How this analysis was created

What question each skill analysis answers

For each published SFIA skill, the analysis asks:

How might AI change the way this professional skill is practised, and what remains the responsibility of the person doing it?

It deliberately does not try to answer whether AI will replace a particular job or role.

That would require information outside the scope of this work, such as how a specific organisation operates, which technologies it uses, how work is distributed between people and systems, and what controls apply. The analysis is therefore concerned with the professional practice described by SFIA, not with predicting workforce displacement or automation percentages.

A controlled evidence base

Each analysis used only the official published wording for the individual SFIA skill:

• the skill name and short description
• the published guidance
• the descriptions for each applicable SFIA responsibility level

No occupational profiles, job descriptions, external AI capability claims, vendor material or assumptions about a typical employer were used as evidence.

This source boundary is important. It means that an analysis of, for example, Audit is based on what SFIA says Audit involves, rather than on general assumptions about what auditors usually do.

A structured analytical method

Each skill was processed using the same structured, source-constrained method.

1. Decomposing the published skill

The published content was first broken down into elements of professional practice, including:

• professional activities and practices
• judgements and decisions
• interactions with other people
• review, assurance and control responsibilities
• authority and accountability
• policy, standards or strategic responsibilities where these were explicitly present

Each source element retained a reference to the exact part of the published skill from which it was derived.

2. Identifying material effects of AI

The analysis then considered where AI might make a material difference to the practice.

It deliberately filtered out routine productivity uses unless they changed something professionally significant. For example, transcription, formatting or drafting were not treated as important findings merely because they could occur while a skill was being practised.

For each skill, the method normally identified a small number of material effects rather than attempting to list every conceivable use of AI.

3. Distinguishing AI contribution from professional capability

A central test was:

Could someone produce a convincing artefact, answer or recommendation with AI without actually demonstrating the professional practice described by SFIA?

Where the answer was yes, the analysis identified what the generated output did not demonstrate.

This distinction is important because producing an artefact is not the same as demonstrating professional capability.

• A generated audit report does not demonstrate independent assurance.
• A generated architecture diagram does not demonstrate architectural judgement.
• A generated stakeholder plan does not demonstrate effective stakeholder relationship management.

The same principle was applied to interactions and other outputs, not only documents.

4. Distinguishing the role AI plays

The method considered whether AI was acting as:

• a method supporting how the skill is practised
• part of the delivery channel through which the work reaches a user or customer
• a subject or context to which the skill itself is being applied

This distinction matters because the impact of AI is not always simply that a practitioner gains another tool.

5. Preserving professional responsibility

The analysis explicitly separated AI contribution from professional judgement, authority, accountability, validation, approval and assurance.

It also distinguished different professional statuses of AI-produced material where this mattered, such as working material, decision input, evidence and an authorised output.

An AI-generated output was not treated as evidence merely because it might eventually contribute to an evidential process. Where professional validation was still required, that distinction was retained.

6. Respecting SFIA responsibility levels

The analysis considered every published responsibility level for the skill.

It did not assume that the same responsibility simply becomes larger or more complex at higher levels. Instead, it used the actual published progression for that skill.

This was particularly important for skills where responsibility changes significantly across the levels, for example from support and execution through management, organisational direction, policy or strategy.

7. Respecting source silence

The method also checked what the published skill does not say.

The analysis was instructed not to import plausible but unsupported assumptions from occupational practice. For example, concepts such as empathy, trust, ethics, regulation, safety or formal approval authority were included only where the published skill supported them.

This was an important safeguard against turning a source-grounded analysis into a generic description of an occupation.

8. Identifying the distinctive skill insight

For every skill, the method looked for the single most important skill-specific insight needed to avoid a superficial interpretation of AI impact.

Examples include recognising that:

• some skills already contain substantial automation, so AI does not simply introduce automation
• some skills already include machine learning or related techniques as part of the professional domain
• generated code is not the same as software design capability
• a plausible safety artefact is not the same as justified safety assurance
• stakeholder records and indicators are not the relationship itself

The purpose was to make each analysis recognisably about that skill rather than a generic commentary on AI.

9. Checking whether automation or AI was already part of the source

The analysis explicitly checked whether the published skill already referred to automation, machine learning, advanced analytics or equivalent technology-enabled practice.

This prevented the analysis from describing AI as introducing something that the skill already included.

10. Identifying possible capability shifts

The internal analysis also considered whether AI might create changes in:

• knowledge
• methods
• professionally significant behaviours

These were treated as analytical inferences from the published practice, not as claims that SFIA already contains AI-specific requirements.

Empty results were acceptable. The method did not force a generic list of AI literacy, ethics or adaptability requirements onto every skill.

How large language models were used

Large language models were used to apply this method consistently across all 147 skills.

Their role was controlled by the method, the structured source material and the required output format. They were not asked simply to answer an open-ended question such as “how will AI affect this skill?”

The workflow had three separate stages.

Source-constrained analysis

The first stage applied the structured analytical method to the published skill and produced a detailed internal record.

The output was constrained to a defined structure rather than free-form prose. Material analytical findings had to reference the specific published source statements from which they were derived.

Controlled public synthesis

A second stage converted the structured analysis into a concise public-facing explanation.

It produced four consistent sections:

1. Potential uses of AI in this skill
2. Why AI-generated output does not prove someone has this skill
3. What practitioners remain responsible for
4. How AI may affect work at each responsibility level

This stage was not asked to redo the analysis. It could only synthesise the findings already produced.

Independent quality evaluation

A separate model then reviewed both the internal analysis and the public summary against the original published SFIA source.

The reviewer applied nine quality criteria:

1. source fidelity
2. materiality
3. skill distinctiveness
4. artefact/capability distinction
5. responsibility fidelity
6. level fidelity
7. source silence
8. boundary integrity
9. AI restraint

Each criterion was rated as either passing or needing review. A review flag was intended to identify material for human attention, not automatically to declare the analysis unusable.

Automatic validation

Model-based review was supplemented by deterministic checks for matters that do not require interpretation.

These checks included:

• confirming that every source reference pointed to a real statement in the published skill
• checking that exactly the applicable SFIA responsibility levels were covered
• enforcing the agreed public-summary length limits
• checking that structured classification fields used only permitted values
• checking for certain internal contradictions, for example classifying an output as validated evidence while simultaneously saying that validation was still required
• ensuring that excluded metadata did not enter the analysis

All 147 skills passed these deterministic checks.

During development, the checks also caught isolated implementation errors that could then be corrected before continuing.

Calibration before scaling

The method was not applied to all 147 skills immediately.

Thirteen pilot skills

The first stage used thirteen deliberately varied SFIA skills to develop and refine the method.

These covered different forms of professional practice, including content creation, audit and assurance, stakeholder relationships, live infrastructure operations, architecture, software design, data science, incident management, professional development, governance, hardware design, customer service and safety engineering.

The pilot work was read in full and used to identify weaknesses in the method before wider use.

This led to several important improvements, including stronger treatment of:

• generated artefacts versus professional capability
• AI as method, delivery channel or subject
• source silence
• evidence and validation
• existing automation already present in a skill
• the distinctive insight for each skill
• responsibility-level scope

Regression against the pilot set

Once the production pipeline had been built, the thirteen pilot skills were regenerated from the published source using the controlled workflow.

The generated analyses were compared with the human-developed pilot outputs.

The purpose was not to reproduce the same wording. It was to establish whether the pipeline independently recovered the important professional distinctions that had emerged during the pilots.

Six previously unseen skills

A further six skills were then analysed for which no pilot answer existed.

These were deliberately chosen to test different patterns of professional work.

This was an important generalisation test. It showed that the method could derive new skill-specific distinctions from source material it had not previously been calibrated against, rather than simply reproducing the surface patterns of the original pilots.

Model-configuration comparison

Alternative model configurations were also tested against the same calibration set while keeping an independent reviewer.

This was done to establish whether the quality of the analysis depended on one particular model configuration. The selected production workflow was based on comparative quality and consistency, not on a single untested model choice.

Human involvement

The analysis was not delegated to AI without human control.

Human work included:

• defining the question and scope of the analysis
• designing the analytical method
• deciding what evidence was permitted
• defining the structured source packets and analytical tests
• reviewing the pilot outputs in detail
• comparing regenerated output against the pilot analyses
• reading and assessing the unseen generalisation batch
• diagnosing recurring weaknesses found by independent QA
• changing the method or writing rules only where there was evidence of a systematic problem
• distinguishing genuine analytical defects from minor editorial issues
• deciding when the method was sufficiently stable for controlled production
• reviewing outputs and exceptions during the full-framework run

The models executed the structured analysis, synthesis and evaluation stages at framework scale. Human judgement established and governed the method and determined how quality findings were handled.

What the review flags mean

Every skill passed the deterministic validation.

The independent quality review was deliberately more demanding. Many skills received one or more “needs review” flags even when the analysis was substantially sound.

A recurring example concerned responsibility scope. A public summary can contain responsibilities that are all genuinely present somewhere in a skill, but still be imprecise if it combines responsibilities from several different SFIA levels and describes them as though they apply to every practitioner.

For example, a skill may progress from support at one level, through operational responsibility and management, to organisational strategy or policy at the highest level. A concise public summary can accidentally flatten these distinctions.

The review process therefore checks not only whether a concept appears in the source, but also:

• which actor holds the responsibility
• at which SFIA level it applies
• whether the wording strengthens the published responsibility
• whether organisational responsibilities have been turned into individual practitioner responsibilities

Residual review flags of this kind are treated as editorial precision issues for human review before publication, rather than as reasons to repeatedly regenerate otherwise sound analyses.

Why this is different from one-shot AI prompting

A single prompt to a general-purpose chatbot can be useful for brainstorming, but it is not equivalent to this method.

One-shot prompting	This analysis
May draw on broad background knowledge	Uses a defined SFIA-only evidence boundary
Usually produces free-form prose	Uses a structured analytical model
No claim-to-source traceability	Material findings reference specific published source statements
No independent evaluation	Separate QA against the original SFIA source
No deterministic validation	Mechanical checks enforce source, level and structural rules
No calibrated baseline	Pilot, regression and unseen generalisation testing
Variable structure between skills	Common method and public-output format across all 147 skills
Difficult to reproduce or challenge	Retains an internal audit trail for each skill

The difference is therefore not simply that a more capable AI model was used. It is the combination of source constraint, structured analysis, traceability, independent evaluation, deterministic validation, calibration and human governance.

What the analysis is, and is not

It is: a source-grounded analysis of how AI may affect the professional practice described by each of the 147 published SFIA skills. It is intended as a starting point for SFIA Foundation review and for organisations considering AI in their own workforce and skills context.

It is not:

• a prediction of job losses
• an estimate of the percentage of work that can be automated
• a statement about the capabilities of any particular AI product
• an assessment of how any particular organisation has deployed AI
• a substitute for organisational judgement about roles, processes, controls or accountability

Actual impact will depend on organisational context and on how AI is designed, deployed and governed.