Accessible AI-assisted healthcare communications

Learn in your browser · Chapter 01 of 06

Begin with what the reader needs to do and the source facts they must retain. AI may draft clearer administrative wording, but it must not create or alter clinical advice, recipient authority or an appointment.

This workbook concerns adult administrative communications at fictional Elmbridge Clinic. Every person, appointment, preference, address and event is invented. The practice record EC-14 describes a review appointment on 18 November 2026 at 14:20, with arrival at 14:05, at North Building reception. The step-free entrance is on Orchard Walk. Bring the appointment letter. If unable to attend, use the existing appointments team route. Clinical preparation instructions, if any, are supplied separately by the care team. The exercise grants no authority to change that wording into treatment advice, to decide whether attendance is necessary or to send a message to anyone.

A communication brief separates the source of truth from the writing task. The source contains the approved facts; the brief states audience, purpose, channel, permitted changes and reviewer. Do not start with make this friendly, because friendliness does not identify what must survive. A rewrite can be shorter yet worse if it removes the early arrival time, the alternative entrance or the route for someone who cannot attend. In this course, material means a fact, condition, instruction or support route whose alteration could change understanding or action. A full clinical content review is outside the supplied exercise.

Ask what the person has said they need rather than guessing from age, name, language or a diagnosis. A person may use a screen reader, require large print, prefer a particular language or share a phone. These facts lead to different design questions. They do not justify inferring capacity, literacy or consent. The fictional personas are prompts for inspection, not real participants. Observing a synthetic persona cannot establish that a design works for people with that lived experience. Real communication planning requires the relevant staff and affected users.

1. Copy the approved source identifier and version into a communication brief.
2. State the reader task: recognise the sender, find the appointment, know the arrival requirement and obtain help.
3. List every material fact and condition before drafting.
4. Record the communication need and allowed recipient using only the necessary supplied information.
5. Specify allowed transformations such as sentence order and plain-language expansion, and exclusions such as clinical advice or changed dates.
6. Name source, language, accessibility and release reviewers; leave unsupplied authority unresolved.

Scroll sideways to see every column.

Worked example: a weak brief says make this suitable for older patients. It gives the model an age stereotype instead of a requirement. The corrected brief says retain every EC-14 v2 material fact, use a clear heading and short sentences, preserve the separately supplied clinical-instruction boundary, and prepare a large-print layout for a fictional reader who requested it. The output is a draft with a source-fidelity ledger and an unapproved status. Verification checks the eight material facts, not whether the model calls the result accessible.

• Review 1: Why write a brief before asking for a friendly rewrite?
• Review 2: What makes a fact material?
• Review 3: Can a person’s name identify their preferred language?
• Review 4: Does a synthetic persona prove accessibility?
• Review 5: What stays outside this drafting task?
• Review 6: Why retain a source version?

Scroll sideways to see every column.

Source context: NHS-CONTENT recommends a documented approval and review process with user needs. The EC-14 brief and personas are original fictional exercises.

Learn in your browser · Chapter 02 of 06

A successful rewrite is easier to use while preserving every material fact, condition and distinction. Count fidelity before celebrating fewer words. A readability score cannot determine whether an instruction is correct.

The EC-14 appointment time is 14:20 and arrival time is 14:05. Those are different facts. A model may collapse them into arrive at 14:20, especially if asked to shorten the letter. It may change if unable to attend, contact the appointments team into cancel online, although no online cancellation route was supplied. It may interpret separate preparation instructions as an invitation to invent generic advice. These errors can be fluent and grammatically correct. Compare against a ledger of source meaning rather than trusting a second model to agree with the first.

A fidelity ledger has one row per material fact, its source locator, the draft wording and a disposition: retained, changed, omitted, ambiguous or unsupported addition. Numbers, dates, negation, exceptions and actor names deserve deliberate attention. Retain the difference between an appointment time and a requested arrival time; between a proposed action and a completed action; and between contact the team and your appointment has been cancelled. Use unambiguous written dates in these fictional letters. Do not convert a source date whose meaning is uncertain until the owner resolves it.

Plain language can change grammar while preserving meaning. Please present yourself for registration may become Please go to reception. A sentence may be split into two. An unfamiliar administrative term may be explained if the explanation is approved and does not add a new condition. Avoid using a fixed reading age as the only target. It is possible to achieve a low score by dropping necessary qualifications. The NHS messaging guidance encourages simple wording, important information early and channel-specific design [NHS-MESSAGES]. It does not make AI output self-approving.

1. Number the material source facts and record their exact values.
2. Draft an answer-first heading followed by date, appointment time, arrival time and place.
3. Use short sentences and familiar administrative wording, preserving conditions.
4. Compare each source fact with the draft and classify any difference.
5. Search separately for new claims that were not in the source ledger.
6. Hold unresolved facts and send questions to the source owner before any approval.

Scroll sideways to see every column.

Worked corrected draft: Your Elmbridge Clinic appointment. Your appointment is on 18 November 2026 at 14:20. Please arrive at 14:05. Go to North Building reception. The step-free entrance is on Orchard Walk. Bring this appointment letter. If you cannot attend, use the existing appointments team route. Your care team supplies any clinical preparation instructions separately. This wording remains a fictional draft: a real letter would need the verified operational contact route and all relevant qualified approvals. Verification maps each sentence to EC-14 v2 and confirms that no preparation instruction was invented.

• Review 1: Why are arrival and appointment time separate ledger rows?
• Review 2: Can a lower readability score prove correctness?
• Review 3: What should happen to an ambiguous numerical date?
• Review 4: What is an unsupported addition?
• Review 5: Why inspect omissions separately?
• Review 6: Can the draft create a cancellation rule?

Scroll sideways to see every column.

Sources: NHS-MESSAGES and NHS-CONTENT, checked 10 October 2026. All letter facts, defects and marking rules are fictional teaching material.

Learn in your browser · Chapter 03 of 06

Treat translation as a separate transformation requiring competent language and subject review. A fluent translation, machine back-translation or user language preference does not establish clinical accuracy or consent.

The exercise uses an English gloss of a simulated translation so learners need not claim competence in another language. Source clause: If you cannot attend, contact the appointments team. Simulated target-language review gloss: Contact the appointments team only if you can attend. The words are individually plausible, but the condition has reversed. A bilingual reviewer comparing the actual source and target would need to identify and correct that reversal. This English gloss demonstrates the review problem; it is not a real translation and cannot support any language-quality claim.

A translation brief supplies source version, target language and locale, audience needs, material facts, permitted terminology and reviewer roles. The source owner answers administrative and subject questions; a competent translator or interpreter checks meaning in the target language; the communications owner confirms the approved delivery process. A staff member who speaks some of a language may help identify a concern without being the qualified reviewer for a clinical document. Local service policies determine the required competence and assurance. AI assistance must not be used to disguise the absence of that review.

Back-translation can expose a discrepancy, but it is not an independent proof of fidelity. The second model or translator may reproduce the same assumption, smooth over a mistake or restore a familiar phrase from context. Keep the original target wording and compare it directly with the source through the appropriate human review. Ask reviewers to check negation, time, sequence, conditional wording, action owner, support route and tone. A draft that says help has been arranged when the source says request help changes a future request into a completed commitment.

1. Confirm the language or communication support need through the authorised process; do not infer it from identity.
2. Freeze the approved source and its material-fact ledger.
3. Provide a translation brief with locale, purpose, terminology and excluded clinical rewriting.
4. Have the competent reviewer compare actual source and target text, recording material changes.
5. Resolve queries with the source owner and repeat review on the corrected target version.
6. Bind approval to the exact translated version and channel; keep accessible human language-support routes available.

Scroll sideways to see every column.

Worked review record TR-7: source EC-14 v2; target version T1; target locale recorded in the hypothetical brief; reviewer status pending. The reviewer gloss flags the reversed attendance condition and the claim that support is arranged. The source owner confirms the original conditional and request status. T2 corrects both. Release remains held until the competent reviewer accepts the actual T2 wording and the communications owner checks the intended channel. Verification records which version was checked, the two defects resolved and any remaining question. A model back-translation score is not substituted for the missing reviewer.

• Review 1: Does the English gloss establish language quality?
• Review 2: Who resolves an unclear source fact?
• Review 3: Why is back-translation insufficient alone?
• Review 4: What is wrong with request support becoming support arranged?
• Review 5: Can language preference be inferred from a name?
• Review 6: What binds a translation approval?

Scroll sideways to see every column.

Source context: NHS-CONTENT requires translated clinical content to remain accurate and to receive relevant review. The gloss exercise and review sheet are original educational material; no professional language validation is claimed.

Learn in your browser · Chapter 04 of 06

Accessibility belongs to the whole communication journey, from receiving the message to finding help. A technically structured page does not resolve an unreadable letter, an unsafe shared phone or an unavailable human alternative.

Consider the same message in paper, HTML, text-message and audio forms. Paper can work without an account but may need large print or another requested format. HTML can support headings, zoom and assistive technology when implemented correctly, but the route to it may require inaccessible authentication. A short text message may expose appointment information on a shared phone and cannot carry every detail. Audio can be useful for some people but is hard to scan for a time unless controls and a text alternative are available. Select channels through the approved service process with the person’s needs; do not declare one universal solution.

For the fictional HTML outline, use a meaningful page title, a main heading, separate headings for appointment details and help, and labelled links whose purpose is clear. Keep content in a logical reading order. Keyboard focus should be visible, controls operable without a pointer and text usable when enlarged. W3C WCAG 2.2 provides web accessibility success criteria [WCAG22]. These selected checks are not a full conformance assessment and do not cover every healthcare communication need. Actual testing must include the delivery and support journey, not just an automated page scan.

The NHS Accessible Information Standard describes communicating in ways that covered service users can access and understand [NHS-AIS]. Its scope is not identical to every language-translation need. A language preference alone should not be presented as proof of a disability or impairment. In this exercise, P1 asks for large print, P2 uses a screen reader and a private digital route, P3 requests a language reviewer, and P4 has no internet access. Their profiles are invented inspection prompts. Results are design observations, not claims about real patient satisfaction, comprehension or compliance.

1. Map receipt, recognition, reading, action and help for each supplied persona.
2. Choose an approved primary channel and a usable fallback, preserving the same material facts.
3. Inspect headings, reading order, link purpose, keyboard operation and enlargement for digital formats.
4. Check print size, spacing and whether critical content is clipped for a paper layout.
5. Check privacy at the channel boundary, including shared devices and recipient authority.
6. Record actual observed tasks separately from assumed persona responses; arrange real user testing before claiming usability.

Scroll sideways to see every column.

Worked journey: D2 is an HTML letter whose only help link says click here and opens a portal that requires a phone the persona does not have. The body passes a simple heading scan. The journey still fails. Repair the label to identify the appointments support route, confirm the actual approved destination and provide the service’s usable non-portal alternative. In this private exercise the actual route is not supplied, so the delivery field remains a release blocker. Verification records that the draft design is improved while real channel testing and contact validation remain pending.

• Review 1: Why can a page pass automated checks yet fail the journey?
• Review 2: Does WCAG alone cover every communication need?
• Review 3: What is wrong with click here as the only help label?
• Review 4: Why inspect shared-device use?
• Review 5: Can P1 to P4 establish patient satisfaction?
• Review 6: What must a non-digital alternative contain?

Scroll sideways to see every column.

Sources: WCAG22 and NHS-AIS, checked 10 October 2026. The four-persona journey matrix is an original inspection exercise, not a conformance or user-testing report.

Learn in your browser · Chapter 05 of 06

An approved message is a particular version for a particular purpose, recipient and channel. A model-generated draft, staff preference or delivery receipt cannot substitute for the required authority and review.

The practice approval record uses separate fields for source verification, clinical boundary, language review when relevant, accessibility review, recipient/channel authority and release decision. These fields are fictional workflow design, not a universal consent form. In a real service, qualified governance staff determine the lawful basis and applicable confidentiality, consent and proxy arrangements. Do not assume that consent is always the data-protection basis, or that choosing a communication format authorises every disclosure. This course supplies no real patient consent and no permission to contact anyone.

Minimise the draft-review record without destroying necessary traceability. A policy-correction ticket may need the source version, draft version, changed field, responsible role and disposition. It usually does not need a copied clinical history merely to explain that an appointment time was wrong. ICO guidance describes the data-minimisation principle in terms of adequacy, relevance and necessity for the purpose [ICO-MIN]. Real retention periods, access rules and lawful processing require local determination. The workbook avoids inventing a universal number of days or a claim that local hosting alone makes processing acceptable.

A review can expire through change even without a calendar expiry date. If EC-14 v2 becomes v3 with a different arrival time, approved draft D3 must be checked again. The same applies if a translated target version, support route, recipient or channel changes. Keep earlier records distinguishable from the corrected version. A checksum can help compare bytes but does not authenticate the approver or prove the content is true. A checkbox labelled clinician approved is only a statement unless the service has an appropriate identity, authority and evidence process behind it.

1. Confirm the approved source version and identify material fields for verification.
2. Assign the relevant subject, language, accessibility and communications roles.
3. Record each reviewer’s scope and unresolved findings; do not combine incomplete checks into a green status.
4. Confirm the intended recipient and channel through the authorised service process.
5. Keep only necessary review evidence with appropriate access and retention arrangements.
6. On any material change, create a new version and reopen affected reviews before release.

Worked approval: D3 uses EC-14 v2 and passes a staff fidelity check. Its translated version T2 still awaits competent language review, and the private-digital recipient route has not been confirmed. The correct state is HOLD. The team may retain the source/fidelity work while the other checks continue. It cannot send T2 and seek approval later. When the route is confirmed and T2 is reviewed, a designated release owner records the exact versions and scope. Verification checks that missing approvals remain visible and that release is not inferred from the number of completed checks.

• Review 1: Does a preferred format authorise all disclosure?
• Review 2: Why is a universal retention period inappropriate here?
• Review 3: What does a spelling check establish?
• Review 4: Can one missing mandatory review be averaged away?
• Review 5: What does a checksum fail to establish?
• Review 6: What follows a changed appointment time?

Scroll sideways to see every column.

Sources: ICO-MIN supplies data-minimisation context; NHS-CONTENT supplies review-process context. Approval fields, role names and hold rules are original fictional design.

Learn in your browser · Chapter 06 of 06

Testing should check whether the intended reader can find and act on the right information, while correction must reach affected recipients through an authorised route. Sent, delivered, opened and understood are different claims.

The synthetic test pack includes four inspection cards, eight material source facts and three drafts. D1 contains changed facts and invented instructions. D2 preserves the main facts but hides them in an image-only PDF behind a portal. D3 is the corrected plain-language text, still awaiting real layout, channel and user checks. None was sent. Use task questions such as What time is the appointment, when should you arrive, where do you go and how can you obtain help? A yes/no question asking whether the letter is clear can miss a confidently misunderstood time.

A real usability study needs an appropriate protocol, accessible participation, relevant users and approved data handling. In this workbook learners inspect fictional personas and record anticipated barriers. Label those as inspection findings. Do not fabricate interviews, participant quotations, test timings or success percentages. An AI simulation of a screen-reader user is not an assistive-technology test. A tool can help prepare tasks and organise observations, but the observations must describe what was actually done. A technical test also needs its device, browser, assistive technology and settings recorded.

For a worked numerical audit, suppose a fictional inspection set has twelve delivery cases. Nine have delivery receipts; three have no confirmed receipt. Of the nine delivered cases, six have an observed correct response to the arrival-time task, one has an incorrect response and two were not assessed. Report 9/12 delivery confirmation, or 75 percent; task success among assessed delivered cases is 6/7, about 85.7 percent. Do not call this 6/9 comprehension, because two cases are unassessed, and do not generalise seven assessments to all recipients. These numbers are invented for arithmetic practice and do not describe the supplied personas or any real clinic.

1. Define the source facts and reader tasks before inspecting drafts.
2. Record actual actions, observations and unmeasured parts separately.
3. Compare drafts against the eight-fact ledger and the full channel journey.
4. If an error is found before release, hold the affected version and preserve the correction record.
5. If an error was already sent in a hypothetical incident, identify affected version and recipients through authorised records, stop further use, and assign qualified correction ownership.
6. Use the established communication route to issue and verify a correction; record delivery and remaining uncertainty rather than assuming understanding.

Worked incident: a hypothetical sent version D1 used December instead of November. The communications owner stops reuse of D1, preserves its identifier and checks the authorised dispatch record. The source owner confirms the correct date and considers consequences with the relevant service professionals. A corrected message identifies the correction clearly and uses the approved accessible route. A failed delivery remains unresolved and follows the service’s escalation process. Merely replacing the web file does not establish that someone who already received D1 has seen the correction. This is a planning tabletop, not an instruction to send messages from the exercise.

• Review 1: Why is delivered different from understood?
• Review 2: What is 9 confirmed deliveries out of 12?
• Review 3: What is the task-success denominator in the worked audit?
• Review 4: Can synthetic personas supply user quotations?
• Review 5: Why is replacing a web file insufficient after a sent error?
• Review 6: What does a knowledge-exam pass establish?

Scroll sideways to see every column.

Keep this workbook for notes and self-review. The private knowledge bank uses twenty equally weighted questions per sitting and strictly more than 80 percent to pass: sixteen correct fails; seventeen passes. The capstone is ungraded. Check the published course and assessment pages for current exam and certificate availability. Local files do not establish issuance or public availability. An educational knowledge result does not verify identity or professional competence.

Sources: NHS-MESSAGES recommends testing messages with users; the scenarios, numbers and correction tabletop are original synthetic teaching material.