AI
Corporate AI Training
For AI training for employees to become hands-on practice, the organizer prepares a package before the session: repeatable work tasks, files cleared for use, verified access to the tools, and a clear criterion for judging whether the output is good enough. Without this package, the facilitator works from their own examples and the group just watches.
Why does an AI session turn into a demo?
An AI session turns into a demo the moment participants have nothing to work on themselves. Standard pre-workshop instructions look like this: register for the service, bring a laptop, come with questions. Technically everyone complies. Practically, the key question goes unanswered: what tasks will the group actually practice on?
When participants have no material of their own, the facilitator switches to their own examples. The group watches someone else's process, takes notes, asks questions. After the session, everyone thinks "That was interesting, I should try it" - and does not, because they have no idea where to start on their own work. The session became a demo, though nobody planned it that way.
The fix is a concrete package assembled before the session. A task passport is a document that describes one practice task: what to do, which files to use, and who will check the result. The package the facilitator receives before the session can include one or more such passports. This article explains how to put one together.
How do you choose a task for AI training for employees?
A good practice task is a slice of a real, repeating process where the AI creates or transforms something useful, and the participant can check whether it worked.
Start with a map of the department's repeatable processes. Not wishes - "we want to automate everything" - but a concrete list: what gets done regularly, by whom, with what input and what output. For each process, answer four questions: how often it happens, what the input is, what the output is, and who accepts the result.
From that list, candidates emerge. Good candidates have enough frequency that participants know the context well, a specific file or text input, a clear output format, and a person who can say "this works" or "this is wrong." The frequency benchmark depends on the group: a task someone runs every week will feel more familiar than one they run once a month.
Not every work task fits a session. Here are three common cases where a task is better left out.
The task is too rare. An employee prepares a consolidated report for the board once a quarter. The task is real and important. But without fresh data and enough context, it cannot be practiced in a single session. Exception: if a completed archived report exists with cleared data, it can serve as the input material.
The task is too large. "Analyze the customer base" or "write a strategy" is a project, not an exercise. It has no clear starting point, no measurable finish line, and no realistic completion time inside a session. The participant never knows when they are done.
No owner, no data permission. Someone brings a file with real client data because "it is the best example." Nobody checked whether uploading that data to an external service is allowed. NIST AI 600-1 (2024) is a voluntary cross-industry framework from the US National Institute of Standards and Technology - not a law and not a ready-made company policy. It describes re-identification - recovering a person's identity from indirect data - as a distinct risk class for generative systems (full text). That is reason enough to check the company data policy before the session, not after.
Three distinct problems sit behind these cases: developing a skill, ensuring data quality and permission, and integrating with working systems. They are solved differently and, as a rule, not in a single session. Mixing all three into one exercise risks solving none of them in the time available.
After building the list of candidates, assign each one of three statuses:
| Status | What it means |
|---|---|
| Ready to practice | The input is accessible and cleared, the output is verifiable, the exercise is completable in one session |
| Needs a specific fix | The task makes sense but has a removable blocker: an inconvenient file format, missing data permission, a regulation document too long and unstructured. The blocker is described and an owner is assigned |
| Not suitable for this format | The task is important but cannot be practiced here: too rare without available archive material, too large, requires connecting to systems that are not ready, or the data cannot be used in any form |
"Not suitable for this format" is not a judgment of business importance. An important task can receive this status and move to a separate track - a dedicated preparation program with the right environment and data in place.
How to prepare materials: eight steps
Preparation is not a one-off action by the organizer. It is a sequence of steps involving four roles: the department head, HR or the training organizer, IT, and the facilitator. The participant can also help prepare their own task where the format allows.
Step 1. Collect repeatable processes
The department head lists the tasks participants do regularly. Not ideal AI use cases, just what actually gets done. Honesty matters at this stage: if a task happens once a year, mark it as rare and evaluate it separately, checking whether archive material exists.
Step 2. Describe current steps and the bottleneck
For each candidate, describe how the task is done today: where the input comes from, what happens to it, where the result goes, and where time or quality leaks out. The bottleneck is the point where AI could help. If no bottleneck can be stated, the task probably does not fit this session format.
Step 3. Select one specific slice for practice
From the described process, pick one step, not the whole thing. Not "prepare a commercial proposal" but "write the cover letter to a proposal using a template and client data." The participant must be able to finish the exercise in one session and get a checkable result.
Step 4. Prepare cleared files
The organizer checks each file: can it be uploaded to an external service? Does it contain client personal data, trade secrets, or internal correspondence? If the answer is unknown, find out before the session, not during it.
File structure matters more than it looks. A scan with no recognized text is a different input than the same table in Excel. A long PDF without headings creates a different exercise than a structured document. To check whether a file is suitable, upload it and process the relevant section in the chosen tool in advance.
If real data cannot be used, two paths exist: take a public document of the right structure, or prepare a properly anonymized version. Preparing an anonymized input is covered in the section on difficult cases.
Step 5. Name the acceptance criterion
Who decides the result is good, and by what signs? It could be the participant ("the letter is sendable without edits"), the manager ("the structure matches our standard"), or an external recipient. Without a criterion, the exercise ends with "so what now?" A good criterion is two or three concrete signs, not "quality text."
Step 6. Check access under a normal account
IT verifies that every participant can log in to the required service under their own account - not an admin account, not a shared login. A corporate proxy is an intermediate server that filters employees' internet traffic. An admin can often open what this filter or the browser policy blocks for a regular employee.
The check covers the participant's full path: download the file, find it on the device, open it, run the cleared trial task, save the result, and hand it to the recipient. Simply reaching the login screen is not enough.
Problems during sessions often appear before anyone touches the AI. A participant cannot find the downloaded file. The service is blocked by a corporate filter. The device runs a different operating system. These are observations from specific sessions - enough reason to walk the full participant path before the session day, not on it.
If expanding permissions for some participants is impossible, a fallback format is needed - chosen in advance, not improvised. If each pair has a verified device, the pair works together and saves the result on one device with a clear file name. If only the facilitator has a verified device, the format shifts to a demonstration with walkthrough: participants analyze the result, suggest edits, discuss criteria. Both options require a prior environment and data check. Assign the person participants call when something breaks, and set the trigger point for switching to the fallback format, before the session.
Step 7. Run the exercise fully before the session
The organizer or facilitator completes the exercise end to end, using the same files and access the participants will have. The goal is confirming the task is solvable, the input is correct, the output saves, and no unexpected blockers exist. This is a trial run in advance, not a rehearsal of the session.
If the task involves an external service, the run must not automatically write results into the company's working systems. Working systems are the CRM (a customer relationship management database where client records live), shared storage, and email where real data is kept. A practice result must not automatically land in the CRM, send a message to a client, or modify a shared document without explicit confirmation. IT confirms this condition in advance.
Step 8. Hand the facilitator the agreed package
The end product of all preparation is a package the facilitator receives before the session. It includes the task passports, the files, and confirmation of access. A facilitator who was not involved in preparing it should be able to understand the task, find the files, run the practice, and judge the result without extra questions. If that is not possible, the package is not ready.
How do you handle difficult cases with files and data?
Sometimes a task makes sense but a file or access problem creates a blocker. The general rule: check permission and file format first; if in doubt, replace the input or move the task to a separate track. Here are five common cases.
Data exists but permission does not
An employee brings a real file that fits the exercise well but contains client personal data or confidential information. Uploading it to an external service requires checking company policy and the service's terms of use.
Surface-level anonymization - replacing client names with "Client A" and amounts with placeholder numbers - does not make the file safe automatically. Addresses, contact details, dates, free text, hidden sheets, and indirect identifiers may remain. Re-identification is the process of recovering a person's identity from such indirect data. NISTIR 8053 (2015) explains that replacing direct identifiers alone does not eliminate this residual risk (full text). The authorized data owner must confirm the specific input after checking that copy for indirect identifiers and hidden content.
If a safe copy cannot be confirmed, two paths remain. First, build an agreed synthetic example that preserves the structural dependencies: field types, format, calculation logic. Second, remove the task from this session and move it to a track where the environment and data are prepared properly. This article does not replace a legal review of company policy or a specific service's terms.
If the company uses an enterprise version of the service with a confidentiality agreement, that is a separate question for IT and legal. Public availability of a document is not automatic permission to upload it: check the licensing terms of the specific source for the chosen environment. NIST AI 600-1 (section GV-6.1-001) separately requires accounting for third-party rights when working with external systems (full text).
File exists but the format is inconvenient
A scanned contract is not the same input as the contract in text. A table saved as an image is not the same as an editable table. A long PDF without headings creates a different exercise than a structured document.
If the file is in an inconvenient format, that is a blocker with status "needs a specific fix": convert the file, choose a different input, or make working with the format part of the exercise - but then it is a different exercise than the one planned.
Participants hold different roles
If the group is mixed - managers, analysts, support specialists - one task will not suit everyone equally well. Either choose the most universal task that everyone understands, or prepare two or three input variants with the same exercise structure.
If a manager works in a different tool or on a different device, prepare the same files in a compatible format for them. That is also a valid preparation option.
The participant sees no link to real work
The organizer selects a task by formal criteria, but the participant does not understand why this process was chosen. The task passport should explain: why this process was picked and what could change if it is done with AI. That explanation lowers the chance the participant treats the exercise as a formality.
Combining a skills check with an environment check
This is possible, but it requires an explicit separation of goals: what exactly is being assessed - the ability to work with the tool, or the readiness of the infrastructure? Mixing the goals makes it harder to judge the result. Decide on this format in advance, not mid-session.
The task passport: a template for one practice exercise
A task passport is a document that records everything needed for one practice exercise. The four roles agree on this document, not on a verbal arrangement. The passport is written for each task, not for each participant.
Passport fields and what to put in them:
Process identifier - a short label so the task can be found in email threads and file lists.
Owner - a specific person responsible for this process in the department who can say whether the result is correct. Assigned by the organizer.
Frequency - how often the task is performed: daily, several times a week, once a month. Affects how familiar participants are with the context.
Current steps - how the task is done today: where the input comes from, what happens to it, where the result goes. This field explains the connection to real work for the participant.
Input files - specific files with formats. Not "a table with data" but exact references: file name, sheet, range.
Approved environment - which service or tool is used in the session, and whether IT and the data owner have confirmed it for these specific files.
File source and usage terms - where the file came from and whether licensing conditions have been checked for uploading to the chosen environment. For public documents this is not obvious by default.
Expected output - what should result: text, table, list, filled template. Format and approximate volume.
Recipient - who receives the result and in what form.
Acceptance criteria - by what signs the result counts as good. Prefer two or three concrete signs over "quality text."
Baseline time - how long this task currently takes an experienced employee. This is a reference point for the participant, not a promise of speedup.
Known constraints - what must not be done with these data, which formats do not work, which steps require manual checking.
Technical check - a note that a trial run was completed: by whom, when, and with what result. Empty "owner" and "date" fields do not mean readiness.
Readiness decision - one of three statuses: "ready to practice," "needs a specific fix" with a description of what exactly, or "not suitable for this format" with an explanation.
Owner for the fix - if the status is "needs a specific fix," a specific person and a deadline.
Check date - when the passport was last reviewed substantively. Updating the date is not the same as running a check: if environment rules, material rights, or acceptance criteria changed, the relevant fields need a substantive review.
A partially filled passport for a financial exercise
Below is a model example of a filled passport for an exercise using a public financial document. It illustrates structure - it is not a record of a real session and not a description of Apple's real workflows.
The input material is a public document: Apple's Form 10-K for the fiscal year ending September 27, 2025 (FY2025). A 10-K is the annual report that US companies are required to file with the SEC (the Securities and Exchange Commission, the US capital markets regulator). Available on the SEC EDGAR public filing registry: https://www.sec.gov/Archives/edgar/data/320193/000032019325000079/aapl-20250927.htm. Values from the Consolidated Statements of Operations (the income statement): Net sales (revenue) 416,161 million USD, Gross margin (gross profit) 195,201 million USD, period FY2025.
In the proposed training scenario, the participant passes an excerpt of the report to the tool to structure the named line items, checks the extracted values against the source, and calculates the percentage themselves. Below are author-written samples of a prompt and a response showing the shape of the exercise. These are author examples, not a record of a real model run.
Author-written sample prompt:
Structure two values from Apple FY2025 Consolidated Statements of Operations (year ended September 27, 2025): Net sales and Gross margin. Output a table with columns "Line item," "Value, million USD," "Period." Recipient of the table: the session facilitator for evaluation.
Author-written sample response:
Line item Value, million USD Period Net sales 416,161 FY2025 Gross margin 195,201 FY2025
After this, the participant opens the same section in the SEC document (printed page 29, Consolidated Statements of Operations), checks both figures against the report, and calculates gross margin by hand: 195,201 divided by 416,161, multiplied by 100, equals 46.9%. The AI extracts and structures the lines. The person confirms the values against the source and does the math.
A finished example does not close the technical check. In the passport, "Technical check" stays "not completed" until the organizer runs the full trial in the chosen environment.
Process identifier: financial-summary-gross-margin
Owner: assigned by the organizer from among participants with a finance background
Frequency: an archived annual FY2025 document is used for this example; the organizer fills in the real frequency based on the department's process
Current steps: an analyst receives the 10-K, locates the Consolidated Statements of Operations (the income statement), writes out the key line item values, checks them against the source, calculates derived figures by hand, and formats a brief summary. The organizer fills in the specific steps of the department's real process. This is the original manual work; AI is practised in the training exercise.
Input files: public Apple 10-K FY2025 HTML document (SEC EDGAR), Consolidated Statements of Operations section; Net sales and Gross margin values for the period ending September 27, 2025
Approved environment: requires IT and facilitator confirmation for the chosen tool; public availability of the document does not mean automatic permission to upload it to any external system
File source and usage terms: SEC EDGAR, public regulatory filing; the usage terms of the specific tool are checked separately before the session
Expected output: a two-row table (Net sales, Gross margin) with columns "Line item," "Value, million USD," "Period," plus the Gross margin percentage formula and the checked value; the calculation is performed by the participant by hand, not accepted on the basis of model output without verification
Recipient: the facilitator for evaluation; in a real process, the manager or financial analyst
Acceptance criteria: values match the report (Net sales 416,161 million USD, Gross margin 195,201 million USD); units and period stated explicitly; Gross margin percentage calculated correctly (195,201 / 416,161 x 100 = 46.9%); no new interpretations or commitments added to the table
Baseline time: assessed by the organizer based on the department's real process; not set for this example
Known constraints: document is in English; the HTML version requires checking that numbers copy without formatting artifacts; the percentage calculation is verified by the participant by hand
Technical check: not completed; the organizer records who, when, and with what result before the session
Readiness decision: needs a specific fix - technical check not completed
Owner for the fix: assigned by the organizer
Check date: not set
This example shows how the passport fields work together. The input material is specific, the acceptance criteria are checkable, and the participant's role is described explicitly. The readiness decision honestly reflects that the technical check has not yet happened. A facilitator receiving such a passport knows exactly what to do before the session.
For more on building a full training program, see the AI training page.
Frequently asked questions
How many tasks should be prepared for one session?
It depends on session length and group composition. For a two-hour practical session, a reasonable starting benchmark is two to three tasks with different input types - but that is a parameter of the proposed format, not a universal rule. One complex task with an unfamiliar data format or a mixed group can fill the whole session. One well-prepared task beats several with status "needs a specific fix." Narrow the list and prepare the chosen tasks thoroughly, rather than taking many and failing to clear the blockers before the session.
Can a participant help prepare their own task?
Yes, and it is useful. The participant knows their process well and can help describe the current steps more accurately, name the real input, and formulate the acceptance criterion. In the process described here, the facilitator represents participant interests by default. But a participant's direct involvement in preparing and checking their own task is not excluded if the format allows it and the organizer agrees. This is especially valuable when the task is highly specialized and only the employee knows the details.
Can tasks from other departments be used?
Yes, if the process owner from that department participates in preparing the passport and confirms the acceptance criteria. Without an owner there is no feedback - the participant cannot know whether the result is correct. That lowers the training value of the exercise. The owner from another department may also know data constraints the organizer is unaware of. Agree with the owner before preparing the passport, not after.
What if IT cannot verify access before the session?
Choose a fallback format in advance, not on the day of the session. If each pair has a verified device with the right access, the pair works together and saves the result on one device. If only the facilitator has a verified device, the format shifts to a demonstration with walkthrough: participants analyze the result, suggest edits, discuss criteria. Assign the person responsible for switching formats and set the decision point before the session.
Is the passport per participant or per task?
Per task. The passport describes the whole exercise: what to do, which files to use, who checks the result and by what signs. The whole group works from one passport - each participant completes the task independently but follows the same instructions. If the session includes several different tasks, each task gets its own passport.
What if real data cannot be used and no public analog exists?
Prepare an agreed synthetic example that preserves the structural dependencies: field types, format, calculation logic. The structure must mirror the real one so participants work with familiar data patterns. The authorized data owner decides whether such an input is acceptable, after checking the specific copy - not the organizer unilaterally. If even that cannot be confirmed, move the task to a track with a properly prepared environment.
Must the whole process be redone for every session?
A passport, once prepared, can be updated. But updating the date is not the same as running a check. If environment rules, material rights, participant composition, or acceptance criteria changed, the relevant fields need a substantive review. A blank "check date" field means one thing: the check has not been done. Over time the department builds a library of verified tasks, and future sessions become much easier to prepare.
Conclusion
An AI session becomes real practice when every participant has a concrete task, cleared files, verified access, and a clear criterion for the result. The task passport is the form that forces you to gather all of this before the session rather than sorting it out during. Start with one task from a real department process, walk through all the preparation steps, and check that the passport is clear to the facilitator without extra questions.
To discuss task preparation for a specific department or the session format, write on Telegram @shimaoz or at hello@majento.ai.