Guide path
How to Practise Using AI at Work Without Exposing Confidential Company Information
Use this evidence-led article to understand the topic, compare practical options, and choose a concrete next step. Then continue with the relevant guide, prompt library, or course only when it matches the work you actually need to complete, without random browsing, unsupported claims, or unnecessary purchases that do not fit your goal.
Open the curated guide layer before you pick a course or prompt pack.
Jump to the most relevant AI path for your profession.
Turn article ideas into reusable prompt systems.
Download free prompt packs tied to roles, workflows, and use cases.
Compare options before you spend more time or money.
A practical method for rehearsing workplace AI tasks with synthetic inputs, approval gates and a final privacy check.
Key Takeaways
Guide stack
Most readers should leave with one of three next steps: a role guide, a prompt library section, or a course that matches the same problem.
Reader FAQ
If you want faster execution, open the prompt library. If you want a bigger decision, open the role guides or the course catalog.
Yes. Start with the guide hub, then use the sample lesson path or the prompt library before committing to membership.
Choose the next step that matches your job to be done, not the most popular page.
Keep learning
Continue with practical courses connected to this topic.
Free flagship course: learn the portable system for asking, choosing, reviewing, and delivering with ChatGPT, Gemini, and Claude.
View course →
The flagship TakeAICourse program for applying AI at real work in 30 days.
View course →
The short answer: practise the shape of the task, not the live company material. Start with a clearly defined work outcome, classify every input before it reaches an AI tool, and replace real names, values, events and documents with invented equivalents. If the task cannot remain useful after that replacement, stop and seek approval for an authorised method. Before submitting anything, check both the prompt and any attachment as though they were about to become visible outside your organisation.
This is a rehearsal method, not permission to process workplace data. An employer’s policies, contracts, approved-tool list and approval process still apply. Removing obvious names can reduce exposure, but it does not make an otherwise sensitive document safe by default.
The SHAPE framework separates the skill you want to learn from the information you must protect:
The key distinction is between structural fidelity and factual fidelity. A good practice pack can preserve the number of sections, conflicting priorities, missing fields, awkward formatting and decision points in a task. It does not need to preserve the real people, customers, figures, dates, product names or incidents. Preserve the difficulty; replace the facts.
Use this deliberately conservative decision tree for each proposed input, including text typed from memory:
FAQ
Sources
“Public”, “internal” and “confidential” labels can mean different things in different organisations. Use your organisation’s own classification scheme where one exists; this tree is only a practice filter.
| Proposed material | Default practice decision | Safer rehearsal move | Why the obvious edit may be insufficient |
|---|---|---|---|
| A real customer email | Replace | Write a fictional message with a different issue, tone and history | Deleting the name can leave order details, wording or circumstances that identify the sender |
| An internal report | Replace | Recreate its section structure with invented projects and values | Titles, comments, hidden context and exact combinations of facts may remain distinctive |
| A spreadsheet export | Replace | Make a small synthetic table with invented columns and plausible-but-fictional entries | Cells, file names, formulas, notes and metadata can carry information beyond the visible rows |
| A contract, legal note or personnel record | Stop | Practise on a blank template or a wholly fictional scenario | The learning benefit rarely requires the real record |
| Credentials, tokens, private keys or access links | Stop | Use unmistakable placeholders such as [DEMO_TOKEN_NOT_REAL] | These are access mechanisms, not practice content |
| Material from an approved public page | Check, then use | Save the public URL and extract only what the exercise needs | Public availability does not by itself establish that every tool or reuse is approved |
| A brief recalled from a recent internal incident | Replace from first principles | Change the setting, actors, sequence, values and outcome | Memory-based prompts can disclose protected details even without an attachment |
| A fictional brief made specifically for training | Use after review | Mark it “FICTIONAL PRACTICE DATA” and inspect it for accidental resemblance | Fiction can still be too close to a real person or event |
Redaction is best treated as a review technique, not the foundation of the exercise. Starting from a blank page makes it easier to avoid preserving hidden or distinctive details.
A useful pack has four parts: the fictional input, the task instruction, a quality rubric and a privacy boundary. The following templates can be reused across common work formats.
FICTIONAL PRACTICE DATA
Skill: [summarise / restructure / compare / draft]
Audience: [invented role, not a real colleague]
Input structure: [sections, length, conflicts and missing details to simulate]
Invented subject: [fictional organisation/project]
Must include: [three or four synthetic facts]
Must not include: any real names, wording, values, events or file excerpts
Quality check: [accuracy to synthetic input, clarity, unanswered questions]
Create the document from scratch. If matching the original layout matters, describe the layout generically rather than uploading the original as a model.
FICTIONAL PRACTICE DATA
Skill: [clean / categorise / explain a formula / propose a chart]
Columns: [generic field names]
Rows: [small set of invented records]
Deliberate complications: [blank cell, duplicate, inconsistent date format]
Expected checks: [totals reconcile, assumptions labelled, no invented rows]
Boundary: no copied cells, identifiers, account numbers or exact internal values
Use obviously invented entity names and change more than the numbers. Recreate column names where those names are themselves internal, and avoid reproducing a unique combination of categories.
FICTIONAL PRACTICE DATA
Skill: [write a response / adjust tone / extract actions]
Sender: [invented person and organisation]
Situation: [generic fictional event]
Constraints: [deadline, tone, two unresolved questions]
Output: [draft plus assumptions and questions]
Boundary: no real message text, signature, contact detail or case history
A synthetic pack should be difficult enough to exercise judgement. Add ambiguity, incomplete information or competing priorities—but invent those complications rather than borrowing them.
This example is entirely fictional.
Suppose someone wants to practise turning a troubled project update into an executive summary. Their live source contains colleague names, a customer name, budget figures, a delayed delivery and comments about performance. Pasting it into a general AI assistant would conflict with the practice boundary.
First, define the transferable skill:
Produce a six-sentence status summary that distinguishes facts, risks, decisions and missing information.
Next, create a new brief without looking line by line at the original:
FICTIONAL PRACTICE DATA
Project: Cedar Lantern, an invented office move
Audience: fictional steering group
Status: furniture delivery moved from 12 May to 19 May
Budget: use categories only—on plan, watch, exception—not currency values
Risk: the training room may be unavailable for the first week
Decision needed: choose remote induction or rent a temporary room
Missing information: supplier has not confirmed weekend access
People: Project Lead A, Facilities Contact B, Supplier C
Then give the practice instruction:
Using only the fictional brief, draft a six-sentence steering-group update.
Separate confirmed facts from assumptions. State the decision required and the
missing information. Do not add causes, costs, names or dates not present.
After the draft, list any statement that needs human verification.
Finally, score the result against the synthetic source: Did it keep the revised date accurate? Did it label the unconfirmed access issue? Did it avoid inventing a cause or cost? This tests summarisation, uncertainty handling and review without transferring the live brief.
Notice what was preserved: a schedule change, a downstream risk, a decision and an information gap. What was replaced: the organisation, project, people, values, setting and wording. The exercise keeps the reasoning challenge while discarding factual fidelity.
Pause before any move from fictional practice to live work. A “yes” to every item below is the minimum gate for continuing; a “no” or “not sure” means stop and use the synthetic exercise only.
This checklist does not create approval. It helps reveal when approval is missing.
Run a separate review after drafting the prompt and before selecting Send or uploading a file:
For guided learning beyond this kit, the course catalogue provides the site’s current structured course paths. Keep using fictional practice inputs unless your workplace has authorised something else.
Synthetic practice is useful for learning a workflow, but it cannot prove that the workflow is safe for real company information. It may omit the messy relationships, rare cases and contextual clues present in live data. It also cannot determine whether a particular tool, account or task is approved in your organisation.
Redaction and replacement can fail when details remain identifiable in combination, when copied files contain material outside the visible page, or when an apparently fictional scenario closely mirrors a real event. This article does not assess legal duties, contractual restrictions, industry rules, vendor terms, retention behaviour or technical security controls. Those are context-specific and require current, authorised guidance.
There is also a learning trade-off: the more aggressively you simplify the practice data, the less representative the task may become. Resolve that by preserving structural complexity—missing fields, conflicting priorities and review criteria—rather than restoring sensitive facts.
No. Names are only one route to identification. Exact values, dates, wording, unusual events, file properties and combinations of details can still reveal the source. For practice, rebuild the material with invented facts rather than editing a copy of the live document.
A newly created synthetic spreadsheet is the safer rehearsal format. A real workbook may contain informative file names, hidden sheets, formulas, comments, links or structure. Create a clean file and reproduce only the generic complexity needed for the skill.
Separate them. Record the approved public source if using it is permitted, but recreate the analysis task with fictional notes. Do not assume the combined document is public because some source material is public.
The boundary applies to the information, not merely the file format. Typed, pasted, summarised and memory-based details all need the same classification and approval check.
Realistic enough to test the decisions, not realistic enough to reproduce the case. Preserve field types, contradictions, missing information and quality criteria. Change identities, exact values, chronology, setting, wording and outcomes.
Stop when the task depends on retaining real facts, when you cannot confidently classify the input, when the scenario remains recognisable after editing, or when approval for the tool and use is absent. Continue learning with a different fictional exercise while the workplace question is resolved.