Buy vs build AI training is the wrong first question, and asking it in that form is why so many of these decisions get re-litigated six months later. A training program is not one thing you either purchase or produce. It is four separable assets, and the right answer is usually different for each one.
The four assets inside any AI enablement program are curriculum, facilitation, measurement, and maintenance. Most companies buy a bundle of all four from a single vendor, or try to produce all four in-house with a person who has a day job. Both routes fail for the same reason: the parts have different economics.
Curriculum for general model skills is close to a commodity and the baseline price is zero. The judgment about which of your documents an assistant may touch cannot be bought at any price. Those two facts should pull your decision in opposite directions, asset by asset.
This guide splits the decision into those four parts and does the cost arithmetic in the open rather than borrowing unsourced vendor bands. It ends with a 90-day sequence that buys the first session and builds the second.
The short answer
Buy facilitation first. Buy baseline curriculum, because it costs tens of dollars a head. Build the proprietary half of the curriculum, the measurement, and the internal ownership that keeps the program alive after the vendor leaves.
That split holds for most companies between 50 and 500 people, because it matches where their scarcity sits. These companies are not short of content. They are short of a person who can stand in front of forty colleagues, open a real client file, and show what a model does well and where it quietly invents things. Content is abundant and facilitation is not, so buying content and producing facilitation inverts your own supply curve.
The rest of this guide is about the exceptions, the arithmetic, and how to run the sequence without buying the same thing twice.
The four assets inside an AI enablement program
Pull a training program apart and you get four assets with four different price curves.
Curriculum is what gets taught: the module order, the exercises, the worked examples, the prompts people take back to their desks. General curriculum on how large language models behave is widely available and cheap. Curriculum built on your own contracts, approval chains, and client sensitivities is not available anywhere, because nobody outside your company has it.
Facilitation is who stands in the room. A facilitator reads the group, notices the three people who have quietly stopped typing, and adjusts. This is the asset most companies under-price, and the one that decides whether a session changes behavior or fills a calendar slot. Our workshop design guide goes through the mechanics of a session that survives the walk back to the desk.
Measurement is how you know it worked. Seat licenses, course completions, and login counts are not measurement; they are receipts. Real measurement needs a before-and-after on one named workflow, which means someone inside your company has to define the workflow and hold the baseline. No vendor can do that part for you, which is exactly why so few offer to.
The measurement guide sets out the instruments.
Maintenance is who updates all of the above in six months when the products have changed under you. This is the asset nobody prices and the one that quietly decides whether you are doing this again from scratch next year.
Once you see four assets instead of one program, “should we buy or build” becomes four smaller questions, each with a clear answer.
Why the standard build-vs-buy table misleads
Search this question and you will find the same three-column table on seven different sites. Build costs more and takes longer but gives you customization and ownership. Buying is faster and cheaper but brings lock-in. Hybrid is probably right.
The columns are populated with suspiciously round cost bands that run from a few hundred thousand dollars a year for building to a mid-six-figure license for buying.
Trace those numbers. They do not resolve to a survey, a methodology, or a dated primary source. They appear on one vendor blog, get cited by a second, and arrive at the top of the results page as common knowledge. A widely repeated framework in this space attributes a staged buy-then-build sequence to analyst research that cannot be located at all.
Two problems follow from this, and both matter to you.
First, the bands are enterprise-shaped. The case studies in the ranking articles are global banks standing up internal academies for tens of thousands of staff. If you have 200 people, no instructional designer, and no learning function, the build column as written is not an option you can exercise. You are being shown a decision you do not have.
Second, those tables conflate two different questions. “Build vs buy AI” in most business writing means building custom models versus licensing software. “Buy vs build AI training” is a question about teaching people. They share four words and almost nothing else. If a page slides between the two, its cost bands are describing something you are not purchasing.
The useful version of this analysis costs out one thing at a time and shows its work.
What building actually costs
Here is the arithmetic, done in the open with a public number.
The US Bureau of Labor Statistics puts the median annual wage for training and development specialists at $69,280 as of May 2025, with training and development managers at $133,000. Load a specialist’s salary with payroll taxes and benefits at a conventional 1.25 to 1.4 multiplier and you land somewhere around $87,000 to $97,000 a year in true cost.
Now the part that decides the answer: you do not need a year of that person, you need a slice. Designing a credible applied curriculum, building the exercise set, dry-running it twice, and delivering it to four groups is a few months of concentrated work for someone who has never done it before. Call it a quarter of a specialist’s year, and you are at roughly $22,000 to $24,000 in loaded salary for version one. Add a manager’s oversight and the tooling and you are near $30,000 before anyone has learned anything.
That figure already clears the top of the market range for a bought workshop. A session runs about $2,000 to $15,000 depending on scope, group size, and how much of your own material gets worked into it. And the $30,000 buys you version one only, delivered by a first-time facilitator, with no refresh plan.
Two costs are missing from that count and both are larger than they look.
Participant hours. Two hundred people at three hours each is 600 hours. At a conservative $50 loaded hourly rate that is $30,000 of payroll spent sitting in the session, which usually dwarfs the invoice either way. Run this number before you optimize the vendor line, because it tells you that session quality matters far more than session price.
The failed first attempt. Internal first drafts of this curriculum tend to over-teach the technology and under-teach the application. The second version is usually the good one. Budget for two.
None of this says building is wrong. It says the build column is not the cheap column, and anyone who told you otherwise was comparing a salary slice against a license fee and leaving the hard parts out of both.
What buying actually costs
The buy column starts at about $49 a head, and no comparison table in this category mentions it.
DeepLearning.AI’s “AI For Everyone”, taught by Andrew Ng, runs four weeks at two to three hours a week across 35 video lessons, and the provider lists certificate eligibility at $49. Microsoft Learn publishes self-paced interactive modules and learning paths on the same ground.
Baseline literacy means your team can define a language model, knows why it fabricates citations, and can tell a summarization task from a reasoning task. If that is what you need, the need is met for the price of a lunch per person. Any paid literacy curriculum has to beat a $49 Andrew Ng course, and most of what gets sold as corporate AI literacy does not.
Above the free tier, the market sorts into two pricing shapes.
Per-seat catalog subscriptions. Coursera, Udemy Business, LinkedIn Learning, Pluralsight, and Skillsoft sell annual access per user. Coursera’s business plans, for example, split at a headcount line: the team plan covers 2 to 499 employees while 500-plus users move to an enterprise plan that is quoted rather than published. Published self-serve pricing below a threshold, a sales conversation above it: that is the standard shape across this tier. Treat any specific per-seat figure you read as a starting point for your own quote rather than a fact.
The per-seat model has one structural problem for enablement. You are buying access, and access renews whether or not it produced capability. Cost scales with headcount while benefit scales with completion, and completion on assigned asynchronous AI courses is where most of these programs quietly die. This is the strongest objection anyone raises against bought training, and it holds.
Per-session instructor-led delivery. Workshops, cohort courses, and bootcamps price per engagement. Here you are buying a facilitator’s time and attention, which is the scarce asset, and the cost does not renew by itself. The market range is roughly $2,000 to $15,000 for a session. A custom automation build that follows later is a different order of commitment, commonly $25,000 to $150,000, and should never be bundled into a training decision.
The honest summary of the buy column: baseline curriculum is close to free, catalog access is cheap per seat and expensive per outcome, and facilitation is the thing genuinely worth paying for.
The content half-life problem
Built AI curriculum decays faster than any other training content a company owns, and the decay is structural rather than a sign of sloppy work.
Three things move under it. The products change, so a screenshot-driven exercise breaks when a vendor reorganizes a menu. The behavior changes, so a prompt pattern that needed three careful steps last spring now works in one sentence and your exercise teaches a workaround for a problem that no longer exists. Your own policy changes, so the tool you told everyone to use for client material gets restricted and the module that teaches it is now wrong in a way that matters.
Write curriculum at the level that survives this. Teach the shape of the task rather than the sequence of clicks. Name products, never version numbers, because a module that says ChatGPT, Claude, Gemini, or Microsoft Copilot stays true for years while a module naming a specific model release is stale before the second cohort. Keep the volatile material, meaning screenshots, menu paths, and current limits, in a short appendix one person can refresh in an afternoon instead of scattered through forty slides.
Then put a name and a date on the refresh. “Reviewed quarterly by the enablement owner” is the line that separates a program from a deck. A bought curriculum should carry the vendor’s refresh commitment in writing; if the proposal is silent on refresh, you are buying a snapshot.
The dividing line: portable skill vs proprietary context
Here is the test that resolves most of these decisions in one pass.
Portable skills are commodities. Buy them. How a language model predicts text and why that makes it confident when wrong. How to structure a prompt with role, context, and a worked example. How to recognize a task that needs retrieval rather than recall. How to check a model’s output against a source. None of this is specific to your company and all of it is taught well somewhere already. Producing it in-house is paying a premium for the privilege of writing your own version of a $49 course.
Proprietary context cannot be bought at any price. Which client documents may go into which tool. Which approval still needs a partner’s signature even when the draft is good. Which three of your workflows are worth automating first and which look tempting but are too entangled with a system of record to touch yet. The vocabulary your industry uses where a general model will cheerfully use the wrong term. What your policy actually permits, which is a separate question from what the policy document says, and one the safe AI use guide treats at length.
Every hour of training sits on one side of that line. The split we aim for is roughly one part portable to three parts proprietary. If your program is 80% model mechanics and 20% applied, you have bought a course and called it enablement. That inversion is the most common failure in this category, and the reason the license-to-capability gap persists at companies that have already paid for everyone’s seats.
A vendor who will not run the session on your own documents can only ever deliver the portable quarter. That is a legitimate thing to purchase. It is not a program.
The asset-by-asset decision grid
| Asset | Buy when | Build when | What you must keep either way |
|---|---|---|---|
| Baseline curriculum | Almost always. A $49 self-paced course already covers general model literacy. | Never as a first move. Writing your own version of a $49 course is the clearest waste in this category. | The reading list, so a new hire self-serves the baseline in week one |
| Applied curriculum | For the structure, exercise design, and facilitation patterns | For the content: your documents, your workflows, your edge cases. Nobody else can supply these. | The exercise set, written against your real files |
| Facilitation | First and second sessions. This is the scarce asset and the one worth paying for. | From session three, once an internal owner has co-facilitated twice | A recording and the facilitator’s notes, including what fell flat |
| Measurement | Only the instrument design | The baseline and the follow-up. A vendor cannot define your workflow for you. | The before-and-after on one named workflow, with the raw numbers |
| Maintenance | As a contractual refresh commitment, in writing | As a named owner with a quarterly date | The volatile appendix, separated out so refresh is an afternoon |
Read down the “build” column and notice what it is made of. It is not production capacity. It is access to your own context and a named person who will keep it current. Those are the two things you cannot purchase, and they are also the two things a 200-person company can supply without hiring anyone.
The two ways hybrid programs fail
Everyone recommends hybrid. Almost nobody specifies it, and an unspecified hybrid fails in one of two predictable ways.
The seat-and-hope hybrid. A company buys a catalog subscription for everyone, assigns an AI learning path, adds one kickoff session so it feels like a program, and moves on. Completion starts low and keeps falling. Nobody owns the gap between the course library and anyone’s actual Tuesday.
At renewal the dashboard shows adequate activity and the capability has not moved, which is the exact pattern that sends a COO looking for a different vendor rather than a different design. The fault was not the catalog, which was cheap and fine. The fault was buying curriculum and measurement from the same vendor and buying no facilitation at all.
The orphaned internal build. One enthusiastic person builds a genuinely good internal deck, runs it twice to real applause, then gets promoted, reassigned, or simply busy. The deck goes stale in a drawer. Six months later a new executive asks what happened to AI training and the honest answer is that it was one person’s side project. The fault here was treating facilitation as a volunteer role rather than part of someone’s job with a quarterly date attached.
Both failures share a root cause: nothing in either design names an owner with authority and a calendar. Specify hybrid as “we buy facilitation and baseline curriculum, we build applied content and measurement, and this named person owns maintenance with a quarterly review” and both failure modes close.
Who owns this on Monday
At most companies in this size band there is no Chief AI Officer and no AI function. The work landed on a COO, a chief of staff, a head of operations, or a functional VP as a second job, which shapes the buy-vs-build answer more than any cost table.
A person with a full-time role elsewhere can do three things well. They can define the workflow that matters, because they already know it. They can hold the measurement baseline, because they own the numbers. They can co-facilitate after watching twice, because the content is their own work.
The same person cannot write a curriculum from scratch, design exercises, and facilitate cold for forty colleagues on top of their day job. Asking them to is how the orphaned internal build happens.
So build the owner, not the course. Buy the first two sessions and have your named owner co-facilitate the second. By session three they are running it with the vendor in the back of the room, and you now hold something that no purchase produces: an internal person who can teach this.
That is the capability transfer worth paying for, and it is a reasonable thing to write into a proposal as a deliverable. Curricula vary a great deal by role, and the role-based guide covers why one session for the whole company under-serves nine different jobs.
The vendor landscape, sorted by what they sell
“Best AI training companies” lists sort by brand. Sort by what the vendor actually sells instead, because that determines which of your four assets it can fill.
| What they sell | Representative providers | Fills | Does not fill |
|---|---|---|---|
| Self-paced foundational courses | DeepLearning.AI, Microsoft Learn | Baseline curriculum for tens of dollars a head | Facilitation, applied content, measurement |
| Per-seat catalog access | Coursera, Udemy Business, LinkedIn Learning, Pluralsight, Skillsoft | Breadth, self-serve baseline, a credential trail | Applied context, anything about your documents |
| Scheduled instructor-led classes | General Assembly, BrainStation, NobleProg, Certstaffix | Facilitation on a fixed syllabus | Your workflows, unless you pay for customization |
| Cohort courses with a named instructor | Maven | Facilitation and peer learning, one person at a time | Team-level measurement, company context |
| Membership plus software plus services | Section | Adoption tracking, coaching, leadership programs. Its published offering spans an adoption command center, coaching, and bootcamps | Published pricing, which is quote-only |
| Apprenticeships and funded degrees | Multiverse, Guild, Correlation One | A longer talent pipeline | Anything on a 90-day horizon |
| Applied hands-on workshops on your own work | Boutique practitioners, including us | Applied curriculum, facilitation, capability transfer | Breadth. A workshop is not a catalog. |
Three readings worth taking from that table.
Catalog vendors and workshop practitioners are not competitors. They fill different assets and the sane purchase is often one of each: the catalog as a self-serve baseline, the workshop for applied work on real files.
A vendor’s pricing model tells you what it believes it is selling. Per-seat annual means it is selling access. Per-session means it is selling delivery. Neither is wrong, but only one of them stops charging you when nobody shows up.
Verify every pricing and curriculum claim against the vendor’s own current page before it reaches a procurement document. This catalog changes quarterly, published tiers move, and the aggregator sites that rank for pricing queries are often a year stale. We checked each claim above against the provider’s live page at the time of writing and still recommend you re-check at quote time.
Eight procurement questions that separate vendors
Ask these in the first call. The answers sort the field faster than any proposal.
- Will you run the session on our documents, and what is your intake process? A vendor who works on your material has an intake step. A vendor who does not will offer to “tailor examples to your industry”, which means changing the names in a generic deck.
- Who facilitates, by name, and how many sessions have they run? Sales staff on the call and an unnamed contractor in the room is the most common bait-and-switch in this market.
- What do we keep afterward? The recording, the exercise set, the prompt library, the slides. If the answer is a certificate of completion, you bought attendance.
- What is your refresh commitment when the products change? Ask for it in months and in writing.
- What will you measure, on what instrument, before and after? A vendor who answers with engagement metrics is measuring itself, not you.
- How does our internal owner get to the point of running session three? Capability transfer should be a named deliverable, not a hope.
- Is this priced per seat or per session, and what happens at renewal? Then ask what happens if your headcount grows 30%.
- Will you decline to teach something our policy forbids? A vendor who has never pushed back on a client’s policy has not thought about yours.
Our field guide to evaluating an AI partner applies a similar filter to the broader services decision, and the two screens compose well.
The build case that holds up
Four conditions make building the right first move. They are narrower than the comparison tables suggest, and when they are present the case is strong.
You already have a working learning function. Not a person who is willing, but an instructional designer who ships, a delivery calendar that holds, and a measurement habit that predates AI. If that exists, adding an AI track is a marginal cost rather than a new capability, and buying facilitation is paying for something you already have.
Your context is the entire lesson. Some companies operate in regulatory or proprietary territory where the applied portion is 90% of the value and no external facilitator can reach it quickly enough to be worth the ramp. Defense, certain clinical settings, and heavily regulated finance land here more often than most sectors.
You are training continuously at scale. If you onboard a hundred people a year and every one needs this, per-seat and per-session costs both compound while a built program amortizes. Run the arithmetic at your actual intake volume rather than at a single cohort.
You intend to sell it. A few companies build enablement content as a product or a client-facing asset. That changes the calculation completely, because you are now building inventory rather than overhead.
Absent all four, building first is usually a more expensive route to a worse version one. The enablement playbook walks the full path from license rollout to measured capability and is the better starting point if you are not sure which side of this you are on.
A 90-day sequence
This sequence buys the scarce asset, builds the unpurchasable ones, and leaves you with an internal owner at the end of it.
Days 1 to 15. Pick one workflow and baseline it. One workflow, named, with a number attached: hours per week, cycle time, or error rate. Measured now, by you, before anyone trains anyone. Skipping this step is the reason most programs cannot answer whether they worked.
Days 16 to 30. Buy the baseline cheaply and use it as a filter. Assign the foundational course, which lists at $49 a seat. Watch who finishes. The people who work through four weeks of self-directed material unprompted are your co-facilitator shortlist, and you identified them for the price of a few lunches.
Days 31 to 45. Buy session one, on your own documents. One group, real files, a named facilitator. Your intake package goes out a week ahead. Market range for this is $2,000 to $15,000 and the participant-hours cost is probably larger, which is the argument for a good facilitator rather than a cheap one.
Days 46 to 60. Name the owner and have them co-facilitate session two. Same material, second group, your person running two of the modules with the vendor in support. This is the transfer step and it is the single highest-return hour in the whole sequence.
Days 61 to 75. Your owner runs session three alone. Vendor observes and debriefs, or is gone entirely. You now hold facilitation capability in-house, which no purchase could have given you on day one.
Days 76 to 90. Re-measure the same workflow and decide what to renew. Compare against the day-15 baseline on the same instrument. Then make three decisions with evidence: renew the catalog or drop it, bring the vendor back for the next role group or not, and put a quarterly refresh date in your owner’s calendar.
Ninety days, one bought session plus one co-facilitated, and an internal owner who can run the next six. That is the shape of a program rather than an event.
How to price the AI enablement decision
Build a total-cost line for both routes over 24 months, not 12. The 12-month view flatters buying and hides the maintenance liability on building.
Include these lines on both sides:
- Direct cost. Vendor fees or loaded internal salary slices, calculated from a real wage figure rather than a borrowed band.
- Participant hours. Headcount times hours times loaded hourly rate. Usually the largest single line and usually omitted.
- Design and rework. Two versions, not one, if you are building.
- Refresh. Quarterly, for 24 months. Eight refresh cycles is the number people forget.
- Measurement. Someone’s time to hold the baseline and run the follow-up.
- Renewal exposure. For per-seat deals, the cost at your projected headcount in year two, not today’s.
Then compare per-outcome rather than per-head. The right denominator is not people trained, it is the measured change on the workflow you baselined. A $12,000 session that moves one workflow by four hours a week per person beats a $20,000 catalog renewal that moves nothing, and only the per-outcome denominator shows that.
If the two routes come out close, buy facilitation and build ownership. Close arithmetic should break toward the option that leaves capability inside the company.
FAQ
Is it cheaper to buy or build AI training?
Buying is almost always cheaper for the first year at companies under 500 people. Using the BLS May 2025 median of $69,280 as the base, a quarter of a training specialist’s loaded time runs roughly $22,000 to $24,000. That is already above the $2,000 to $15,000 market range for a bought workshop. It also buys only version one, from a first-time facilitator, with no refresh plan. Building becomes competitive when you are training continuously at volume, because a built program amortizes while per-seat and per-session costs compound.
What should we never build in-house?
Baseline model literacy. Self-paced courses, including DeepLearning.AI’s “AI For Everyone” at a listed $49 for certificate eligibility, already cover how language models behave and why they fabricate. Writing your own version of that content is the clearest waste in this category. Spend your internal effort on the applied half instead, which is the part nobody can sell you.
How do we evaluate AI training vendors without a procurement team?
Ask eight questions on the first call. Will you work on our documents, and how? Who facilitates, by name? What do we keep afterward? What is your refresh commitment, in writing? What will you measure before and after? How does our internal owner reach the point of running session three? Is pricing per seat or per session, and what happens at renewal? Will you decline to teach something our policy forbids? Any vendor that answers all eight concretely is worth a proposal.
Will a catalog subscription teach our team to use AI on our actual work?
A catalog subscription is not designed to do that, and will not. Catalog platforms sell access to general content, so they fill your baseline curriculum well and cannot touch your applied context. Pair one with facilitated sessions on your own files rather than treating it as the whole program. Assignment does not produce completion, and completion is not capability even when it happens.
Who should own AI enablement if we do not have a Head of L&D?
Whoever owns the workflow you are trying to improve, usually a COO, head of operations, or functional VP, with a named quarterly review date. They can define the workflow, hold the measurement baseline, and co-facilitate after watching twice. They cannot write curriculum and facilitate cold on top of a full-time role, which is what to buy instead.
How long does built curriculum stay current?
Plan on a quarterly review and expect the volatile parts to break sooner. Products reorganize their interfaces, model behavior shifts so that yesterday’s three-step workaround becomes one sentence, and your own policy changes which tools may touch client material. Keep screenshots, menu paths, and current limits in a short appendix one person can refresh in an afternoon, and teach the shape of the task in the body so the body survives.
Is hybrid actually the right answer?
Hybrid is the right answer only when it is specified asset by asset. Unspecified hybrid fails as seat-and-hope, where a catalog gets assigned and nobody owns the gap to anyone’s real work, or as an orphaned internal build, where one enthusiast’s deck goes stale after they get promoted. Specify it asset by asset: buy facilitation and baseline curriculum, build applied content and measurement, name one person for maintenance with a date.
What does a corporate AI training program cost in total?
Add six lines over 24 months on both routes: direct cost, participant hours, design and rework, eight refresh cycles, measurement time, and renewal exposure at projected headcount. Participant hours are usually the largest line and the most commonly omitted. Two hundred people at three hours each is 600 hours, which at a conservative $50 loaded hourly rate is $30,000 of payroll in the room, often more than the vendor invoice either way.
How do we know the decision was right?
Re-measure the workflow you baselined before training, on the same instrument, 60 to 90 days afterward. Completion rates, login counts, and satisfaction scores answer a different question. One named workflow with a before number and an after number is enough to defend a renewal or kill it, and it is the only evidence that survives a change of executive sponsor.
Key takeaways
- Decompose before deciding. Curriculum, facilitation, measurement, and maintenance have different economics, and bundling them is why these decisions get re-litigated.
- Buy portable skills, build proprietary context. General model literacy is a commodity priced in tens of dollars a head. Which of your documents a tool may touch cannot be purchased.
- Facilitation is the scarce asset. Content is abundant and a person who can run the room is not, so buy the part you are short of.
- Ignore unsourced cost tables. Do the arithmetic from a public wage figure, include participant hours, and run it over 24 months.
- Build the owner, not the course. Buy sessions one and two, co-facilitate the second, run the third internally.
- Specify your hybrid by asset and name a maintenance owner with a quarterly date, or it will fail in one of two predictable ways.
- Measure one named workflow before and after. Everything else is a receipt.
If you want a second opinion on which assets to buy and which to build in your situation, we run a free AI-readiness call. Bring the one workflow you would most like to see move, and we will tell you plainly whether a workshop is the right next step or whether you should spend the quarter on something else.
Arthur Wandzel
