A 200-person company does not have one AI training problem. It has nine, and they share almost nothing but the login. Your controller and your sales director both opened ChatGPT this morning. One of them must never paste what is on her screen into a consumer account; the other should be pasting more than he does. One needs to stop trusting numbers the tool produces; the other needs to stop writing outreach that reads like everyone else’s. A single company-wide session has to tell both of them something true, so it tells them nothing either can use on Monday. This is the case for AI training by role, and the whole argument of role-based AI enablement: one floor for everybody, then nine rooms.
Why AI Training by Role Beats One Company-Wide Deck
The clearest public evidence comes from a company that had every reason to get this right the first time. Moderna launched an internal AI academy in December 2021 and made an AI fundamentals course mandatory for all staff. Fortune reported in March 2024 that the company scrapped it: CHRO Tracey Franklin said some employees found the course too basic and others found it too complex, and the curriculum was rebuilt as six tracks at mixed skill levels. A biotech with deep technical staff and real budget could not make one course fit one company.
The same finding now has hard numbers behind it. BCG’s 2026 AI at Work survey of close to 12,000 frontline employees, managers, and leaders found that 74% of frontline employees describe themselves as AI users, up from 51% a year earlier, while only 36% feel they have received adequate upskilling. Usage went up. Capability did not follow it. The previous edition pinned down why that matters: 79% of people who received more than five hours of training were regular AI users, against 67% of those who got less, and 18% of regular users had received no training at all.
Staff are also asking for the thing most firms have not built. In McKinsey’s Superagency in the Workplace survey, “formal gen AI training from their organization” was the single highest-scoring answer, named by 48% of employees, when asked what would raise their daily use of these tools. The same study found leaders estimating that about 4% of staff use generative AI for at least a third of their work while employees reported a rate three times higher. The gap is not appetite. Most AI training for employees is bought as one company-wide session, pitched at a company, while the work happens in a function.
There is a legal reading too. Article 4 of the EU AI Act has been in force since 2 February 2025. It asks providers and deployers to take measures on AI literacy for their staff “taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in.” Role differences are written into the rule itself. Any US company with EU staff or EU customers inherits that wording, and it says what practitioners already know: literacy is relative to the job.
The broader case for treating this as capability work rather than a license rollout is made in the AI enablement manifesto, and the program-level view of moving a company from seats to measurable skill sits in the AI enablement playbook. This piece is about the layer underneath both: what nine functions each need to be taught.
The Floor Every Role Stands On
Before anything splits by function, everyone in the company needs the same four skills. This floor is short, around 75 to 90 minutes, and it is the only part of the program that should ever be delivered to a mixed room.
Knowing the shape of the tool. What a general-purpose assistant does well is language work: drafting, summarizing, reformatting, explaining, critiquing. What it does badly is arithmetic it was not shown how to do, recall of facts it was never given, and anything that needs a source it cannot see. People who hold that map in their head stop wasting afternoons on the wrong tasks.
Briefing instead of asking. The gap between a one-line question and a real brief is the gap between a generic paragraph and a usable draft. A brief carries context, a role, the output format, the constraints, and an example of good. This is a teachable craft, and it is the single skill that moves output quality most.
Verifying on purpose. Every claim, number, name, date, and citation is unverified until a person checks it against a source. The floor session needs one exercise where the tool produces something confident and wrong, and the room catches it. Reading about hallucination does not build the reflex; being fooled once does.
Knowing what may go in. One rule, written down, about which tier of which tool may receive which category of company data. Business and enterprise tiers of the major products state they do not train on your inputs by default; consumer tiers make no such promise. Most firms have a policy nobody can recite. The policy and training gap, and how to close it without a committee, is the subject of the safe AI use playbook.
Everything after this floor is role work. The design mechanics of the sessions themselves, how long each runs, how a room is staffed, and what exercises hold a group of twenty, are covered in the AI workshop design playbook.
The Nine Curricula
Nine tracks, each described by four things: the work AI touches in that function, what fluency means there, the drill that builds it, and the failure mode that shows up when the training was generic. The drills matter more than the content. In the rooms we run, the exercise where a team brings its own document and leaves with a working output is the part that people are still using a month later.
1. The Leadership Team
The work. Board narratives, strategy memos, market reading, long documents that arrive on Friday, the first draft of an all-hands message.
What fluency means. Executives are the one group whose main use of these tools is judgment about other people’s AI-assisted work. They approve budget for enablement, they read analysis a team produced with AI, and they set the disclosure norm for the company by example. Fluency here is hands-on enough to call a weak output weak.
The drill. Take last quarter’s board narrative, give the tool the underlying numbers, and ask it to build the strongest case against your conclusion. Executives who have watched a model argue the other side stop treating these systems as answer machines and start using them as a check.
The failure mode. Full delegation. A leader who has never written a brief cannot tell whether the enablement spend worked, which is how firms end up renewing seats on a usage dashboard that measures logins.
2. People Managers
The work. Reviewing work their teams produced with AI, one-on-one preparation, project updates, hiring loops, performance notes.
What fluency means. Managers set the review standard for their function: what ships unreviewed, what gets a second pair of eyes, what never leaves the building without a named human owner. They also model habits. A manager who pastes a team member’s self-assessment into a consumer account has taught the whole team that the data rule is optional.
The drill. Hand each manager three AI-assisted deliverables from their own function. Two are sound. One contains a fabricated figure and a confident, wrong claim. Ask them to approve or reject each, out loud, with reasons. The polish bias is visible within minutes and it never fully goes away once named.
The failure mode. Approving finished-looking work. Managers are the layer where an unchecked output becomes a company position, and they are the audience generic training serves worst, because the skill they need is review, not prompting.
3. Sales
The work. Account research, call preparation, follow-up email, proposal and RFP drafting, CRM notes, objection handling.
What fluency means. Grounding the tool in real account context rather than asking it to invent plausible context. A fluent rep assembles a pre-call brief from the CRM record, public filings, and recent news in five minutes, and knows that any fact about the prospect’s business that did not come from a source is a liability in the room.
The drill. Pick a live deal. Build the pre-call brief with the tool, then compare it against what the rep walked in with last time. Then write the follow-up from the actual call notes, and strike every sentence that could have been sent to any other account.
The failure mode. Generic outreach at scale. Faster production of the same undifferentiated email is a worse outcome than no AI at all, and buyers notice first.
4. Marketing and Communications
The work. Briefs, long-form drafts, repurposing one asset into five, campaign copy, search content, asset production.
What fluency means. Holding the brand voice while raising volume. That means a written voice brief used as a reusable input, a claims rule about what may be asserted without a source, and a disclosure standard for anything customer-facing. Marketing teams also need one person who can tell when AI-assisted content has flattened into the house style of the entire internet.
The drill. Take a page you published last quarter. Have the tool rewrite it against your voice brief, then have it critique its own rewrite against that brief and fix what it flagged. The second pass is where marketers learn that critique prompts outperform generation prompts.
The failure mode. Sameness, and unsourced claims entering the market faster than anyone can check them.
5. Finance and Accounting
The work. Variance commentary, reconciliation narratives, board pack prose, policy answers, collections email, documentation of models somebody built three years ago.
What fluency means. A hard line on data, and a hard line on arithmetic. Finance is the function where the tier rule has teeth: unmasked ledger detail, payroll, and anything under an audit does not go into a consumer account. And the tool drafts the commentary, never the number. Finance fluency is reaching for explanation and language work while keeping calculation where it belongs.
The drill. Take the month-end variance table. Draft the commentary with the tool, then verify every figure in that commentary back to the ledger, counting how many were wrong. The count is the lesson, and it is different for every team.
The failure mode. A transcription error inside a fluent paragraph, signed by a controller who read the prose and not the digits.
6. HR and People Operations
The work. Job descriptions, screening summaries, policy answers, interview guides, onboarding material, internal announcements.
What fluency means. Knowing exactly where the tool is allowed near a decision about a person, which is nowhere, and where it earns its keep, which is structure: better interview guides, clearer rubrics, faster policy drafting, consistent onboarding content. HR also holds the most sensitive data in the company, including medical and investigation material, so this track carries the strictest handling rules outside of legal.
The drill. Build an interview guide and a scoring rubric for a live opening. Then run two anonymized past interview summaries through the rubric and compare the result against the rating the human panel gave. The exercise teaches both the value of structure and the limit of inference.
The failure mode. Letting the tool rank candidates. It will do it willingly, it will sound reasonable, and it creates legal exposure that no productivity gain covers.
7. Legal and Contracts
The work. First-pass contract review, clause comparison, duty extraction, plain-language explanation of a redline, research triage.
What fluency means. Verification discipline at a level no other function needs. The public record here is countable: the AI Hallucination Cases database maintained by Damien Charlotin at HEC Paris had identified 2,202 court decisions as of 19 October 2026 in which a party relied on AI-generated content that turned out to be fabricated. A fluent in-house lawyer treats every citation as unverified and every clause reference as a lookup.
The drill. Take a signed agreement. Have the tool list every duty with the clause number it came from, then check each reference against the document. Teams discover that the summary is good and the pointers are sometimes invented, which is the pattern that puts cases in that database.
The failure mode. Invented authority surviving into something filed or sent.
8. Customer Support and Success
The work. Reply drafting, macro maintenance, ticket summarization, account health notes, renewal and review preparation, knowledge-base articles.
What fluency means. Grounding every answer in the actual help center and policy, not in the model’s sense of what a reasonable policy would be. Support fluency is also tonal: holding a brand voice while a customer is angry, and knowing which three categories of ticket must never be answered with a draft nobody read.
The drill. Pull three real tickets. Draft replies with the knowledge base supplied as context, then sort them into send, edit, and escalate. The sorting is the skill.
The failure mode. A confident, invented policy sent to a customer in writing. It is the fastest way an enablement program loses executive support.
9. Operations and Administration
The work. Standard operating procedures, vendor comparisons, data cleanup, scheduling, meeting notes, internal communication, the long tail of recurring tasks nobody owns.
What fluency means. Turning a repeated task into a reusable prompt plus a checklist, and recognizing the moment a workflow has outgrown a chat window. Operations is where fluency training most often surfaces a real automation candidate, because the person doing the task for the fortieth time can describe it precisely.
The drill. Choose one weekly recurring task. Write the reusable prompt, run it twice, and time both runs. Save the prompt where a colleague can find it.
The failure mode. Perpetual one-off prompting. Nothing is saved, nothing is shared, and the same work is reinvented every week by three people.
| Track | Core skill | Walk-out artifact |
|---|---|---|
| Leadership | Judging AI-assisted work | A counter-case to a decision already made |
| Managers | Review standards | A written rule for what ships unreviewed |
| Sales | Grounded account context | A pre-call brief for a live deal |
| Marketing | Voice and claims control | A voice brief plus one rewritten asset |
| Finance | Data lines and number discipline | Verified variance commentary |
| HR | Structure without inference | An interview guide and rubric |
| Legal | Verification discipline | A duty list with checked references |
| Support | Grounded answers | Three triaged draft replies |
| Operations | Reusable prompts | One saved prompt with a timed before and after |
The Tenth Track You Should Cut
Software engineers do not belong in a fluency program. Their AI work is a different discipline: coding agents, review practice, test coverage, secret handling, evaluation of generated code against a real test suite. It moves monthly, it is deep rather than broad, and it belongs inside the engineering function with its own standards and its own budget.
Putting engineers in a session about briefing and verification wastes their afternoon and distorts the program. They are the one group who will tell you the training was too basic, and that feedback gets used to pitch the next session too high for everyone else. Teach them separately or let them teach themselves. Nine tracks is the right number for a company-wide enablement program.
Collapsing Nine Curricula Into an Enablement Program
Nine curricula do not mean nine programs. A 200-person company cannot run nine training streams and should not try. Team AI training works when the people in the room share a document type and a risk profile, which is a looser condition than sharing a job title. Group nearby teams into rooms that share a risk profile and a document type, then let the differentiation happen inside the room through the material each team brings.
The structure that holds up at this size is six sessions:
- The floor, 75 to 90 minutes, everyone, repeated as often as room size demands. Cap it at 40 people if you want hands on keyboards.
- Managers, 90 minutes, before their teams go through the role rooms.
- Revenue room, two hours: sales, marketing, support and success. Shared ground is customer-facing language, voice, and claims.
- Numbers and records room, two hours: finance, accounting, operations, administration. Shared ground is data handling and verification against a source.
- People and policy room, two hours: HR, legal, compliance. Shared ground is sensitive material and decisions about people.
- Leadership, 90 minutes, which can run first or last but works best after the floor and before the role rooms, so executives know what their people were taught.
Six sessions, nine tracks, around eleven hours of delivery for the whole company. Note what that buys per person: the floor at 90 minutes plus a two-hour role room plus an hour of clinic is four and a half hours, which sits just under the five-hour mark BCG ties to regular use. That gap is the argument for the clinic and for a second role room in the heaviest functions, and it is the reason a single all-staff session changes nothing.
Add one clinic at week four: open room, bring whatever broke. Attendance will be lower and value per attendee higher, because the people who come have hit a real wall.
One caution on grouping. Put finance and legal in the same room only if the company is small enough that both functions are two people. The handling rules overlap, the drills do not.
Who Goes First
Managers go before the teams they supervise. This is the sequencing decision most companies get backwards, and the cost shows within a month. Individual contributors come out of a good session with new habits, submit work produced a new way, and run into a manager who has no standard for reviewing it. Some of that work gets waved through because it looks finished. Some gets rejected because it looks like a shortcut. Either way the training is spent.
After managers, sequence by volume of language work rather than by seniority or enthusiasm. Sales, marketing, and support handle the highest volume of drafting in most companies, so they convert training into visible output fastest, which funds the rest of the program politically. Finance and legal come next because their risk is the highest and their drills are the most specific. Operations often wants to be first and is better placed third, once there are examples from other teams to copy.
Resist the volunteer-led rollout. The people who raise their hands are already fluent, and a program built around them produces a thin layer of enthusiasts and no change in the teams where the work piles up.
What Role-Based AI Enablement Costs
Three honest cost lines, and only one of them is the invoice.
Delivery. Live hands-on enablement work for a company this size runs in the $2,000 to $15,000 range in the current market, driven by session count, how much of your own material goes into the build, and remote versus on site. A single floor session for a 60-person company sits at the bottom of that band. The six-session structure above, built on your material, sits in the middle to upper part of it.
Internal time. Eleven hours of delivery across 200 people is not eleven hours of cost. The floor at 90 minutes for everyone is 300 hours. The role rooms add 400. A one-hour clinic that half the company attends adds another 100. Call it 850 person-hours before anyone writes a check. At a blended internal rate this is the biggest number in the business case, and it is the number that argues against running a tenth track nobody needs.
The follow-on that is not training. Role rooms surface automation candidates, especially from operations and finance. Building one is a project, not a workshop, and custom automation work of that kind runs in the $25,000 to $150,000 range in the current market. Keep it out of the enablement budget and out of the enablement conversation until a specific workflow has earned it. Whether to build that capability inside or buy it is the subject of the buy-versus-build playbook.
Market ranges are a starting point, not a quote. Session count, headcount, and how much of your own material has to be prepared move the number more than anything else.
Where Catalog Pathways Stop
Catalog platforms have real role-based content and it is worth knowing what is there before paying anyone to build something similar. LinkedIn Learning publishes a named pathway library for exactly this. It carries AI for Managers and AI for Organizational Leaders, aptitude paths split for new, middle, and senior managers, and professional certificates for generative AI in sales, marketing, and customer service. There are essentials courses for marketers, sales professionals, project managers, and UX designers, a human resources track, and separate developer and cybersecurity paths. Coursera for Business offers Skill Tracks on its enterprise plan across software and product, IT, data teams, and generative AI, with per-seat pricing quoted rather than published.
So the honest comparison is not catalog versus role-based. Both are role-based. The difference is what can enter the room.
A pathway cannot open your contract, your variance table, your CRM record, or your ticket queue. It cannot tell your controller which of her twelve monthly reports is the one to start with, or watch your top rep write a follow-up and point at the three sentences that could go to any account. It teaches the general shape of the skill and leaves the application to the learner, which is why completion rates on self-paced AI content are a poor predictor of changed behavior.
The practical read: use catalog seats for breadth, reference, and the people who learn well alone, and use live sessions for the parts that only work with your documents on the screen. Firms that have already bought seats are usually not short of content. They are short of the two hours where someone applies it to the actual work.
Proof by Role and the Artifact Test
Course completion tells you nothing about capability, and usage dashboards tell you slightly less. Both measure access. The test that survives contact with a CFO is whether each function can produce something it could not produce before, and that artifact is different for every track.
Ask for it two weeks after the room, not on the day itself. Sales shows a pre-call brief for a live deal. Marketing shows one asset rewritten against a written voice brief. Finance shows commentary with every figure traced. HR shows an interview guide and its rubric. Legal shows a duty list with checked clause references. Support shows the triage of three real tickets. Operations shows a saved prompt and two timings. Managers show their written review rule. Leadership shows the counter-case.
Nine artifacts, no survey, and a clear answer to whether the money moved anything. The full version of this, including what to baseline before the first session and how to report it upward without inventing a productivity figure, is laid out in the AI enablement measurement playbook.
One number worth tracking alongside the artifacts: how many hours each person has logged. BCG’s finding that regular use rises sharply above five hours of training is the most useful planning input in the public data, because it tells you that a single session is below the threshold where behavior changes, no matter how good the session was.
Data Rules Change by Role, Not by Policy Document
Most companies write one AI policy, circulate it, and treat the distribution as the control. The policy is usually right and almost never operational, because the question a person has at the moment of pasting is specific to their function and the document in front of them.
Four rules, one per room, will outperform a nine-page policy:
- Revenue room. Customer names and deal detail may go into the business tier. Anything a customer gave you under an NDA stays out of consumer accounts, and no claim about a customer’s business leaves the building unsourced.
- Numbers room. Unmasked ledger detail, payroll, and audited material do not go into consumer accounts. Numbers are drafted around, never computed.
- People and policy room. Medical information, investigation notes, and compensation detail do not go in at all. The tool structures the process; it does not assess the person.
- Leadership. Unannounced material information, legal matters, and anything that would move a valuation stay out until the right people have seen it.
Each rule fits on a slide, is taught in the room it applies to, and is tied to a document the team handles that week. That is the version people follow. The policy document still needs to exist, and the gap between having one and having it obeyed is where most governance programs fail.
A 30-Day Build
For a company of 50 to 500 that has licenses, has run one session that did not stick, and needs this done without a dedicated owner.
Week 1. Map the nine to your org chart. Write down which of the nine tracks exist in your company, who owns each, and which teams are small enough to merge. Most firms find seven, not nine. Pick the four rooms. Name the manager cohort.
Week 2. Collect the material. One real document per track, cleared for use: a contract, a variance table, a ticket set, a live deal, a job opening, a recurring ops task. This is the week that decides whether the program lands, and it is the week most teams skip. Also write the four data rules and have whoever owns risk sign them.
Week 3. Run the floor and the manager session. Everyone gets the 90-minute floor. Managers get their room before anyone they supervise sees a role room. Set the artifact expectation out loud, in both sessions, with the date.
Week 4. Run the role rooms. Three role rooms at two hours each plus the leadership room at 90 minutes, all built on week-two material. Book the clinic for two weeks out and tell people to bring what broke.
Day 30 onward. Collect the nine artifacts. Publish them internally, which does more for adoption than any announcement. Then decide two things: which function needs a second room, and whether anything that surfaced in the numbers room deserves to become an automation project.
If the honest answer to any of this is that nobody internally has the time to run week two, that is the usual reason these programs stall, and it is worth a conversation before you buy anything. A free AI-readiness call walks your org chart against the nine tracks, finds which teams are furthest behind, and tells you which rooms to run first and what to bring into them. You will leave with the sequence and the material list, whether or not we deliver the sessions.
Frequently Asked Questions
How many AI training tracks does a company need?
Six delivered sessions covering nine role curricula, for a company of 50 to 500. The curricula differ by function, but the delivery collapses: one shared floor for everyone, one manager session, one leadership session, and three role rooms that group teams sharing a risk profile and a document type. Companies that try to run a separate program per department stall after the second one, and companies that run a single all-staff session teach nobody anything they can use.
What should every employee learn regardless of role?
Four things, in about 90 minutes: what these tools do well and badly, how to brief one instead of asking it a question, how to verify output against a source, and which company data may go into which tool tier. That floor is the only part of an enablement program that belongs in a mixed room. Everything past it depends on what the person works on.
Do managers need different AI training from their teams?
Managers need their own track, and they need it before their teams. A manager’s main AI task is reviewing work other people produced with AI, which calls for a review standard rather than prompting craft. When teams are trained before their managers, new work arrives in front of someone with no standard for judging it, and it either gets waved through because it looks polished or rejected because it looks like a shortcut.
Should engineers be included in a company-wide AI program?
Keep them out of it. Engineering AI work is a separate discipline: coding agents, review practice, test coverage, secret handling, evaluating generated code against a real suite. It moves faster than a fluency curriculum and runs deeper. Putting engineers in a session on briefing and verification wastes their time and skews the feedback used to pitch the next session, which then lands too high for everyone else.
How much does role-based AI training cost?
Live hands-on delivery for a company of this size runs in the $2,000 to $15,000 range in the current market, driven by session count, how much of your own material gets built into the sessions, and remote versus on site. The larger cost is internal time: roughly 850 person-hours for a 200-person company running the full structure. Automation projects that surface from the sessions are a separate budget, in the $25,000 to $150,000 range.
Can we just buy catalog courses instead?
For breadth and reference, yes, and the role pathways on platforms like LinkedIn Learning and Coursera are genuinely organized by job family. What a pathway cannot do is open your contract, your variance table, or your ticket queue, which is where the applied skill is built. Most firms that have already bought seats are not short of content; they are short of the two hours where someone applies it to their own work.
How do we know the training worked?
Ask each function for one artifact two weeks after the session: a pre-call brief, verified variance commentary, a duty list with checked clause references, three triaged ticket replies, a saved and timed ops prompt, a written review rule. Completion rates and usage dashboards measure access, not capability. Track one number alongside the artifacts, which is hours of training per person, because BCG found regular use rises sharply above five hours.
Which department should go first?
Managers, then whichever functions do the most drafting, which in most companies means sales, marketing, and support. High-volume language work converts training into visible output fastest, and visible output is what keeps the rest of the program funded. Finance and legal come next on risk grounds. Avoid building the rollout around volunteers; they are usually fluent already, so nothing changes in the functions where the work piles up.
Does the EU AI Act require role-based AI training?
Article 4 has applied since 2 February 2025 and asks providers and deployers of AI systems to take steps on AI literacy for their staff “taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in.” It does not mandate a specific curriculum or certificate, and it does not require guaranteeing any individual’s literacy level. The role-sensitive wording does mean a single uniform course is a weaker answer than differentiated tracks if the rule applies to you.
What do we do about people who are already fluent?
Give them the role room and skip the floor, then use them in the clinic. Internally fluent people are the cheapest source of role-specific examples you have, and asking one of them to bring a real artifact to their own function’s room does more for adoption than any announcement. What does not work is building the program around them, because a volunteer-led rollout produces a thin layer of enthusiasts and no change in the teams that need it most.
Arthur Wandzel
