A custom AI curriculum and an off-the-shelf corporate AI course are not competing on price or quality. They compete on information gain: how much either one tells your team something it cannot already get on its own, for nothing. Score the two that way and most paid generic AI courses are worth close to zero, and so is a good share of what gets sold as custom.
The test takes four minutes. Take the syllabus of any course you are weighing, paste it into the assistant your company already pays for, and ask the model to teach you module one. If what comes back is good enough, you have measured that course’s information gain and the answer is nothing. The model read the same public material the course was written from.
What information gain means when the material is a curriculum
The phrase has a precise home. Google holds a granted US patent, 11,354,342 B2, “Contextual estimation of link information gain,” filed by Victor Carbune and Pedro Gonnet in October 2018 and granted in June 2022. It scores a document by “additional information that is included in the document beyond information contained in documents that were previously viewed by the user,” worked out per user and per session.
That last part is the whole idea. Information gain is not a property of the thing you bought. It is the gap between that thing and what the reader already holds, so the same page scores high for one person and zero for the next.
A curriculum behaves the same way. A ninety-minute module on writing better prompts has real gain for a room that has never opened ChatGPT, and none for a sales team that has used it daily since the licenses landed. Nothing about the module changed. The room did.
So the question for a vendor is not “is this good training,” because most of it is competently made. It is what this room lacks, and whether the proposal contains any of it.
The generic module is already priced at zero
The market has settled this for you, in public.
| Off-the-shelf option | Published price | What it covers |
|---|---|---|
| Google AI Essentials on Coursera | $49 a month after a 7-day trial, financial aid available | Five courses, under 10 hours total, 2,044,479 enrollments |
| Coursera for Teams | $399 per user per year, 2 to 499 licenses | 12,000-plus courses, 140-plus certificates, 1,500-plus labs |
| Pluralsight Complete | $29 a month, or $299 billed yearly | Full library across tech, AI, cloud, data, security |
| Section AI Premium | $62.50 a month billed annually | Live and on-demand courses, coaching, monthly office hours |
Those prices were on the providers’ own pages on October 22, 2026. Google’s introduction to working with generative AI, one of the most widely taken items in the category, costs less than a team lunch and has passed two million enrollments. The AI literacy layer is a commodity sold at commodity rates, which is the right price for it.
Note what is missing from the table: every provider on it publishes an individual price and routes team pricing to a sales call. You can benchmark the generic layer by reading a web page and cannot benchmark the tailored layer at all, which is much of why seats feel like the safe default.
Run the syllabus through the assistant
Four steps.
- Copy the course outline out of the proposal or the vendor’s page.
- Paste it into ChatGPT, Claude, Gemini or Microsoft Copilot, whichever your team licenses.
- Ask the model to deliver module one as if it were teaching your team, then ask for the exercises.
- Read what comes back as if a vendor had sent it.
For the literacy half of most curricula, the output is roughly what you would have paid for. No knock on the model or the course: both draw on the same public corpus, and one of them is already on your invoice.
Now change one input and run it again. Ask the model to build that module around a real artifact: last month’s underwriting memo, a half-finished statement of work, the intake form your ops team retypes forty times a week. The answer gets thinner and more careful, because the model has never seen those documents and does not know your review path, your tone rules, or which of the three numbers in that memo the client will argue about.
That thinning is your information gain, and the only part of a curriculum worth paying a facilitator for.
The part a catalog course cannot contain
A course is written before it knows who buys it, so its contents are the intersection of every buyer, which is the material already on the open web. Five things sit outside it.
Your documents. The lease abstract, the clinical note, the board deck, the pricing sheet. People do not learn to work with AI in general. They learn on one artifact and then recognize its shape elsewhere.
Your bad outputs. The draft that read well and was wrong, and why. A course teaches hallucination as a concept. Your own worst output teaches it as a reflex.
Your review path. Who checks AI-assisted work before a client sees it, and what they look for.
Your constraints. The data that may not be pasted anywhere, the client who banned the practice in writing, the regulator with a view. These decide whether anyone still uses the tool by week three.
Your objections. The two people who think this is a layoff in a nicer sweater. A course cannot answer them because it cannot hear them.
In sessions we run, the first real unlock almost always comes from a document somebody brought under protest. It is also where stock material visibly stalls, usually inside the first half hour, and our guide to designing a session that changes Monday morning works through why.
Custom with no gain is the costlier mistake
The argument cuts both ways, and the second cut is the expensive one.
Plenty of AI enablement work sold as custom is a standard deck with your logo on slide one and your sector named in the case study. Zero information gain, at several times the price of a seat. A catalog course that teaches nothing new costs $399. A bespoke day that teaches nothing new costs $2,000 to $15,000 at market rates, plus the payroll of everyone sitting in it.
Four questions separate the real thing from the reskin, all askable before you sign.
- Which of our documents will you read, and when do you need them? A vendor doing the work names a deadline before the session date. A vendor that is not will ask for logo files.
- What will each person produce by the end, using their own work as input?
- Who builds the exercises, and have they done our kind of work?
- What changes in the material if we send you a different set of documents?
A vendor who cannot answer the last one is quoting a fixed curriculum, which is a legitimate product at a catalog price. The gap shows fastest in a leadership room, as our guide to planning executive sessions that are not condescending covers.
Cost per person who changed what they do
Price both options in the same units, for a team of 20.
Twenty Coursera for Teams licenses at $399 is $7,980 a year, and every person gets 12,000 courses. Completion is where it goes. The hardest data on self-paced learning comes from Reich and Ruipérez-Valiente’s 2019 study in Science, which covered 565 course runs and 12.67 million registrations on edX. Certification among registrants fell to 3.13% in 2017-18, down from nearly 6% in 2014-15, and did not improve across six years.
Corporate libraries with managers chasing dashboards beat open enrollment, but not ten times over. Even at a generous 25% finish rate, $7,980 works out near $1,600 per completion, and completion is not capability.
A tailored day for the same 20 people sits near $7,500, with payroll beside it running larger than the invoice, as our look at what a one-day session costs end to end breaks down. Attendance runs far higher because the date is on the calendar, and the output is checkable: people leave holding something made from their own material.
BCG’s 2025 AI at Work survey, 10,600-plus leaders, managers and frontline staff across 11 countries, puts a floor under this. About a third of employees say they have been properly trained, frontline regular use has stalled at 51%, and regular use climbs sharply past five hours of training. Five hours is a half day of focused work, not a hundred days of the fifteen-minute module nobody finishes.
Both routes fit a normal budget: ATD’s 2026 State of the Industry report put direct learning spend at $846 per employee in 2025 across 340 organizations, and $7,500 for 20 people is $375 a head. Only one has an output you can inspect.
When the catalog course is the right buy
A subscription wins outright in four situations. Scale past a few hundred people, where facilitated sessions become a sequencing problem rather than a budget one. A documented completion record, when a client or regulator wants evidence rather than capability. Continuous churn, where new hires need the same baseline monthly. And reference after the fact, since nobody rewatches a workshop but people do reopen a module.
What a subscription will not do is move a team that already holds licenses and is not using them. That gap is the subject of the whole case against treating license rollout as enablement, and catalog access does not close it.
Buy two layers, fund one
Treat AI enablement as two layers with different economics and the decision stops being a coin flip.
The literacy layer is a commodity. Buy it at commodity prices, or send people to Google AI Essentials on a monthly subscription. Make it pre-work with a date on it, not an event.
The application layer is where the information gain sits, and the only layer worth a facilitator, a design budget and a room full of salaried attention. Build hands-on work from documents your team already argues about, and hold it to an output you can read afterward.
Teams that do this well spend almost nothing on layer one and put the workshop budget into layer two, which also shortens the session, because nobody is paying a facilitator to explain what a large language model is. Where it happens matters less than what it is built from, and our comparison of on-site and remote delivery weighs that trade.
If two quotes are on your desk, run the syllabus test on both before you read the prices. A free AI-readiness call is a faster way to find which layer your team is short of.
Frequently asked questions
Is custom AI training worth the extra cost over an online course?
Only when the custom version contains your material. A tailored day runs $2,000 to $15,000 against $299 to $399 a seat a year, so for a room of twenty the two are close and for a room of five the day costs several times the seats. That premium buys one thing: content your team cannot obtain free. A proposal that names no documents it will read beforehand charges it for the same material.
How do I tell whether a vendor’s curriculum is genuinely custom?
Ask what changes in the material if you send a different set of documents. A vendor building from your work describes a different session. A vendor reskinning a deck talks about adding your branding and naming your industry in the examples.
Can we just use ChatGPT or Claude to train the team ourselves?
For the generic half, yes, which is the point of the syllabus test. Self-service breaks down on the part that needs someone to read your own output and say why it is wrong. Running that in-house takes 20 to 40 hours of prep from a person who has a full job already.
What does a custom AI curriculum cost?
Market rates run about $2,000 to $15,000 depending on the size of the room, how much design sits in front of it, and whether anything happens afterward. Custom automation projects, a different purchase, run $25,000 to $150,000. Published catalog seats sit between $299 and $399 a person a year.
How many hours of training does a team need?
BCG’s 2025 survey found regular use climbing sharply past five hours, which is a half day of focused work. Spreading those five hours over twenty self-paced modules is not equivalent, because most people never reach module four.
We bought licenses nine months ago and nothing changed. Where do we start?
With the gap, not the curriculum. Find three weekly tasks that are mostly text, then check whether anybody does them with AI. If the licenses are idle, a catalog subscription adds a second idle purchase.
Does certification from a course prove anything?
It proves attendance and a passed quiz, which is sometimes exactly what a client or regulator is asking for, and a fair reason to buy one. It is not evidence of capability, and a certificate with a 3% completion rate behind it evidences little.
What should we send a vendor before a tailored session?
Five to ten real documents your team works on weekly, redacted if needed, plus one output that went wrong and why. Vendors who can work from that will say so. Vendors who cannot will ask for a headcount and a date.
Key takeaways
- Judge any AI curriculum by information gain: what it tells this room that this room cannot already get free. Google’s patent on the idea scores a document against what the reader has already seen, and curricula work the same way.
- The literacy layer is a commodity. Google AI Essentials runs under 10 hours on a $49 monthly subscription; Coursera for Teams is $399 a user a year.
- Run the syllabus test before the price comparison. If the assistant your team licenses teaches module one adequately, the course has no gain for you.
- Only your documents, your failed outputs, your review path, your constraints and your skeptics sit outside the public corpus, which makes them the material worth a facilitator.
- Custom with no gain is the worse buy: a reskinned deck costs several times a seat and teaches the same nothing. Subscriptions still win on scale, completion records, churn and reference value.
Pull the syllabus from the quote on your desk and run it through your own assistant this afternoon. What survives is what you are buying, and a free AI-readiness call will tell you which layer your team is short of.
Arthur Wandzel
