Claude training: an AI Act Article 4 curriculum
What Claude training has to cover to close the AI Act literacy duty: a curriculum per role, the settings generic AI courses miss, and how to prove it happened.
TL;DR
Claude training for a European company has two jobs: make people productive, and close the AI Act literacy duty that has applied since 2 February 2025. Those need different content. Productivity training teaches prompting and tools. Literacy training teaches what the system does with company data and where it fails. Build both, and record the second.
What does AI Act Article 4 actually require?
Article 4 obliges providers and deployers of AI systems to take measures so that their staff, and other people operating AI systems on their behalf, reach a sufficient level of AI literacy. It is scaled to the role and the context of use. It does not require you to guarantee a specific level for each person, and it does not prescribe a curriculum, an exam or a certificate.
Two features of the duty shape how you should design the training. First, it applies to deployers, not only to the companies building models, so an ordinary company rolling out Claude chat is in scope. Second, it covers contractors and anyone else operating the system on your behalf, which is the part most training plans miss.
The date matters more than the content. The European Commission’s regulatory framework page puts the AI literacy duty in force since 2 February 2025, alongside the prohibited practices, with broad application of the Act from 2 August 2026. The 2026 Omnibus moved the high-risk deadlines to December 2027 and August 2028, and left the literacy duty where it was. If your Claude rollout started in 2025 and the training is still scheduled, that gap is already open.
What this means in practice:
| Question | Answer |
|---|---|
| Who is in scope? | Staff and other persons operating the system on your behalf, contractors included |
| What level is required? | Sufficient for the role and context, not a defined standard |
| Is a certificate required? | No, but you need to be able to show the measures you took |
| When did it apply? | 2 February 2025 |
| Does the Omnibus change it? | No, only the high-risk timelines moved |
Our AI Act checklist for LLM deployers covers the other deployer duties that sit alongside this one.
What goes in the curriculum, by role?
Four audiences need different content, and bundling them into one session wastes everyone’s time. Chat users need data boundaries and failure modes. Developers need cost behaviour and permission models. Admins need the settings that enforce policy. The DPO and risk owners need the audit surface and the retention picture.
The curriculum below is our recommendation, sized for organisations rolling Claude out across departments:
| Audience | Duration | Core content |
|---|---|---|
| Chat and Cowork users | 90 minutes | What leaves the company, what the retention setting is, where Claude is confidently wrong, when to escalate to a human |
| Power users and champions | Half day | Projects, skills, connectors, prompt structure, reviewing output against a source |
| Developers on Claude Code | Full day | Permissions and sandboxing, context management, model choice, cost behaviour, review discipline on generated code |
| Admins and platform owners | Half day | Residency settings, groups and spend limits, audit logs, the Compliance API, offboarding |
| DPO, risk and internal audit | Half day | Processor chain, retention per hop, what the audit export contains, which features are eligible for which posture |
Two content decisions separate a useful curriculum from a generic one. Teach the failure modes with your own material, not with toy examples: take ten real outputs from the pilot, half of them wrong, and have the group find the errors. And teach the cost model to developers, because they are the ones who control it.
Which settings does generic AI training always miss?
Generic AI courses teach prompting. They do not teach the four settings that decide whether your deployment is lawful and affordable: where inference runs, what the retention period is, who can read the conversations, and what makes a session expensive. Those four belong in every role-specific session, at the depth that role needs.
Where inference runs. On the first-party Claude API, the inference_geo parameter accepts only "global" and "us", and US-only inference costs 1.1x the standard rate on Claude 4.6 and later models. Workspace geo offers only "us" and cannot be changed after creation. EU-region processing comes from regional endpoints on Bedrock, Google Cloud or Microsoft Foundry. Users who do not know which route they are on cannot make sensible decisions about what to paste.
Who can read the conversations. The Enterprise Compliance API gives programmatic access to activity logs, chat histories and file content, filterable by user and time range. Staff should be told that this exists before they use the tool, not after an incident. In several jurisdictions, including Germany, the works council will ask about it anyway.
What makes a session expensive. This is the part developers can act on. Anthropic’s cost management guidance names the behaviours that drive spend: long sessions that are never cleared, Opus left as the default model, cache misses after a break, and agent teams, which use approximately 7 times more tokens than standard sessions when teammates run in plan mode. Background token usage stays under 0.04 USD per session, so idle sessions are not the problem. Context size is.
What the per-developer cost actually is. Teach the real numbers rather than leaving developers to guess. Claude Code averages around 13 USD per developer per active day and 150 to 250 USD per developer per month across enterprise deployments, and stays below 30 USD per active day for 90 percent of users. Our Claude Code pricing guide breaks the routes down further.
Three habits do most of the work, and they fit on one slide:
- Clear the session with
/clearwhen switching to unrelated work, because the full conversation is re-sent on every request. - Match the model to the task. Sonnet handles most coding work; reserve Opus for architecture and multi-step reasoning.
- Keep
CLAUDE.mdunder 200 lines, because it loads into context at session start whether it is relevant or not.
How do you prove the training happened?
Prove it with three records: an attendance list per cohort tied to the role-specific curriculum, the material itself with a version and date, and a measurement that shows the behaviour changed. The first two satisfy the duty. The third is what tells you the training was worth running.
For the measurement, the platform gives you the baseline for free if you switch it on before the training rather than after. Team and Enterprise analytics report weekly active members, adoption level, product stickiness, skill and connector use, and estimated time saved, with a spend report that refreshes daily on a one-day delay and exports one CSV row per person per model. Custom date ranges reach back up to 90 days.
Two limits to design around. The visibility is role-dependent: on Team plans only Owners and Primary Owners see analytics, and on Enterprise, Admins see analytics but not Spend. And on seat-based Enterprise plans the spend report appears only when usage credits are turned on, covering overage spend only, because usage inside the seat allowance is not metered in dollars.
For developers there is a sharper instrument. The /insights command analyses recent local sessions and writes an HTML report on friction points such as misunderstood requests and buggy code, covering up to 200 previously unseen sessions per run. Run it before and after the training with the same cohort and you have a before and after on the actual failure modes, not on self-reported confidence.
Capacity has one practical trap. Rate limits apply at the organisation level, and Anthropic’s recommended per-user allocations fall as teams grow: 200,000 to 300,000 TPM for 1 to 5 users, 50,000 to 75,000 for 20 to 50 users, and 10,000 to 15,000 for 500 or more. The same page notes that live training sessions with large groups may need higher TPM allocations per user, because concurrency spikes in a way normal usage does not. Request the increase before the session, not during it. Our Claude usage limits guide separates plan windows from API rate limits and cloud quotas.
Run it yourself or outsource?
Run the training yourself once you have champions who use Claude daily and can teach with your own material. Buy it for the first wave, for the developer curriculum, and for the DPO and audit session, where the content depends on knowing the platform documentation rather than your processes. The literacy records stay your responsibility either way.
What delivery costs in effort, as our estimate rather than a sourced benchmark:
| Activity | Effort | Note |
|---|---|---|
| Build the role-specific material from pilot output | 5 to 10 days | Needs real outputs, including bad ones |
| Deliver a 90-minute user session | 0.5 day per cohort of 25 | Scales linearly, so champions matter |
| Deliver the developer day | 1 day per cohort of 12 | Hands-on, needs working accounts in advance |
| Admin and DPO sessions | 1 day total | One-off, repeat on plan or feature changes |
| Records, attendance and material versioning | 2 days, then 1 day per quarter | This is the compliance artefact |
| Refresh after a model generation change | 2 days | Cost behaviour and tokenizer change |
Our view: external trainers earn their fee on the developer and admin curricula, because those depend on documentation that changes monthly. They are poor value for the chat user session, which works far better delivered by a colleague from the same department using the department’s own documents.
What to demand from a Claude training engagement:
- Separate curricula per role, not one deck relabelled four times.
- Your own material in the exercises, including outputs where Claude got it wrong.
- The four settings above covered explicitly, with your actual configured values.
- Attendance and material records handed over in a form your compliance function accepts.
- A measurement plan using the analytics that already exist, switched on before the first session.
For the skills and connector content in the power user curriculum, our Claude Skills guide covers the governance side, and Claude Code security covers the permission model developers need to understand.
Training is one phase of a rollout rather than the whole of it. The 90-day adoption plan shows where the curriculum belongs in the sequence, and the identity work that has to come before the first cohort is covered in our SSO and SCIM guide.
FAQ
What is Claude training in a company context?
Claude training is role-specific enablement for the people who will use Claude: data boundaries and failure modes for chat users, permissions and cost behaviour for developers, settings for admins, and the audit surface for the DPO. In the EU it also serves as the measure that closes the AI Act Article 4 literacy duty.
Does the AI Act require Claude training?
Article 4 requires providers and deployers to take measures so staff and others operating AI systems on their behalf reach a sufficient level of AI literacy, scaled to the role and context. It has applied since 2 February 2025. It does not prescribe a curriculum, an exam or a certificate, but you need to be able to show the measures you took.
What does Claude training cost?
We have no sourced market price to quote. Judge an offer against the internal effort: roughly 10 to 15 person-days to build role-specific material, then half a day per 25-person user cohort and a full day per 12-person developer cohort. Build the material once, deliver through internal champions, and the per-head cost drops sharply after the first wave.
Internal champions vs external trainer: which works better?
Champions work better for chat and Cowork users, because they can teach with the department’s own documents and are available for follow-up questions. External trainers are worth paying for on the developer and admin curricula, which depend on platform documentation that changes monthly, and for the DPO session.
How many people can we train at once?
Rate limits apply at the organisation level, and recommended per-user allocations fall with team size, from 200,000 to 300,000 TPM for 1 to 5 users down to 10,000 to 15,000 for 500 or more. Anthropic notes that live training sessions with large groups may need higher allocations because concurrency spikes. Request the increase before the session.
How do we show the training worked?
Switch on analytics before the training. Team and Enterprise reporting gives weekly active members, adoption level, product stickiness and estimated time saved, with a daily spend report on a one-day delay. For developers, run /insights with the same cohort before and after; it analyses up to 200 sessions per run and reports friction points rather than token counts.
Sources
- European Commission: Regulatory framework for AI (28 October 2026)
- Claude Code: Manage costs effectively (28 October 2026)
- Anthropic Support: Usage analytics for Team and Enterprise plans (28 October 2026)
- Anthropic: Data residency (28 October 2026)
- Anthropic Support: What is the Enterprise plan (28 October 2026)
- Anthropic: Model and feature pricing (28 October 2026)