From task to human-reviewed output

Worked Examples

See the whole workflow: appropriate input, draft output, what AI did well, what it missed and how a person produces the final version.

Worked example 01

Use NotebookLM with drug guidance

Compare what a selected set of public UK guidance documents says about a focused practice question.

The AI input

A defined question and a deliberately selected collection of current, public documents. No client material.

The draft output

A source-linked comparison of recommendations, differences, limitations and unanswered questions.

What AI did well

The tool can keep the discussion bounded by the selected sources and make passages easier to locate.

What AI missed

It cannot identify guidance you failed to include, decide which source governs locally or guarantee that every qualification has survived summarising.

Human checks

Confirm the edition and review date of every source
Open the cited passage
Check tables, footnotes and exceptions
Record sources that were excluded
Have an appropriate practitioner review the interpretation

Final version

A human-reviewed evidence note that links each important statement to the original document and states what the source set cannot answer.

Worked example 02

Prepare a staff briefing

Explain an approved change in one page so staff can see what they need to know and do.

The AI input

The approved policy or decision, effective date, affected roles, required actions, owner and support route.

The draft output

A concise draft organised into change, rationale, actions, unchanged arrangements and questions.

What AI did well

AI can reduce duplication, improve hierarchy and translate formal wording into plainer language.

What AI missed

It may remove an exception, imply an unapproved assurance or turn an unresolved issue into a definite instruction.

Human checks

Compare every action with the source
Confirm the effective date and owner
Preserve exceptions and escalation routes
Check accessibility and tone
Obtain accountable approval before circulation

Final version

An approved briefing with traceable source references, clear actions and a named route for questions.

Worked example 03

Compare two public guidance documents

Understand where two current documents agree, differ or apply to different populations and settings.

The AI input

The full documents, a focused comparison question and agreed fields such as scope, recommendation, evidence and uncertainty.

The draft output

A comparison table with page references and a list of questions requiring professional interpretation.

What AI did well

AI can impose a consistent structure and reveal differences in wording or coverage.

What AI missed

A real difference may reflect publication date, legal status, population or method rather than disagreement.

Human checks

Read each executive summary and methods section
Verify quoted recommendations
Check jurisdiction and date
Identify superseded material
Resolve applicability through the correct governance route

Final version

A transparent comparison that separates extraction from the accountable interpretation and decision.

Worked example 04

Create a fictional training case study

Develop a realistic scenario that supports discussion without reusing a real person’s story.

The AI input

Learning outcome, audience, generic UK setting, difficulty, time and boundaries.

The draft output

A fictional scenario with strengths, uncertainty, several possible responses, questions and facilitator notes.

What AI did well

AI can quickly vary context and generate more than one perspective.

What AI missed

It may reproduce stereotypes, create a single ‘correct’ answer or include implausible treatment and legal details.

Human checks

Remove stereotypes and moralising language
Confirm it cannot identify a real person
Verify factual details
Check psychological safety
Ensure questions fit participants’ responsibilities

Final version

A clearly labelled fictional case that encourages reflection, uncertainty and humane practice rather than diagnosis.

Worked example 05

Structure a service evaluation

Create a balanced framework for understanding implementation, experience, outcomes and unintended effects.

The AI input

The intervention logic, intended users, agreed outcomes, available data, constraints and stakeholder priorities.

The draft output

Evaluation questions, possible evidence, equality considerations, limitations and review points.

What AI did well

AI can widen the question set and connect process, outcome and experience measures.

What AI missed

It cannot establish causation, select measures without context or represent the views of people who were not involved.

Human checks

Agree the purpose before choosing measures
Use validated measures where appropriate
Involve service users and staff
Set a baseline
Document missing data and rival explanations

Final version

A proportionate evaluation plan agreed by accountable leads and affected stakeholders before implementation begins.

Worked example 06

Simplify technical information

Adapt a technical explanation for a frontline or public audience without weakening safety information.

The AI input

Verified source text, audience, reading aim, terms that must remain and essential warnings.

The draft output

A clearer version with short sections, explained terminology and a final list of points to check.

What AI did well

AI can shorten sentences, explain jargon and improve navigation through complex material.

What AI missed

It may remove uncertainty, change a threshold or make a conditional statement sound universal.

Human checks

Compare meaning sentence by sentence
Preserve cautions and uncertainty
Check reading level with real users
Verify all health and legal content
Approve the final version through the usual route

Final version

Accessible text that remains faithful to the authoritative source and is visibly dated and reviewed.