An updated SOP in healthcare or pharma: reviewing learning materials without losing control

Source → Verification → Decision → Evidence

A standard operating procedure changes. The document owner approves it, and the team must understand what the new version means for their work. Preparing learning material takes time: someone reads the revision, finds the affected roles, writes questions and checks the answers. An assistant can shorten parts of that preparation, but it can also turn an incomplete comparison into a confident instruction. In healthcare and pharmaceutical settings, the consequences of that mistake depend heavily on the procedure's purpose.

Begin with a nonclinical or clearly bounded administrative procedure and a draft-only workflow. The output is reviewed learning material connected to an approved source version. The organisation keeps responsibility for applicability, practical assessment and work authorisation. This guide describes Brillnet's recommended review process. It does not authorise clinical use, batch release or a change to validated systems.

Define the intended task before selecting documents

Write down what the assistant will produce and who will use it. “Prepare five draft questions about the approved records-handling procedure for review by its owner” is concrete. “Train the hospital” leaves the material, people and consequences undefined. A narrow purpose makes access decisions and evaluation possible.

Select a procedure whose source owner can participate in the pilot. Confirm the approved revision and the previous revision used for comparison. Include attachments only if the owner confirms that they belong to the controlled procedure. A missing flowchart or exception table can change the meaning of a question even when the main document appears complete.

Keep patient records and identifiable case histories outside the initial source set whenever they are unnecessary. An administrative training draft often needs process instructions rather than personal health information. If the intended task later changes, reassess the data flow and system impact before expanding access.

Establish the applicable legal and quality context

The GDPR governs personal data processing, with additional conditions for health data under Article 9. An AI feature does not provide a legal basis for processing. Consider purpose, legal basis, access, processor arrangements, security and any required impact assessment for the actual workflow. Articles 28, 32 and 35 may be relevant depending on the circumstances. [1]

EU GMP guidance applies to pharmaceutical activities within its relevant scope. Annex 11 concerns computerised systems and includes a lifecycle and risk-based context. A hospital's administrative procedure is not automatically a GMP process, while an integration that affects GMP work requires assessment of that impact. The small size of a pilot does not remove an applicable obligation. Use the current official guidance and your organisation's qualified interpretation. [2]

Under the AI Act, purpose and use determine classification and the duties associated with the relevant role. Some health-related uses and products can enter high-risk areas. An internal learning draft does not receive its classification solely because a hospital uses it. Do not quietly extend this project into diagnosis, treatment selection, triage or a medical-device function. Such a change needs a separate assessment. [3]

Produce a change summary with visible evidence

Ask for a summary organised by affected action, source location and unresolved question. The reviewer should be able to open the old and new passages behind each claimed change. A general statement such as “the procedure is stricter” does not tell staff what they must do differently.

Separate a document observation from a proposed learning response. The changed source may introduce an additional verification step. The assistant can propose a question about that step. The owner decides whether the question accurately represents the work and whether a practical demonstration is also needed. Preserve that distinction in the packet's statuses.

Do not let the draft resolve conflicts between an approved procedure and an older local instruction. Show the conflict with both sources and ask the responsible owner to resolve it. Otherwise, a polished training page can conceal a problem in the underlying document system.

Build questions that check the changed action

Each question should have a purpose. If a changed procedure requires staff to verify a recipient before sending a document, ask about the verification decision. A question about the procedure's publication date checks memory of a date, not the changed action. Use scenarios that the role actually encounters and avoid irrelevant clinical detail.

The answer record needs the accepted answer, the relevant source location and an explanation of why the answer follows from that procedure. Where several responses are acceptable, record that fact. A single-choice format can create a false rule when the procedure permits alternatives under defined conditions.

For a fictional records-handling procedure, a useful scenario is a request that lacks a confirmed recipient identity. The approved response might require referral to a designated person before release. The exact answer must come from the organisation's approved procedure. This example illustrates question design and does not establish a universal healthcare disclosure rule.

Review the draft with the source owner

Use a review sheet that makes the decision explicit. Record whether the source is current, the role is affected, the wording is understandable and the answer is supported. A reviewer should also assess whether the draft contains an instruction absent from the source. The model can produce a plausible answer that the organisation never approved.

This sheet is Brillnet's recommended design. It can be incorporated into an existing document-control or learning process. It should not become a second approval system with different owners and contradictory release states.

Keep learning completion within its meaning

Acknowledgement shows receipt of a particular message or document. A knowledge test samples selected understanding. Observation can assess practical performance under specified conditions. None of these records should automatically claim more than the organisation actually assessed. Define the meaning of each completion state before implementing dashboards.

CrewShift, offered in Polish as Zmianowo, is being developed in early access for training and tests based on customer documents with human approval. Confirm the available workflow and functions before adoption. A course completion or test score does not by itself establish official qualification, practical competence or permission to work independently.

When a procedure changes again, preserve which learning version each person received. Updating the source link on an old completion record can create a misleading history. Connect the new revision to a new decision about affected staff and any required follow-up instead of assuming an earlier completion covers the change.

Enforce the boundary in the tool layer

The assistant needs access to the selected approved sources and a draft location. Publication, changes to controlled originals and work-authorisation updates should require the appropriate person's action. Enforce those rights in the integration rather than relying solely on a prompt that asks the model to be careful.

Test a source attachment containing an instruction to disclose an unrelated document. The attachment should remain untrusted source material. OWASP's prompt-injection guidance describes this class of risk. Brillnet recommends a negative test that verifies the tool cannot widen access or choose an unapproved destination based on document text. [4]

Use the minimum useful operational record: operation identifier, source versions, draft version, review decision and outcome. Copying complete patient information into general debugging logs adds exposure without necessarily improving the review. Decide access and retention for the actual evidence needed to reconstruct an event.

Assess system impact before connecting existing records

An integration may seem limited because its visible output is a training page, yet it can reach a controlled document repository, staff database or validation record. Identify those connections and their permissions. Ask what could change if a credential is misused or the wrong document is selected.

For work within GMP scope, assess and validate the relevant system impact according to the applicable framework and the organisation's process. Treat changes in configuration, model, prompt, retrieval method or source set as changes that may need review. Do not assume that a supplier's generic test report covers your intended use. [2]

Pulsar GRC is being developed in early access to organise requirements, controls, risks, actions and evidence from customer documents. A proposed use is linking the change packet to the organisation's own control and follow-up records after scope confirmation. It does not decide compliance or guarantee an audit outcome.

Evaluate the pilot against difficult examples

Prepare reference cases before generating drafts. Include an ordinary revision, an unchanged procedure, an obsolete attachment, a missing answer and a conflict between two sources. The source owner should know the expected response for each case. A draft that says “information missing” can be the correct outcome.

Count unsupported answers, wrong revisions, omitted changed actions and unnecessary personal data. Measure reviewer time and the number of questions that need substantial rewriting. If staff cannot understand the reviewed question, record that issue too. Source accuracy and usability both matter to the learning purpose.

Try an unavailable repository and a reviewer without access to the source. The process should preserve the draft status and show a route to the authorised owner. It should not publish a question because a system failure prevented checking it. Define a manual preparation route so essential work can continue without the integration.

Retain a reconstructable release record

At release, retain the approved source identifiers, material version, reviewer decision and intended audience. Later, someone should be able to establish what was approved and why without asking the model to recreate an explanation from memory. Store corrections as dated additions with reasons rather than replacing the historical decision.

A fictional pilot could compare preparation time for ten questions under the existing method and the reviewed-draft method. Ten is an illustrative sample size, not a validated threshold. Report the actual cases and limitations. A reduction in drafting time is useful only if source checking, rewriting and maintenance do not consume the gain.

Bring an anonymised procedure, its previous revision and the existing learning-completion definitions to the first implementation discussion. The practical deliverable is a reviewed learning packet with clear ownership. Expanding toward clinical decisions, product quality decisions or formal qualifications requires its own scope and assessment.

Test a question whose answer depends on an exception

A fictional administrative SOP can specify a standard route and an exception that requires the procedure owner's decision. An assistant drafts a single-choice question from the standard paragraph and ignores the exception in an attachment. The accepted answer looks clear, yet it teaches staff an unconditional rule the organisation did not approve. This is a useful review case because the error comes from incomplete interpretation rather than an obviously invented sentence.

Give the reviewer the draft question, the current procedure and its approved attachment. Ask that person to identify the source supporting the accepted answer and any condition that changes it. The review record should show the actual decision: correct the scenario, use a different question format or remove the question until the source is clear. A general approval tick does not explain which problem was resolved.

If the attachment was missing from retrieval, correct the source set before generating another draft. Rewriting the question alone can hide the retrieval problem. Test the corrected workflow with another question that also depends on the attachment. The aim is to establish whether the integration reaches the approved material needed for the task, not whether the model can produce one acceptable answer after detailed coaching.

When the reviewer accepts a corrected question, retain the learning version and the source versions behind it. If staff have already received an earlier approved version, the organisation decides what follow-up is appropriate under its process. Do not edit a past completion record so it appears that the person saw the corrected question earlier. The change and its actual timing must remain distinguishable.

Measure source checking, rewriting and follow-up alongside drafting time. The owner should be able to explain which errors originated in missing material, conflicting sources or unsuitable question design. Those causes need different corrections. Combining them into one accuracy percentage can hide the action required to improve the workflow. Record the actual sample and the limits of the test.

This exercise stays within the defined administrative learning task. It does not authorise a clinical decision, grant a qualification or remove any applicable assessment and validation requirement. Brillnet recommends it as a practical test of source-grounded preparation and human review. The responsible organisation determines the procedure's scope, the appropriate learning method and any practical assessment needed before independent work.

Sources

The sources support the identified legal facts, guidance and public product scope. Workflows, fictional examples and evaluation measures are Brillnet's operating recommendations. Source review: 2026-10-04.

1. European Parliament and Council — GDPR: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02016R0679-20160504
2. European Commission — EudraLex Volume 4, GMP guidance and Annex 11: https://health.ec.europa.eu/medicinal-products/eudralex/eudralex-volume-4_en
3. European Commission — AI Act questions and answers: https://digital-strategy.ec.europa.eu/en/faqs/ai-act-questions-answers
4. OWASP — prompt-injection prevention guidance: https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html
5. Pulsar GRC — public product scope and early-access status: https://pulsar-grc.com/en/
6. CrewShift / Zmianowo — training, tests and the meaning of a completion record: https://crewshift-app.com/en