From AI training to workflows employees actually use

Imagine a team finishing an AI workshop with good feedback and several impressive demonstrations. On Monday, the sales administrator opens the same inbox, the operations manager updates the same spreadsheet and accounts starts chasing the same missing information.

Nobody has decided which task should change, what a good AI-assisted result looks like or who will review it. The training may have been useful. The operating routine is still intact.

AI training for employees needs a bridge between learning the tool and relying on it during ordinary work. That bridge is a recurring task with approved inputs, a useful output and a person accountable for the result.

At Bosar, we view training as one part of implementation and adoption. Completing lessons establishes exposure to the material. Repeated, competent use on relevant work is the stronger test.

There are three levels of progress

Tool familiarity, task competence and a repeatable team workflow are different achievements. Measuring them separately helps managers see what support is still needed.

Someone may know how to upload a document but struggle to spot an incorrect summary. Another employee may produce excellent work in a personal chat, yet have no reusable instructions that a colleague can follow. Both have made progress. Neither automatically means the department has changed how it operates.

Tool familiarity

Employees should be able to navigate the approved tool, provide suitable context and understand basic access boundaries. They should also know that a plausible answer can be wrong.

Anthropic’s AI Fluency framework covers delegation, description, discernment and diligence. That gives training a broader foundation than writing a clever prompt: people need to choose suitable work, communicate it, evaluate the response and remain responsible for its use. Anthropic AI Fluency framework.

Task competence

The employee can use AI to complete a specific task and explain how the result was checked. For a quote brief, this includes recognising an outdated revision or a customer request that has not yet been agreed.

Competence also includes stopping. An employee who notices that the source data is incomplete and asks for clarification may be using the system better than someone who produces a polished answer quickly.

A repeatable workflow

The method has an owner, current instructions, a consistent output and a route for exceptions. Another trained colleague can use it without reconstructing the original author’s conversation history.

This is the point where personal experimentation starts becoming dependable team capacity. It does not require every task to run automatically.

Choose a recurring task with a clear finish

Start with work employees already perform and can judge. A broad assignment such as “use AI to improve operations” leaves too many choices. “Prepare the unresolved items for tomorrow’s dispatch meeting” gives the learner a defined purpose.

Good early candidates have accessible source information, a recognisable output and a reviewer who knows what correct looks like. They occur often enough to practise and are bounded enough that mistakes can be caught before they affect customers.

Begin with preparation and comparison

For sales, that could be a brief combining the current quote, latest recorded interaction and proposed next action. For operations, it could be a comparison between an accepted order and its delivery instructions. For accounts, it could be a draft explanation of why an invoice and purchase order do not match.

These are suggested applications, not reported client outcomes. Each requires a check that the source information is available and appropriate for the approved tool.

Avoid beginning with a task whose output the learner cannot evaluate. Giving a non-estimator responsibility for approving AI-generated manufacturing costs creates a judgment problem that a training session cannot solve.

Make each role’s practice specific

Generic examples are useful for introducing a tool. Department examples reveal the details that determine whether somebody will use it again.

The following three exercises show how a learning task can connect to business work while preserving review and accountability.

Sales: prepare the next conversation

Give the learner an approved quote and a small set of relevant communications. Ask for a brief containing the customer’s objective, current quote revision, unresolved questions, latest commitment and recommended next action.

Require a source reference beside commercially important statements. A requested discount must remain a request unless somebody authorised it. A customer saying a decision may happen next week must not become a promise to buy next week.

The salesperson checks the brief before using it and approves any resulting message. The exercise succeeds when the preparation becomes easier without introducing false certainty.

Operations: identify an order discrepancy

Use a sample order, acceptance email and delivery instruction. Ask the learner to identify differences in quantities, destination, delivery date and contact details.

The expected output is a short discrepancy table: agreed value, current system value, source and required confirmation. It should not silently rewrite the order.

Include one example where no discrepancy exists. Otherwise learners can mistake finding something to change for doing the task well. The reviewer needs to see that the method can return a clean result as well as a useful exception.

Accounts: prepare an invoice query

Provide approved copies of the invoice, relevant shipment and customer purchase order. Ask for an explanation of the apparent mismatch and a draft clarification request.

The learner verifies amounts, quantities and references against the originals. The AI must distinguish an absent freight line from evidence that freight was waived. Accounts decides whether any correction or credit is appropriate and approves the customer-facing message.

That distinction preserves the professional judgment of the accounts team while reducing the work required to assemble the facts.

Turn a useful exercise into a workflow card

A successful prompt can disappear inside one employee’s account. A short shared workflow card makes the method discoverable and maintainable.

The card should be specific enough to run the task without becoming a manual nobody reads. Here is a proposed structure for the sales brief:

Field	What to record
Trigger	Before an agreed follow-up or account review
Inputs	Current quote and approved, relevant customer history
Instructions	Separate confirmed facts, requests and unknown information
Output	One-page brief with source references and proposed next action
Checks	Customer, revision, amounts, commitments and contact preference
Owner	Sales manager responsible for the method
Reviewer	Rep responsible for the customer relationship
Exception	Missing or conflicting source goes back for clarification

Keep instructions and business knowledge current

Store shared guidance where employees can find the current version. Name who maintains product definitions, commercial rules and approved examples. If an old price list remains alongside a new one without a clear status, better prompting will not resolve ownership.

The business knowledge solution addresses that foundation. Useful knowledge has a source, an owner and a review process. A large folder of documents is not automatically reliable context.

When a workflow changes, update its card and test the examples again. Training materials should follow the operating process, rather than preserving a demonstration that no longer matches the business.

Choose the learning format around the work

Self-paced learning and workshops solve different practical problems. The right choice depends on employee confidence, available time, the task and how much company-specific discussion is needed.

Self-paced material gives people a repeatable foundation they can revisit. A workshop creates shared time to apply that foundation to a real process and resolve disagreements about how the process should work.

Our team training can involve self-paced learning or workshops depending on the engagement. These formats, supported practice and implementation work should be scoped explicitly; they are not automatically bundled together.

Protect time for application

Learning competes with daily work. A manager who allocates a course but no time for practice is asking employees to carry the transition on top of their existing workload.

Schedule a small number of real tasks for supervised application. Review the output, revise the method and repeat. If the new workflow adds effort without a clear benefit, investigate it rather than insisting that employees use AI for its own sake.

Microsoft’s employee-enablement guidance similarly emphasises ongoing, scenario-based learning and measures beyond purchased licences. It distinguishes assistants that support a person’s decision from systems that take actions across tools. That boundary is useful when deciding what the training must cover. Microsoft employee AI enablement.

A proposed 30-day adoption plan

Thirty days can provide a useful first review period for a small set of tasks. The following is a proposed plan, not a promise that every department will be transformed within a month.

Choose one accountable manager and a small group of employees who perform the work regularly. Agree what evidence will support continuing, changing or stopping each workflow.

Week one: establish the task and baseline

Choose one recurring task per participating role. Record the current inputs, steps, review requirements and typical problems. Time a small sample from start to accepted output, including interruptions and corrections where practical.

Confirm approved tools and data access before practice begins. Provide the foundation material and an example of an acceptable result. Also show a result that should be rejected, with the reason.

Week two: practise with completed work

Use past examples whose outcomes are known. Have each learner run the task, inspect the output and record the corrections needed. Reviewers should look for missed facts, unsupported conclusions and excessive checking effort.

Refine the instructions around recurring problems. Preserve difficult examples for later checks instead of deleting them to make the workflow appear more successful.

Week three: apply it to controlled live work

Move to current tasks with the normal accountable employee reviewing the result. Keep sending, approval and consequential system changes under the existing business controls.

Ask employees where they still fall back to the old method and why. The cause may be missing access, unclear instructions, slow output or a genuinely unsuitable task. Those findings are more useful than a blanket request for more enthusiasm.

Week four: decide what becomes standard

Compare accepted outputs, total task time, rework and repeated use with the baseline. Select the workflows that deserve to become normal practice and assign maintenance ownership.

For the rest, decide whether to simplify, support further or stop. A reliable result may justify a later implementation project to connect systems. The decision should follow demonstrated use, not the number of course certificates issued.

Measure use, quality and where capacity goes

Course completion, active users and task frequency show whether employees are participating. They do not establish business value on their own.

Measure accepted output quality and review effort alongside time. If a first draft takes two minutes but needs twenty minutes of correction, the two-minute figure is misleading. If preparation is faster but work waits for the same approval, report preparation time separately from end-to-end turnaround.

For an illustrative capacity calculation, assume ten employees each complete two recurring tasks a week and reduce total task time, including review, by ten minutes per task. That releases 200 minutes weekly. At an assumed fully loaded labour cost of AUD $60 an hour, the capacity is valued at AUD $200 a week, or AUD $9,600 across 48 working weeks. Training fees, practice time, licences and ongoing coaching must also be counted. The calculation is neither a client result nor a promised cash saving; managers still need to decide how that time will be used.

Expect differences between people and tasks

A field study of generative AI in customer support found that effects differed substantially across workers. It studied a particular deployment, so it cannot supply a universal improvement target for your sales, operations or finance team. NBER, Generative AI at Work.

Use your own baseline and record what happens to any released capacity. Does the rep prepare more carefully, handle a backlog or make more useful customer calls? Does accounts resolve queries sooner? Those are business outcomes worth examining.

Training may also reveal that the next improvement is software configuration or an integration. The configure, buy or build framework helps distinguish that requirement from a need for more lessons.

Frequently Asked Questions

Is self-paced learning enough, or do we need workshops?

Self-paced learning can provide the foundation when employees have time to practise and a manager supports application. Workshops are useful for company-specific processes and shared decisions. Choose the format around the work and the team’s needs, with the scope and follow-through agreed in advance.

Does every employee need to learn Claude Code?

No. Match the tool to the task, access requirements and employee’s responsibilities. Some roles need an approved assistant for reviewing documents or preparing briefs. Others may benefit from deeper workflow-building skills. Tool sophistication is not the measure of competence; reliable output is.

How do we protect company information during training?

Use approved accounts and confirm the access and data-handling arrangements for the selected service. Start exercises with suitable samples and limit inputs to what the task requires. Employees should know which information they may use and who can resolve uncertainty before they share it.

Who reviews AI-assisted work?

The person accountable for the business result should retain review responsibility, with specialist input where necessary. Early practice may need closer coaching. Customer communications, prices and consequential changes should follow the organisation’s approval rules. Training does not transfer responsibility to the model.

How do we know the training has worked?

Look for repeated use on relevant work, acceptable outputs, manageable review effort and a clear benefit compared with the previous method. Employees should be able to explain their checks and recognise when to stop. Completion records support that assessment, but cannot replace it.