In-house AI training is a programme your own staff design, deliver, maintain and record. External AI training is the same programme bought from a provider who does those four jobs for you.
Buy when you need people trained this quarter, when nobody internal has the hours to keep pace with the tool releases, or when you need an audit trail. Build when you already have a functioning L&D team, a technical champion with capacity, and workflows too specific for an outsider to teach without a long scoping exercise.
That is the short version, and most comparisons stop at cost. Cost is the least interesting variable here. Since 2 August 2026 the AI literacy duty in the EU AI Act has been enforceable, which adds a criterion almost nobody is applying: whichever route you take now has to produce a record of who was trained, on what, and when.
In-house vs external AI training at a glance
Criterion | In-house | External provider |
|---|---|---|
Time to first session | 3 to 6 months, including design, a pilot and sign-off | 2 to 6 weeks, most of it scoping |
Cost shape | Internal hours, front-loaded, flattens after year one | A fee per cohort or per head, recurring |
Keeping it current | Your build owner, against a release cycle measured in weeks | The provider, inside the fee, if the contract says so |
Credibility with sceptics | Strong on context, weak on authority | Strong on authority, weak on context until scoped |
Evidence for an Article 4 check | Whatever your LMS captured, if you have one | Attendance and content records, if you asked for them |
Main failure mode | The person who built it leaves | The content is generic and nobody applies it |
What does running AI training in-house actually involve?
Running AI training in-house means owning five jobs, not one: curriculum design, content production, delivery, refresh, and record-keeping. Most internal business cases price the third and forget the other four.
Design is where the hours go. Someone has to decide what a finance analyst needs from Microsoft Copilot that a recruiter does not, write exercises against real work rather than toy prompts, and get the resulting material past whoever owns data policy. Refresh is where programmes die. The tools move on a cadence of weeks, so a deck built in March describes a product that no longer exists by September, and your build owner is the only person who can fix it.
The internal advantage is real and it is about context. An internal team can build an exercise around your actual CRM, your actual client confidentiality rules, and the specific report that eats a Tuesday every month. No provider knows that on day one, and some never learn it.
What does an external AI training provider actually give you?
An external provider gives you a curriculum somebody else has already tested on other organisations, a delivery schedule you did not have to build, and the refresh burden off your team's desk. What you are buying is mostly time.
This is the majority route in the UK, and a shrinking one. Of employers who trained staff in the previous 12 months, 54% provided external training, down from 57% in 2022, according to the Department for Education's Employer Skills Survey 2024, a telephone survey of 22,712 UK employers. Where employers did go external, 76% sourced it from commercial organisations rather than colleges, regulators or suppliers.
The weakness is generic content, and it is a fair criticism. A provider running the same ChatGPT Enterprise session for a law firm and a logistics business is selling awareness, not capability. The test is whether the exercises reference your work or somebody's idea of your sector, and you can settle that before signing by asking to see the actual materials for one role.
What does each option cost?
The honest answer is that in-house looks cheaper because its cost arrives as hours rather than an invoice. UK employers spent an average of £1,700 per employee on training of all kinds in 2024, down from £1,960 in 2022 in real terms, on the same Department for Education figures. An AI programme is competing inside a budget that has been falling for over a decade.
For the build route, the sum is: design hours plus rehearsal hours, at a loaded internal day rate, plus delivery hours multiplied by the number of cohorts, plus a refresh block every quarter for as long as the programme runs. The last term is the one that gets left out, and it is the one that never stops. Work it out before the business case goes in.
For the buy route, the number is a quote, so the risk sits in what the quote excludes. Ask whether scoping is billed separately, who owns the materials afterwards, and whether next year's refresh is inside the fee or a new statement of work. Our post on rolling out AI training across a 100-person team covers how to compare provider quotes without anchoring on headline price.
How long does each take to get running?
External training gets to a first session in two to six weeks. In-house takes three to six months to reach the same point, and that is with a willing champion, because design, pilot and internal sign-off run in sequence rather than in parallel.
Both numbers describe when training starts, not when it works. Those are separate clocks and conflating them is how budgets get set wrong, which we pulled apart in how long AI training actually takes. Whichever route you pick, first competence still lands weeks after the first session and habit change lands months after that.
Speed matters more than usual right now. The Employer Skills Survey found 14% of UK employers using AI at their site in 2024, rising to 24% at sites with 100 or more employees, and 86% of those users planned to embed it into operations within three years. Adoption is arriving faster than most L&D calendars are built to handle.
What happens when the person who built it leaves?
The in-house programme stops being maintained, and usually nobody notices for two quarters. Key-person risk is the single most underrated line in the build case, because the champion who designs an AI curriculum is by definition one of your most employable people.
Three things reduce it, and none are free. Write the curriculum so a second person can deliver it without the first one in the room. Keep source files somewhere the L&D team owns rather than in a personal drive. Name a deputy from the start and give them a share of the delivery, so the knowledge has somewhere to sit.
Providers have a version of the same problem and it is worth naming. Your account team churns, the engagement ends, and if nothing was transferred, the capability walks out with them. The fix is symmetrical: contract for the materials and a train-the-trainer session, so the last thing the provider does is make itself less necessary.
Which option can you evidence under the EU AI Act?
External training evidences more easily, because a provider produces attendance records and dated content as a by-product of billing you. In-house programmes generate the same evidence only if somebody deliberately builds the record, and most do not.
This stopped being theoretical in August 2026. Article 4 of the EU AI Act has applied since 2 February 2025 and became enforceable on 2 August 2026, requiring providers and deployers to ensure a sufficient level of AI literacy among staff using AI systems on their behalf. It reaches UK organisations with EU staff, EU customers, or EU-facing systems.
The European Commission's own guidance on Article 4 is unusually practical about what this means. There is no mandated level, because sufficiency is contextual to the role, the risk and the system. Relying on the instructions for use, or asking staff to read them, "might be ineffective". No certificate is required, and no AI officer post: organisations can keep an internal record of trainings and other initiatives, and that record is what you will be asked for.
So the criterion that decides it is whether the route you choose leaves a trail. If you build, decide on day one where completion is logged and who owns that system. If you buy, put attendance records and dated materials in the contract rather than assuming they arrive.
Who should choose which?
Build in-house if you have three things at once: an L&D function that already runs programmes, a technical champion with protected time rather than goodwill, and work specific enough that generic material would be useless. Financial services firms with heavy supervisory rules often sit here, because the compliance overlay is theirs alone and no provider will learn it quickly.
Buy externally if you need people capable this quarter, if AI expertise is thin internally, if your L&D team is already at capacity, or if you want somebody else carrying the refresh. Buy if the sceptic you need to convince is your COO, because an outside voice settles arguments an internal one cannot.
Do both if you are like most organisations of 100 to 500 people. Buy the first two cohorts, take the materials, and internalise delivery once your champion has seen it run twice. That sequence gets you speed first and ownership second, which is the right order.
Where ivee is the wrong answer: if you have fewer than about 20 people to train, a self-serve course will get you most of the way for a fraction of the cost. If your real constraint is that nobody has decided what AI is allowed to touch, buy governance work before training, because teaching people to use tools you have not sanctioned creates a problem rather than solving one. And if you already have a working internal programme, ask us to review it rather than replace it.
What people get wrong about this choice
The most common error is treating it as permanent. Build versus buy is a decision about this year's cohort, not a constitutional position, and the organisations that handle AI upskilling well tend to switch routes as their own capability changes. The second error is pricing the build at design and delivery while ignoring refresh and record-keeping, which is how a cheap-looking internal programme becomes an expensive stale one.
The third is assuming external means generic. It means generic by default, and specific if you scope it and check the materials. That is a procurement problem, and it is fixable in one meeting. Our guide to training your team in AI covers what to settle before either route starts.
If you want a second opinion on which side of this line you sit, talk to us. We will tell you if the answer is to build it yourself.
