Draft for approval.
Problem Starts
The problem starts after generic training
A company runs an AI workshop for finance, sales, HR and creative staff. Everyone sees the same demonstration and learns the same prompting formula. The session feels useful, then each team returns to work with different information, responsibilities and risks.
A finance manager needs figures traced to an approved source. A salesperson must avoid inventing a discount or delivery promise. An HR manager handles policy questions that can affect a colleague. A creative team needs original work that still respects the brief, brand and usage rights.
Shared foundations help, but useful practice has to reflect the job. The OpenAI Academy resource hub separates customer feedback analysis, campaign work, customer replies and staff training into distinct workflows. Each receives its own knowledge, instructions, guardrails and review step.
The practical lesson is to train around a bounded piece of work rather than a general tour of an AI tool.
Small Trial
Run one small trial per role
Choose a task that already happens, has a clear owner and can stop before anything reaches a customer, employee or financial decision. Use fictional, anonymised or otherwise approved material for the first attempt.
| Team | Smallest useful trial | What the tool receives | What remains human-led | |---|---|---|---| | Finance | Draft commentary on a small set of budget versus actual variances | An approved sample table, definitions for each column, reporting period and materiality rules | Recalculating figures, interpreting accounting treatment and approving the commentary | | Sales | Draft a follow-up email from an anonymised meeting note | The note, approved offer details, audience, tone and permitted next steps | Confirming customer facts, commercial terms, commitments and whether to send | | HR | Turn an approved policy excerpt into a manager checklist for a fictional scenario | The policy text, intended reader, scope and required escalation points | Interpreting the real case, checking current policy and deciding what happens to an employee | | Creative | Adapt an approved campaign brief into one channel-specific draft | The brief, audience, brand guidance, channel limits and approved claims | Selecting the idea, checking originality and rights, refining the work and approving release |
Each trial should produce one reviewable output. If the team cannot identify the person who will check it, the trial is too vague to run.
Give Tool
Give the tool enough to work with
OpenAI Academy describes a prompting checklist built around task, context and expectation. For team practice, those elements need operational detail.
The task should name the action and its boundary. “Draft variance commentary from these approved figures” gives the finance reviewer something concrete to inspect. “Help with the monthly report” leaves too much room for interpretation.
Context should include the audience, source material, definitions, relevant policy and examples of acceptable work. It should also explain unfamiliar fields and internal language. A sales tool cannot distinguish an approved package from an informal suggestion unless the team provides that distinction.
The expectation should describe the output’s format, length, tone and required sections. It should state what the tool must do when information is missing. Asking for a short clarification or flagging an unresolved field is safer than allowing a plausible guess.
Give every trial a compact workflow card containing the task owner, approved inputs, prohibited inputs, expected output, clarification rule and review method. This card matters more than a polished prompt because it records how the team intends to use the output.
Keep Sensitive
Keep sensitive and consequential information outside the first trial
The first exercise does not need live confidential data. Each team can test the method with a small approved sample while the organisation confirms its data policy, tool configuration and access controls.
| Team | Keep outside the trial | Why the boundary matters | |---|---|---| | Finance | Bank details, credentials, personal payroll data, identifiable invoices and unreleased results | The exercise only needs enough structure to test analysis and traceability | | Sales | Personal contact details, confidential negotiations, undocumented discounts and private customer material | The draft should rely on approved terms rather than hidden commercial context | | HR | Employee case files, health information, grievance details and identifiable performance records | A fictional scenario can test the workflow without exposing a colleague’s circumstances | | Creative | Unreleased strategy, customer likenesses, rights-uncleared material and licensed assets that the tool is not approved to receive | The team can test briefing and review with cleared or synthetic material |
These are starting boundaries for the trial. An organisation may later permit some categories inside an approved system with suitable governance. That decision belongs to the people responsible for security, privacy, legal obligations and the relevant business function.
Match Human
Match human review to the consequence
A quick read for tone is not enough. Reviewers should trace the output back to the material the tool received and record what they changed.
Finance reviewers should recalculate every figure mentioned in the commentary. They should check signs, periods, totals and the meaning of each variance against the source table.
Sales reviewers should compare names, needs, prices, dates and next steps with the approved meeting note and offer information. Any new commitment should be removed or confirmed before the message leaves the business.
HR reviewers should locate each policy statement in the approved source. They should check whether the scenario needs a manager, HR specialist or legal adviser rather than allowing the generated checklist to decide a real case.
Creative reviewers should assess factual claims, brand fit, channel requirements and usage rights. They should also decide whether the idea is good enough. An AI tool can produce options, but the creative lead remains responsible for taste, context and release.
A simple review log can mark each material statement as supported, unsupported or missing context. It should also capture the reviewer’s corrections. Repeated corrections reveal which instruction, source file or boundary needs improvement.
Test Workflow
Test the workflow before colleagues rely on it
The OpenAI Academy build recipe includes normal cases, a missing-information case, an edge case and a human check. Apply the same test structure to each role rather than judging the workflow from its best example.
| Team | Normal test | Missing-information test | Realistic edge case | |---|---|---|---| | Finance | All columns are defined and the totals reconcile | The reporting period is absent | A refund arrives after the reporting cut-off and could be assigned to the wrong period | | Sales | The meeting note and approved offer contain all required details | The customer’s required delivery date is absent | The note mentions a verbal concession that does not appear in the approved offer | | HR | One approved policy clearly covers the fictional scenario | The employee’s role or location is absent | Two policy excerpts appear to point towards different next steps | | Creative | The brief contains an approved message, audience and channel | The required call to action is absent | The strongest concept depends on an image or claim that has not been cleared for use |
The expected behaviour should be explicit. The tool completes the normal case from the supplied evidence. It asks for the missing fact or leaves a labelled gap in the second case. It flags the edge case for a named human rather than smoothing over the conflict.
A fluent answer can still fail the test. Pass the workflow only when the output stays within its sources, exposes uncertainty and gives the reviewer a practical route to a decision.
Turn Useful
Turn a useful trial into team practice
Repeatability comes from preserving the working method, not from asking colleagues to copy a successful chat.
Once a trial passes its tests, save the workflow card, approved source set, output format, test cases and review checklist together. Give them an owner who can update the material when a policy, offer, reporting definition or brand rule changes. Record the version used for consequential work so a reviewer can reconstruct what the tool knew.
Keep training close to the task. Finance staff should practise tracing claims to cells and definitions. Sales staff should practise spotting invented commitments. HR staff should practise escalation and source checking. Creative staff should practise evaluating choices, claims and rights. A shared introduction can cover basic use and company policy, then each function needs examples drawn from its own work.
Anthropic’s Enterprise AI Transformation Guide frames adoption around governance, targeted pilots and structured training. Anthropic also says its guide covers measures across adoption, efficiency, quality and satisfaction. A small business can apply that structure without turning the exercise into a large transformation programme.
For each workflow, record whether staff use it, how long review takes, how many material corrections reviewers make and why an output is rejected. Those observations help the team decide whether to revise the instructions, improve the source material or stop the workflow.
Practical Starting
A practical starting point
Choose one team and one low-risk task that produces a draft rather than a decision. Prepare a fictional or approved sample, name the reviewer and write down what must stay outside the tool. Then run the normal case, the missing-information case and the edge case.
Keep the workflow only if the reviewer can trace its claims, identify its gaps and remain in control of the consequential judgement. After that, document the useful version and teach it with examples from the team’s real responsibilities.
Finance, sales, HR and creative teams can share an AI policy. Their practice should still reflect the work they do, the information they handle and the mistakes they are accountable for catching.
Checked sources