's full marketing description.
The downstream quality decision and people potentially affected by product consequences.
The model's broad task capabilities without operational handoffs.
The specialist's training record without the output destination.
Answer: 2. Option 1 is tempting because supplier context is relevant, but marketing does not trace consequences. Option 2 is right because it connects use to the affected workflow. Option 3 is tempting because model capability matters, but it leaves the decision unexplained. Option 4 is tempting because competence supports controls, but a training record cannot identify where the output goes.
If you missed this: Review Map information, people, and decision handoffs - focus on the action supported after the output is displayed.
Quick Recall:
What closes a workflow map? The downstream action or decision and relevant affected people.
True or false: A supplier outside organizational control can still be a mapped dependency. True.
Complete the phrase: Human review requires defined tasks and _____ . Authority.
1.1.1.3 3. Assign provisional use categories without overclassifying
A Explanation: choose use-based triage over vendor-based conclusions
WHY THIS IS TESTED: Similar software can create different regulatory questions when its intended function or output destination changes. The decision conflict is whether to accept a supplier's broad category or record the organization's specific use. Choose the specific use and preserve uncertainty for the responsible review.
Table 2. Three use categories for initial triage
Provisional category
Operational discriminator
Initial review question
AI-enabled medical-product function
The proposed function is part of a medical-product use that requires qualification assessment.
What is the intended function, and which product framework may apply?
Drug-development evidence tool
The output is used to generate, analyze, or support evidence in drug development.
What decision relies on the output, and what credibility evidence is needed?
Internal operational tool
The output supports an organizational process such as quality, manufacturing, or administration.
Does this process affect GxP operations, required records, or other regulated activities?
Mixed workflow
Separate functions or output routes span more than one category.
Which function needs which owner, evidence, and interface controls?
These categories organize review; they are not statutory classifications. A medical-product possibility does not establish a device class. A development-evidence use does not determine evidence adequacy. An internal use does not settle electronic-record applicability. Record the jurisdiction and operator-role questions separately. See Fig 3.
State the intended function
Locate each output destination
Assign provisional use categories
Flag mixed or uncertain functions
Route applicability questions
Fig 3. A use-category decision screen Exam takeaway: Screen every relevant use rather than forcing a mixed workflow into one exclusive category.
flowchart TD; A[Intended function and output destination] --> B[Screen medical-product function]; B --> C[Screen drug-development evidence use]; C --> D[Screen internal operational use]; D --> E{More than one relevant route?}; E -->|Yes| F[Document mixed workflow]; E -->|No or uncertain| G[Record provisional category and questions]; F --> H[Coordinate applicability review]; G --> H;
Different uses of a shared model Exam takeaway: A shared model can support functions that require different review pathways.
{"prompt":"Square educational comparison illustration for AI governance in life sciences. Three clearly separated vertical panels show the same small AI model symbol supporting different uses: medical-product interface, drug-development evidence report, and internal quality dashboard. Labels Medical product, Development evidence, Internal operations. Use blue and gray with teal accent, white background, crisp outlines, no regulatory approval seals, no figure numbering. Show separate output destinations, not a claim of exclusive legal categories.","caption":"Different uses of a shared model","size":"1024x1024"}
A mixed workflow needs separate routes Exam takeaway: Shared infrastructure does not eliminate function-specific review questions.
{"prompt":"Square clean educational illustration of one life-science AI service with two controlled output routes. One route leads to a development evidence dossier reviewed by a scientist; the other leads to an internal quality record reviewed by a quality specialist. At the split show a small label Mixed workflow. White background, cool blue gray and teal, simple professional icons, generous margins, no legal classifications or figure numbers.","caption":"A mixed workflow needs separate routes","size":"1024x1024"}
B Worked example: one engine, two uses
Scenario: A sponsor uses a text-analysis service to summarize safety narratives for a development evidence report and to organize internal quality complaints. A narrative describes a 72-year-old participant with heart failure and diabetes; a stable body weight is highlighted. The supplier calls the service an internal productivity platform.
Identify the task: Assign initial use categories and review routes. Extract key facts: One route supports development evidence; another supports a quality process. Apply the rule: Document both routes, their destinations, and the reliance placed on each output. Eliminate: The productivity label cannot collapse the evidence route into general administration, and the participant's stable weight does not determine the use category. Verify: Obtain the scientist's evidence-use description and the quality process owner's account of complaint handling. Assign coordinated review questions rather than concluding that every function has the same status.
C Common trap: using hosting location as a category
Wrong answer: Classify all functions as unregulated internal operations because the system runs on the sponsor's network. Why tempting: Internal hosting feels separate from marketed products. Single eliminating clue: One output supports a development evidence report; its use remains relevant regardless of hosting location.
D Checkpoint: distinguish triage from determination
An internal AI application drafts a required quality record. Which statement is most appropriate at initial triage?
Internal deployment settles the application as outside electronic-record review.
The generative function makes every output a medical-device function.
The operational category should be documented, with applicability questions about the record and process.
Supplier validation establishes that the organization's record controls are adequate.
Answer: 3. Option 1 is tempting because the application is not sold, but record and process applicability remain unresolved. Option 2 is tempting because AI may appear in medical products, but generative capability alone does not determine qualification. Option 3 is right because it separates the use category from the next regulatory assessment. Option 4 is tempting because supplier evidence can help, but it cannot establish adequacy for an unspecified organizational workflow.
If you missed this: Review Assign provisional use categories without overclassifying - focus on the distinction between organizing review and reaching a regulatory conclusion.
1.1.1.4 4. Connect change surfaces to ownership and review
A Explanation: choose behavioral impact over component labels
WHY THIS IS TESTED: Teams may monitor model releases while overlooking changes that alter the information, instructions, or reliance surrounding the model. The decision conflict is whether an unchanged model means an unchanged system. Choose a review based on potential behavioral and workflow impact.
Maintain a baseline description of the model or service, instructions, retrieval sources, configuration, interface, review procedure, and intended output use. For each proposed change, ask what could differ, which decision could be affected, and what prior evidence may no longer apply. An identified change is a reason for impact assessment, not an automatic conclusion that a submission, full revalidation, or a particular legal duty is triggered.
Table 3. Change surfaces and discriminating review questions
Change surface
Potential consequence to examine
Question for the owner
Model or service version
Different generated content, scoring, or task performance.
What changed, and does existing evidence cover the deployed version?
Prompt or task instruction
Different selection, emphasis, exclusions, or response format.
Does the instruction change alter the defined task?
Retrieval source or selection
Different information reaches the model or reviewer.
Are source identity, permissions, relevance, and selection still appropriate?
Configuration or interface
Different output handling, visibility, access, or export behavior.
Can users now act differently or bypass a prior gate?
Workflow or population
Different reliance, setting, affected people, or decision consequence.
Has the context of use expanded beyond the assessed scope?
Assign the initial regulatory review owner according to the consequential use. The product regulatory function may lead a medical-product question; the development regulatory function may coordinate an evidence-use question; quality may lead a GxP process assessment with regulatory support. These are practical coordination examples, not universal organization charts. Name an accountable role, participating functions, unresolved questions, and a next decision point. See Fig 4.
Describe the baseline
Identify the changed surface
Trace the possible decision impact
Assign a coordinating owner
Document review and evidence questions
Fig 4. Review a changed system, not only changed weights Exam takeaway: A stable model version does not establish that the assessed workflow remains unchanged.
flowchart TD; A[Proposed change] --> B{Only model weights considered?}; B -->|Yes| C[Expand impact assessment]; B -->|No| D[Review all changed surfaces]; C --> D; D --> E[Trace output and decision impact]; E --> F[Owner defines evidence and approval questions];
An unchanged model with changed retrieval Exam takeaway: Different source material can alter system behavior without a model update.
{"prompt":"Square side-by-side educational comparison. Left show an AI model connected to a small approved document library and a draft output. Right show the identical model icon connected to an expanded library with new document folders and a different draft output. Labels Same model and Changed retrieval. White background, cool blue gray with teal accent, clear arrows, no numerical performance claims, no figure numbers, generous margins.","caption":"An unchanged model with changed retrieval","size":"1024x1024"}
From proposal to accountable review Exam takeaway: A named coordinating owner turns a classification question into an actionable review task.
{"prompt":"Square operational educational illustration of a life-science governance review desk. Show a workflow map, change log, and question register being handed to a named role card labeled Review owner. Three nearby professional role icons labeled Regulatory, Quality, and Data science collaborate. Use white background, blue gray palette, teal accents, crisp outlines, no approval stamps, no figure numbers, no excessive text.","caption":"From proposal to accountable review","size":"1024x1024"}
B Worked example: a retrieval-only update
Scenario: A quality team uses an assistant to draft investigation summaries. A linked complaint describes a 64-year-old person with chronic kidney disease and hypertension; a normal sodium result is included. The vendor keeps the model version unchanged but replaces the retrieval library and changes ranking configuration. The release note calls the update maintenance only.
Identify the task: Decide whether the proposed update needs impact review. Extract key facts: Both available information and its selection order change. Apply the rule: Assess the retrieval and configuration changes against the baseline task, review procedure, and downstream quality record. Eliminate: A maintenance label does not show that output behavior is unaffected; the sodium value is unrelated to change scope. Verify: Ask quality to coordinate with the system owner and relevant regulatory function on affected evidence, permitted implementation, and unresolved applicability questions. Do not infer a specific submission requirement from this change description alone.
C Common trap: confusing ownership with permission
Wrong answer: Deploy immediately once a review owner is named. Why tempting: Accountability can look like the final governance gate. Single eliminating clue: The evidence-impact and applicability questions remain unresolved; assignment starts the assessment rather than completing it.
D Checkpoint: identify the meaningful change
The model and prompt remain unchanged, but an approved draft can now be exported directly to a quality record without the former review step. What is the most appropriate next action?
Record a workflow change and assess the altered reliance and control gate.
Close the change because model behavior has not been intentionally modified.
Repeat only the supplier's generic benchmark before authorizing the export.
Classify the change as a new medical device before examining the function.
Answer: 1. Option 1 is right because the route to consequential use has changed. Option 2 is tempting because the model is stable, but system behavior includes the review workflow. Option 3 is tempting because testing provides evidence, but a generic benchmark does not address the removed gate. Option 4 is tempting because the change may be consequential, but device qualification requires an intended-function assessment rather than an automatic label.
If you missed this: Review Connect change surfaces to ownership and review - focus on changes to reliance and permissions even when the model is unchanged.
Quick Recall:
Name two non-model change surfaces: prompts and retrieval sources; configuration and workflow also qualify.
Does assigning an owner authorize deployment? No; it establishes responsibility for the next assessment.
What belongs beside a provisional category? The unresolved question and the role responsible for resolving it.
TRAPS
Exam Traps & Differentiators
Common traps & distractors:
Choose the framework from the word AI. Tempting because a technology label is easy to recognize. Eliminating clue: the intended function and decision remain unspecified.
Copy the vendor's classification into the approval record. Tempting because the supplier appears authoritative about its product. Eliminating clue: the organization's actual output destination differs from the vendor's described use.
Treat a numerical output as inherently lower risk than a narrative. Tempting because numbers look objective. Eliminating clue: the score directly determines prioritization or another consequential action.
Use one owner to erase multiple review routes. Tempting because a single intake record simplifies coordination. Eliminating clue: distinct functions create distinct applicability questions even when one person coordinates them.
Table 4. Looks similar but requires a different conclusion
Surface similarity
Discriminating fact
Boundary or triage response
Both outputs are summaries.
One supports a development evidence decision; the other organizes general administrative material.
Record different destinations and reliance; do not infer identical review needs.
Both systems have human review.
One reviewer can inspect sources and stop use; the other's task is undefined.
Document actual review tasks and authority before judging adequacy.
Both use the same hosted model.
The prompts, connected sources, and downstream interfaces differ.
Assess separate deployments despite shared infrastructure.
Both are internal tools.
One influences a required quality record.
Raise the relevant process and record questions rather than stopping at internal status.
Neither changes model weights.
One removes an output approval gate.
Assess changed workflow reliance and permissions.
If the proposal says:
General-purpose model → ask which bounded application is actually being assessed.
For internal use → ask which internal decision, process, and records are affected.
Human in the loop → ask what the human sees, checks, and can stop.
Same model version → inspect prompts, sources, configuration, and workflow.
Already validated by supplier → ask whether the evidence matches this deployment and reliance.
Fig 5. Replace labels with reviewable evidence Exam takeaway: A governance conclusion needs a function, a decision path, and an assigned question owner.
flowchart TD; A[Vendor label] --> B{Actual use described?}; B -->|No| C[Obtain workflow and destination]; B -->|Yes| D[Compare label with intended use]; C --> D; D --> E[Document classification questions]; E --> F[Assign review owner];
See Fig 5 when a proposal offers confident labels but little operational detail.
Review connection: Return to Assign provisional use categories without overclassifying. A category is a routing aid; the applicability register must later capture the authority, role, timing, and supporting source for the determination.
What is coming next: Evolving guidance and draft materials may change readiness priorities. Keep a source-status field that distinguishes current requirements, guidance, and proposals. The supplied blueprint identifies draft EU GMP Annex 22 as a readiness topic; do not present a draft as an effective obligation. This lesson does not supply enough detail to determine provision-specific EU commencement or transition dates. Address those questions in EU Commencement and Transition Calendar using the controlling sources.
ALGO
Algorithm / Approach
Use this five-step intake sequence before selecting detailed assurance activities. See Fig 6.
1. Bound the function: State the task, starting input, ending output, and intended reliance.
2. Trace the workflow: Map components, interfaces, suppliers, users, affected people, and decision handoffs.
4. Capture change surfaces: Record model, prompt, retrieval, configuration, and workflow baselines and proposed changes.
5. Assign the next review: Name the coordinating owner, participating functions, open questions, and required decision before intended use.
Fig 6. Initial review readiness decision Exam takeaway: Missing destination or ownership information should be resolved before treating triage as complete.
flowchart TD; A[Draft intake map] --> B{Decision destination known?}; B -->|No| C[Clarify intended reliance]; B -->|Yes| D{Categories and owner recorded?}; C --> A; D -->|No| E[Record questions and assign owner]; E --> A; D -->|Yes| F[Proceed to context-of-use and applicability review];
Reviewable intake record: Include a use-case identifier; intended function; scope exclusions; input sources; output destinations; users and affected people; supplier dependencies; provisional categories; baseline change surfaces; open classification questions; and accountable review role. A scope exclusion should explain a real boundary, not hide a connected decision.
Applied document exercise: Rewrite this proposal: 'Use a validated AI platform to improve study summaries.' A reviewable version identifies which records enter, what the application drafts, which scientist verifies which source elements, where approved summaries go, and who resolves evidence-use applicability. Do not invent those facts. Mark each missing fact as an explicit question for the process owner.
Readiness check: Could another reviewer reconstruct the path from source information to consequential use without asking what the supplier meant by its marketing label? If not, refine the map before moving to a context-of-use statement.
QUIZ
Self-check quiz
Select the most appropriate response in each scenario.
1. A general-purpose model is configured to predict document-routing priority in a quality workflow. Which description best supports initial governance review?
General-purpose model, because capability determines the complete boundary.
Predictive routing application, including inputs, interfaces, reviewer actions, and downstream use.
Internal software, with supplier assurances replacing workflow documentation.
Quality documentation tool, assessed only at the final signed record.
2. A research coordinator uses an AI-generated summary to prepare a worksheet that an investigator reviews before acting. Who should the map distinguish?
Only the coordinator, because that person operates the interface.
Only the investigator, because that person authorizes the decision.
The coordinator, investigator, and people affected by the resulting decision.
The supplier and coordinator, because downstream review is outside AI scope.
3. One service supports both drug-development evidence summaries and internal quality complaint organization. What is the most appropriate initial classification action?
Use the internal-tool label for both because the service is hosted internally.
Use the evidence-tool label for both because it is the more consequential use.
Record a mixed workflow and route function-specific applicability questions.
Determine that both functions are medical products because patient information is processed.
4. A supplier retains the same model but changes document retrieval and output export permissions. What is the best next step?
Assess the retrieval and permission changes against intended use and prior evidence.
Record no relevant change because the model version is stable.
Require a specific regulatory submission before mapping the changed function.
Accept the change after confirming that the supplier calls it maintenance.
5. A workflow map is complete, but an internal quality-record function has unresolved applicability questions. A quality lead agrees to coordinate review. Which statement is best?
The named owner establishes that deployment may begin.
The map establishes that no additional applicability assessment is necessary.
The owner should document the questions, involve relevant functions, and obtain the required decision before intended use.
The supplier's generic classification should settle the outstanding questions.
ANSWERS
Answer key & explanations
1. Answer: 2.
Option 1: Tempting because general-purpose accurately describes broad capability; wrong because capability does not define the configured quality workflow.
Option 2: Correct because it names the deployed task and the surrounding operational components needed to assess its use.
Option 3: Tempting because internal software may use supplier evidence; wrong because neither internal status nor supplier assurances replaces the organization's map.
Option 4: Tempting because signed records are consequential; wrong because assessing only the endpoint omits transformations and controls.
If you missed this: Review Separate the model from the deployed system - focus on the operational unit rather than the model label.
2. Answer: 3.
Option 1: Tempting because the coordinator is the direct user; wrong because the later decision-maker and affected people remain relevant.
Option 2: Tempting because authorization is important; wrong because upstream operator actions influence what reaches the investigator.
Option 3: Correct because it distinguishes the user, decision-maker, and affected people across the complete handoff.
Option 4: Tempting because supplier and operator interfaces are tangible; wrong because the output's downstream reliance is part of the workflow boundary.
If you missed this: Review Map information, people, and decision handoffs - focus on following the output to its consequential use.
3. Answer: 3.
Option 1: Tempting because common hosting suggests a single category; wrong because hosting does not erase development-evidence use.
Option 2: Tempting because one consequential category seems conservative; wrong because it conceals the separate quality-process questions.
Option 3: Correct because it preserves both uses and permits coordinated, function-specific review.
Option 4: Tempting because patient information can appear in medical products; wrong because data content alone does not determine product qualification.
If you missed this: Review Assign provisional use categories without overclassifying - focus on retaining multiple relevant routes.
4. Answer: 1.
Option 1: Correct because retrieval affects available information and permissions affect how outputs reach decisions; both may alter the assessed system.
Option 2: Tempting because version tracking often centers on the model; wrong because other change surfaces remain relevant.
Option 3: Tempting because consequential changes can have regulatory implications; wrong because the required response cannot be determined before examining the function and applicable framework.
Option 4: Tempting because maintenance terminology suggests low impact; wrong because the label does not demonstrate unchanged behavior or reliance.
If you missed this: Review Connect change surfaces to ownership and review - focus on non-model changes that affect information or action.
5. Answer: 3.
Option 1: Tempting because ownership is a governance milestone; wrong because responsibility is not authorization.
Option 2: Tempting because a complete map is useful evidence; wrong because mapping makes applicability review possible rather than resolving every question.
Option 3: Correct because it converts uncertainty into an accountable review with a decision before intended use.
Option 4: Tempting because suppliers know their products; wrong because a generic classification may not address this organization's process, records, roles, or jurisdiction.
If you missed this: Review Connect change surfaces to ownership and review - focus on the distinction between assigning review and granting permission.
RAPID
Rapid review / cheat-sheet
Capability statement → ask which deployed task it enables.
Boundary completeness → test whether another reviewer can reconstruct the decision path.
Human oversight claim → request observable review tasks and stop authority.
Mixed use → coordinate review without collapsing function-specific questions.
Change intake → identify which prior evidence might no longer cover the use.
Supplier assurance → compare its scope with the organization's actual reliance.
Unresolved classification → record the missing fact, accountable role, and next decision.
Key points:
Define the complete deployment before choosing a regulatory route.
Follow information through interfaces to decisions and affected people.
Use provisional categories to organize the next assessment.
Keep change review and ownership connected to the actual workflow.
See also: Context-of-Use Statement Workshop; Risk Classification Across Regulatory Frameworks; Regulatory Sources and Applicability Register. The next topic converts this map into a concise statement of intended use, setting, population, and reliance.
End of your free lesson
Keep going with the full AI Governance in Drug and Device Development Professional Certificate
You just studied the way members do, in the same lesson interface. Full access unlocks every lesson, the question bank and flashcards for a full year.