General or industry-specific AI training: how to decide

A different question from format: what the training actually teaches

“General or industry-specific” is one of the first questions that comes up when preparing AI training, but it answers a different problem than the channel, the length of the programme, or who delivers it. This page deals only with content scope: should training build a general ability to work with AI, target the specific content of one department, or combine the two.

The decision tends to collapse into an assumption that industry content is always more thorough and general content is always cheaper and faster to prepare. Neither claim holds universally. The right answer depends on seven specific properties of a role and its work, and each one can be checked against a real team.

Three candidate answers

Before the matrix, it helps to name the three options actually on the table:

• General training. Teaches a shared ability to work with AI across roles: what a model can and cannot do, how to phrase a request, how to recognise an uncertain or incorrect output, and how to check it before using it. The content does not depend on a specific department.
• Industry-specific training. Targets the working content of one specialisation: its terminology, its typical tasks, its own standard of correctness, and the checks that belong only to that field.
• A combined core and layer. A general part shared by every participant, with an industry layer that varies by department or role and builds on top of the general core.

The formal category of general digital and AI literacy is described by the European DigComp 3.0 framework. It does not itself decide when a particular organisation should add industry content on top; that is the subject of the matrix below.

Decision matrix: seven content criteria

For each criterion, the matrix describes what that property of a role or a job means for the choice of content. The columns are not points to add up; each criterion can point in a different direction, and resolving a conflict between them is covered in the next section.

Criterion	General	Industry-specific	Combined core and layer
Specificity of the team’s vocabulary	Work vocabulary is shared and common, such as email, minutes or summaries; general AI concepts are enough.	Vocabulary is narrow and technical, such as legal drafting, clinical terminology or a technical specification; general content will not cover it.	The general core teaches universal AI concepts; the layer adds the department’s own vocabulary.
Existence of an industry standard of correctness for output	Checking output is universal: does it make sense, is it readable, does it match the request.	Output is subject to its own industry standard, such as legal precision, accounting classification or a technical tolerance, that general content does not cover.	The core teaches how to recognise model uncertainty generally; the layer teaches the specific industry criteria for checking output.
Share of shared office work in the role	Most of the role consists of activities common across departments.	Most of the role consists of narrowly specific tasks, and the shared component is small.	The size of the layer matches the actual share of industry-specific work in the role.
Industry-specific severity of a wrong output	An error leads to a minor correction with no industry-specific consequence.	An error can create an industry-specific consequence, such as an incorrect legal classification or a safety failure, that the general rule of “check before use” does not cover on its own.	The core builds a universal habit of verification; the layer adds severity criteria specific to the industry.
Size of the trainable group per specialisation	The group is small or the specialisation changes often; dedicated content may not be worth maintaining.	The group is large enough to justify building and continuously updating its own content.	Everyone gets the core regardless of group size; the layer only where it is operationally sustainable to maintain.
Legal availability of internal industry examples for material	General content can be built on public or synthetic examples without special review.	Real relevance grows with real internal examples, which require assessing purpose, legal basis and scope before they can go into material.	The core stays on general examples; the layer uses internal material only after a positive assessment, otherwise synthetic examples.
Speed of change in industry rules, tools or procedures	General principles of working with AI change slowly and do not depend on one specific tool.	Fast-changing industry rules or tools need an owner for the content and a regular review cycle.	The core needs little maintenance; most of the maintenance load sits with the layer, so naming a review owner in advance is worthwhile.

Reading rule: when the criteria conflict

The matrix does not produce a single summary score, and the criteria do not carry equal weight. A simple order applies when reading it.

The existence of an industry correctness standard and the industry-specific severity of a wrong output decide first. If either applies to a role, industry content, whether standalone or as a layer, is warranted regardless of what the other rows show; what is at stake is whether a participant can recognise an error in that field at all.

Group size and the speed of rule change are operational constraints. Where a specialised group is small, a deferred or lighter form of the layer, such as individual consultation instead of a standalone module, tends to work, with a review once the group grows or the rules settle.

Legal availability of internal examples determines where the material for industry content comes from. A negative assessment means synthetic or otherwise altered material with no identifiable personal data; it does not undo the first two criteria’s conclusion about whether industry content should exist at all.

Empirical grounding for the distinction between general and industry-specific work comes from Dell’Acqua and colleagues: among 758 knowledge workers, using generative AI improved performance on 18 tasks inside the boundary of the model’s capability, but reduced accuracy on the one task outside that boundary. The study covers general knowledge work with generative AI, not company AI training specifically; it shows why general literacy on its own does not guarantee that someone can recognise when a task has moved outside the model’s reliable range. That is precisely the gap an industry layer is meant to close where the first two criteria in the matrix call for one.

Content scope and Article 4 of the AI Act

Article 4 of the AI Act, since its amendment by Regulation (EU) 2026/1744, requires measures supporting the development of AI literacy that take into account knowledge, experience, education, training and the context of use. It does not fix the content scope, the format or the duration of training and does not require a certificate. The European Commission’s current Q&A adds that there is no obligation to measure individual employees’ knowledge, and no single procedure is mandatory.

The choice between general and industry-specific content is a pedagogical and operational decision based on what a role actually requires. The law does not make that choice, and the option a company picks does not by itself demonstrate compliance with the statute. An internal training record, if an organisation keeps one, can exist for either general or industry content and shows only that training took place. It does not show whether the content actually matched what the role needed.

Where to go next

The seven criteria above are best checked role by role. Wider background on the Article 4 obligation behind this decision sits in the AI Act’s compliance deadlines from February 2025 to August 2027. The legal-availability criterion for internal examples connects directly to the broader question covered in is it safe to put company data into ChatGPT, and to keeping AI use governed once training has been delivered, in securing AI use across the enterprise. A concrete example of what an industry-specific content decision looks like for one function is set out in the comparison of the best AI tools for Czech content creation, where tool choice already tracks the specific content needs of one team.

Sources and limits

The decision matrix in this article is an editorial tool built by CIAD. The cited sources support individual claims within the text; the matrix as a whole, the ordering of its criteria and the reading rule are a CIAD editorial synthesis that none of the cited sources directly confirms. The current wording of Article 4 and the European Commission’s Q&A were verified on 5 and 6 August 2026 respectively; they describe what the provision requires and does not require, and no recommended content scope for training follows from them. The content choices recommended in this text are CIAD’s own recommendation; the law does not prescribe specific content. The cited Dell’Acqua study concerns general knowledge work with generative AI; it is used here for its finding about the boundary of model reliability, with no conclusion drawn about the effectiveness of any particular AI training content. Whether a specific internal example may be used in training material is a decision for the organisation’s own lawyer or data protection officer.