Back to Guides
Enterprise Software 15 min read

The case for building an internal AI academy, and when not to

The case for building an internal AI academy, and when not to

Build an internal AI academy when you have a named owner with protected hours, more than roughly 150 people to train, and the capacity to rewrite the material every quarter for years. Below that line, building one converts a purchase into a staffing problem. For most companies between 50 and 500 people the better answer is to buy the facilitation and build exactly three things: the policy, the examples drawn from your own documents, and the record of who was trained on what. Those three are the parts nobody outside your company can write, and they are what makes bought training hold.

Nearly every published guide to internal AI training programs is written by a party with a financial interest in the answer: platform vendors argue for seats, agencies for delivery, trade press for internal capability. The result is a literature in which “hybrid” is the universal recommendation and nothing ever disqualifies a build. The wider decision is worked through in our guide to buying versus building AI enablement capability, and the reason a license count stopped standing in for readiness is argued in the case against buying seats and calling it capability. This page is narrower: the moment someone proposes building it yourselves.

What an Internal AI Academy Actually Is

Strip away the branding and an internal academy is four assets plus a person who maintains them.

  1. A tiered curriculum. Lloyds Banking Group’s academy, launched in January 2026 to put all 67,000 of its staff through AI training inside the year, sorts roles into users, leaders, builders, and enablers, with a mandatory responsible-use module first. One course for everybody teaches the wrong depth to three quarters of the room.
  2. Material built on your own work. Your contracts, your reports, your intake process — what makes training transferable and what no outside catalog can supply.
  3. A delivery mechanism. Booz Allen Hamilton’s “AI Ready” program gates access: all 33,000 employees must complete internal training and certification before using generative AI on the job, which is only enforceable because the training is internal.
  4. A record. Who was trained, on what, when, and what they could do afterward.

The person is the whole thing. Remove the owner and you have four documents aging at the speed of the tool releases they describe.

Why the Published Playbooks Do Not Transfer

Every named example in the public literature is an enterprise. Accenture’s FY2025 annual report counts over 550,000 of its people equipped with generative AI fundamentals, cumulative over years. Citi runs several thousand internal AI champions across 182,000 employees. Lloyds has an academy team; Booz Allen has a talent development director.

These are not scaled-down versions of what a 220-person company would build. They are a different category of thing — an internal capability with its own staff and budget line, and in Accenture’s case its own commercial logic, since the firm sells what it trains. Reading them as templates is how a mid-market company talks itself into a program it cannot maintain.

The one genuinely mid-market mechanism on record is also the least cited. SHRM’s reporting on closing the AI skills gap describes how Mineral, a compliance guidance provider serving smaller businesses, trained its staff: peer pods of about four people, a red-yellow-green scale for appropriate use, a six-step prompting framework, and an explicit norm of failing fast. No platform, no curriculum team. That is what internal capability looks like at a reachable scale, and it is closer to a reading group than an academy. The same piece notes Boston Consulting Group’s 2023 finding, across 12,800 employees in 18 countries, that 14% of front-line workers had received AI upskilling against 44% of leaders — a real problem, and not by itself an argument for building anything.

The Case For Building

Three arguments hold up, and they are narrower than the literature suggests.

Enforcement. Gated access only works when you control the gate. If nobody may touch a client document with an assistant until they have passed something, that something has to be yours. No external vendor can certify your people against your policy.

Specificity that compounds. A session built on your own documents teaches more than a generic one, and run a dozen times a year the material improves each time. Below a certain volume it does not compound — it goes stale between runs.

Institutional memory. The prompt library, the list of what went wrong, the examples of good and bad output. This accumulates only if someone owns it, and it is both the most valuable artifact in the program and the cheapest to build.

Note what is absent from that list: cost. Building arrives as salary already on the payroll rather than as an invoice, which makes it invisible rather than absent.

The Arithmetic Nobody Runs

Training magazine’s 2025 Training Industry Report puts US training expenditure at $102.8 billion, up 4.9%, with average spend of $874 per learner — $782 at midsize firms and $1,091 at small ones. Employees received an average of 40 training hours across the year, down from 47.

Read that against a 200-person company. At the midsize rate your whole annual training capacity across every topic is roughly $156,000, and each employee has about 40 hours to split between onboarding, compliance, management, tools, and now AI. The question is not whether an academy is affordable in isolation, but what fraction of those two budgets it consumes before anyone has learned anything.

Cost element Build internally Buy facilitation
Curriculum design 80–150 hours, repeated each refresh In the fee
Delivery Facilitator’s hours plus the room’s $2,000–15,000 per session
Material from your documents Yours, and you keep it Negotiable — fix it in the contract
Quarterly refresh Recurring, indefinitely Vendor’s problem until you ask
Certification and gating Possible Not possible
Year-one total Looks low; it is payroll Looks high; it is an invoice

One facilitated session at the low end of that range runs to about two weeks of a mid-level manager’s time — and the manager still has their actual job.

Year Two Is Where These Die

The literature on corporate AI training costs the build and ignores the maintenance, while insisting on a quarterly refresh cadence. That contradiction is where internal academies quietly fail.

AI tools change faster than any other subject in a corporate curriculum. A module recorded in March on how to structure a long document for an assistant is partly wrong by September — not because the principle changed but because the interface did. Almost no guide that recommends building prices four rewrites a year against the one person doing them as a second job.

The failure is undramatic. Nobody cancels the academy. The owner gets pulled onto a product launch, the Q3 refresh slips, new hires watch the March material, and eighteen months later the program exists on the intranet and in nobody’s behavior. Price the maintenance first and the build second.

Four Conditions That Disqualify a Build

Any one of these makes an internal academy the wrong call. Two together make it an expensive way to do nothing.

No named owner with protected hours. Not “L&D will pick it up.” A person, a share of their week in writing, and a manager who agreed to it. Many companies in this band have no standing learning function, and the work lands on a chief of staff or an operations lead as a second job. That arrangement can run a bought program; it rarely sustains a built one.

Fewer than roughly 150 people to train. Below that the design cost never amortizes. A hundred hours of curriculum work delivered four times is worse on nearly every dimension than buying four sessions — quality included, because the external facilitator has run fifty of them.

No capacity for quarterly rewrites. If you cannot name who does the Q2 refresh and when, you are not building an academy. You are building a snapshot.

A tool stack still in motion. If the choice between assistants is unresolved, internal material written now documents a configuration that will not exist. Buy a session on fundamentals that transfer across tools, settle the stack, then revisit.

The Split That Works Between 50 and 500 People

The standard recommendation is “hybrid,” and it stops there. The specific division of labor that holds up:

Build internally, regardless of who trains your people:

  • The acceptable-use policy. What goes into an assistant, what does not, what needs a second pair of eyes, who to ask. Nobody outside can write this, and it should exist before any tool goes wide.
  • The examples. Ten to twenty real artifacts from your own work — a good output and a bad one for each main document type. A few hours of effort, and the highest-return asset in the program.
  • The record. Name, session, date, what they can now do.

Buy, until volume justifies otherwise:

  • Facilitation. A practitioner who runs sessions weekly is better at it than your best internal presenter, and the $2,000–15,000 market range per session makes the comparison straightforward.
  • Curriculum structure. The tiering, the sequence, the exercises — bought, then loaded with your examples.

Revisit at volume. Past roughly a dozen sessions a year — sustained new-hire flow, several departments, a second site — the arithmetic turns and an internal facilitator starts to pay for themselves. Whether you start with one department or go company-wide largely determines how fast you get there, and if you are weighing who delivers the bought half, the choice between a freelance trainer and an enablement firm turns on continuity rather than price.

If you intend to build eventually, say so during procurement and negotiate for the material. A vendor who will not let your examples and the derived exercises stay with you is selling a dependency, so put material ownership on the evaluation checklist. It is cheaper to fix before the first session than after the fourth.

The Record You Keep Either Way

Article 4 of the EU AI Act states that providers and deployers of AI systems “shall take measures to support the development of AI literacy” among staff and others operating those systems on their behalf. The obligation has applied since 2 February 2025 and explicitly does not require you to guarantee any particular level of literacy. It requires measures, which in practice means evidence.

If you sell into the European market, or your software is used there, that reaches you whether or not you have an entity there, and the scope extends to contractors operating systems on your behalf. Article 4 is not on the penalty list in Article 99, so the exposure is indirect: missing training is not the fine, but it is what turns a bad output into a larger finding. Keep the list. It costs almost nothing, no vendor can hand it to you, and that is why it stays in your column even when everything else is bought.

Frequently Asked Questions

How many people do we need before building an internal AI academy makes sense?

Roughly 150, with sustained new-hire flow behind them. Below that the design cost never amortizes: 100-plus hours of curriculum work for a handful of runs, delivered by someone with less practice than any external facilitator. The threshold is delivery volume, not headcount.

What does an internal AI training program cost compared with buying workshops?

Buying costs $2,000 to $15,000 per facilitated session at market rates. Building costs 80 to 150 hours of design time, roughly the same again in refreshes across the year, plus delivery hours. The build only looks cheaper because it is payroll rather than an invoice.

Should we build AI training in-house or buy it?

Buy the facilitation and curriculum structure; build the policy, the examples from your own documents, and the training record. That split holds for most companies between 50 and 500 people. Build delivery too only with a named owner, more than a dozen sessions a year, and quarterly rewrite capacity.

How long does it take to build an internal AI academy?

Three to six months from decision to first session, done properly: needs assessment, curriculum, pilot, revision, sign-off. Buying a session takes two to six weeks. If the budget cycle closes before the six-month mark, the decision is already made for you.

Does the EU AI Act require us to train our employees on AI?

If your AI systems are used in the European market, Article 4 requires measures to support AI literacy among staff and others operating those systems for you. It has applied since 2 February 2025 and sets no specific level. Keeping a record of who was trained on what, and when, is the practical step — internal or bought.

Who should own the program when we have no L&D function?

Someone with operational authority and a protected share of their week: a chief of staff, a VP of operations, or whoever already owns tooling decisions. The person matters less than the protected hours.

Can we start internally without a budget?

Yes, as a first step. In a Gartner peer-community thread on internal academies, a practitioner describes starting with free self-paced material plus a weekly discussion group and says no large up-front investment was needed. That costs only meeting time and tells you quickly whether appetite exists. What it will not do is produce consistent capability across a function.

What should we build internally even if we buy the training?

Three things: the acceptable-use policy, because nobody outside can write your rules; ten to twenty real examples of good and bad output from your own document types, because that is what makes any session transferable; and the training record. A few days of work that decide whether bought training holds.

How often does an internal curriculum need rewriting?

Quarterly — the number every serious guide gives and almost none of them prices. Interfaces move faster than the underlying principles, so material recorded in one quarter is partly wrong two quarters later. If you cannot name who does the next refresh and when, build a shorter program or buy instead.

How do we know whether the program worked?

Measure behavior, not completion. Pick a task the team already repeats, time it before and after, and ask each participant what they can now do that they could not before. Completion rates and satisfaction scores measure attendance and mood, and an unmaintained academy reports both perfectly.

Key Takeaways

  • An internal AI academy is a standing commitment, not a project. The build cost is visible; the quarterly maintenance cost decides whether it survives year two.
  • Every named case study — Accenture, Lloyds, Booz Allen, Citi — is an organization with a dedicated academy team. None is a template for a 200-person company.
  • Four conditions disqualify a build: no owner with protected hours, fewer than roughly 150 people to train, no capacity for quarterly rewrites, and an unsettled tool stack.
  • Buy the facilitation and the curriculum structure. Build the policy, the examples from your own documents, and the training record.
  • Building is not the cheap option. It is payroll instead of an invoice, which makes it invisible rather than absent.

Where to Start

Write two numbers down before deciding anything. Who owns this, and what share of their week is protected for it in writing. Then how many people you will train in the next twelve months, and how many separate sessions that implies. Many companies find they have answered the question by the time they finish the first number.

Then do the part that is yours regardless: draft the acceptable-use policy, collect ten real examples of good and bad output from your own documents, and open the spreadsheet that becomes the training record. If the answer is buy, you have made the bought training better. If it is build, you have the three assets that were never purchasable. Either way, testing one supplier with a paid pilot session teaches you what good delivery looks like more cheaply than your own first attempt will.

If you want a second opinion on those two numbers before the decision reaches a budget meeting, book a free AI-readiness call. We run these sessions for a living, which means we will tell you when your headcount and refresh capacity point toward building it yourselves.

Last Updated: Nov 17, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles