Source: [AI skills inventories: separate inference from demonstrated ability — Skills for Business](<https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/>)

[← All articles](https://skillsforbusiness.ai/blog/)

HR & people · Guide

# AI skills inventories: separate inference from demonstrated ability

Use AI to organize skills evidence without treating a job title, course record or extracted phrase as proof of proficiency. A practical guide for HR and L&D.

[Skills for Business](https://skillsforbusiness.ai/blog/editorial/)1 November 20267 min read

On this page

1.  [Extraction research answers a narrower question](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#extraction-research-answers-a-narrower-question)
2.  [A shared vocabulary does not supply individual evidence](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#a-shared-vocabulary-does-not-supply-individual-evidence)
3.  [Preserve the kind of evidence you actually have](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#preserve-the-kind-of-evidence-you-actually-have)
4.  [A fictional service-improvement team](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#a-fictional-service-improvement-team)
5.  [Make unknown a usable planning state](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#make-unknown-a-usable-planning-state)
6.  [Pilot the evidence transitions before expanding](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#pilot-the-evidence-transitions-before-expanding)
7.  [Sources](https://skillsforbusiness.ai/blog/ai-skills-inventories-inference-demonstration/#sources)

What to take away

-   Keep extracted skill mentions, proposed taxonomy matches and demonstrated proficiency in separate fields.
-   A missing skill in the available documents means unknown, not incapable.
-   Build learning plans from reviewed evidence and unresolved questions rather than a single AI-generated skills score.

Use AI to prepare a skills inventory as a set of claims to review, with the source and evidence type attached. Do not let an extracted phrase, a job-title association or a completed course silently become a proficiency rating. The inventory should help HR and learning teams decide what to investigate and develop next, while showing what they do not yet know.

The practical distinction is between finding a possible skill and establishing an ability to perform particular work. An assistant may identify “data analysis” in a project description. That alone does not establish who performed the analysis, what tools they used, how independently they worked or whether the result met the required standard.

Our proposed method keeps those questions visible. It concerns development planning and team capability discussions, not automated hiring, promotion or other individual employment decisions. The examples below are fictional; we have not tested a skills platform or evaluated employees.

## Extraction research answers a narrower question

The July 2022 NAACL paper [SkillSpan, by Mike Zhang and colleagues](https://aclanthology.org/2022.naacl-main.366/), studies identifying skill and knowledge spans in English job postings. Its annotated dataset comes from three sources, and its experiments compare language-model approaches using span-level precision, recall and F1.

That is a text-extraction task: did the model identify the relevant passage according to the annotation rules? It is not an assessment of whether an employee possesses the ability named in the passage. Job postings themselves describe requirements, rather than demonstrations by particular people.

The paper is historical evidence about a defined task and dataset. It does not establish the accuracy of today’s assistants on your project records, languages or internal terminology. We use it to clarify the extraction boundary, not to estimate a current product’s performance.

For an internal pilot, test extraction separately from the conclusions drawn afterward. A system might correctly identify a phrase and still attach it to the wrong person, map it to an overly broad skill or invent a level. One reassuring overall score would conceal those different failures.

## A shared vocabulary does not supply individual evidence

A taxonomy can make skills easier to organize across teams. The European Commission’s ESCO guidance distinguishes [essential](https://esco.ec.europa.eu/en/about-esco/escopedia/escopedia/essential) skills associated with an occupation from [optional](https://esco.ec.europa.eu/en/about-esco/escopedia/escopedia/optional) ones whose relevance depends on context, employer or country.

Those relationships concern occupational profiles. Our implication for an employee inventory is straightforward: a relationship between an occupation and a skill is useful context for a question, but it does not prove that a specific person has demonstrated that skill. Record it as an occupation-based suggestion when that is its origin.

Keep the local meaning alongside the common label. “Reporting” might mean assembling a standard monthly report in one team and designing an analytical method in another. If both become the same broad label, a dashboard can look consistent while obscuring the work that needs coverage.

A mapping review should therefore ask whether the label preserves the task, scope and relevant conditions. Leave uncertain mappings unresolved rather than forcing every phrase into the nearest available category. A smaller inventory with interpretable entries can support a more useful development conversation.

## Preserve the kind of evidence you actually have

Mercer’s practitioner guidance on [skills measurement](https://www.mercer.com/en-jo/insights/talent-and-transformation/skill-based-talent-management/addressing-common-skills-measurement-challenges/) distinguishes rated, inferred and validated methods. Its inferred category includes AI tagging from records such as CVs, work history and training; its discussion of demonstration concerns skills assessed in work-like scenarios. Mercer also sells related services, so this is provider guidance, not independent validation of its tools.

Our recommendation is to retain evidence categories rather than collapse them into one apparent truth. Self-report, a manager’s rating, course completion and a reviewed work example can all contribute information. They answer different questions and can disagree.

A useful record can remain quite small:

| Field | What to preserve |
| --- | --- |
| Proposed skill and local task | The label plus the work it is intended to describe |
| Source and attribution | Where the claim came from and whom the source actually concerns |
| Evidence type | Mention, self-report, occupation-based inference, training record or reviewed demonstration |
| Scope and conditions | Tools, assistance, responsibility and circumstances covered by the evidence |
| Review status | Proposed, confirmed within a stated scope, disputed or unresolved |
| Date and next question | When the evidence arose and what would resolve the remaining uncertainty |

This is an editorial record design, not a validated assessment instrument. “Confirmed” should never stand alone: confirmed as what, by whom and against which task criteria? An employee confirming that a course record is accurate has not necessarily confirmed independent ability to perform the corresponding work.

## A fictional service-improvement team

Imagine an operations team planning a new reporting service. It has approved project summaries, voluntary skills statements and training records. An assistant proposes that several colleagues have “dashboard development” skills.

The review finds three different situations. One person completed a dashboard course. Another coordinated a project in which an external supplier built the dashboard. A third built and maintained a reporting view, with a colleague reviewing the calculations.

All three records may contain relevant information. They do not support the same conclusion:

| Available record | What it can support | Useful next question |
| --- | --- | --- |
| Course completion | Participation in the named learning activity | Which relevant task can the learner now perform? |
| Coordination summary | Involvement in the project, with responsibility to clarify | Did the person build, specify, review or coordinate the work? |
| Reviewed reporting example | Evidence about the stated work under its actual conditions | Does the new service require capabilities beyond that example? |

The learning plan might offer practice to the course participant, clarify the coordinator’s contribution and arrange an extension exercise for the person with the work example. These are possible development responses, not conclusions about real people.

Notice the benefit of preserving attribution: coordinating a supplier can be valuable work in its own right. Treating it as either dashboard construction or “no technical skill” would lose the actual contribution. Ask what the source establishes before interpreting what it omits.

## Make unknown a usable planning state

No mention of a skill in the available records should remain “not established from these sources.” It should not become “does not have this skill.” Otherwise, differences in documentation can masquerade as differences in capability.

Give employees a clear way to correct attribution, add relevant evidence and explain the conditions of a work example. Do not resolve disagreement simply by asking the same model to issue a more confident rating. Preserve the dispute and identify what additional information would help.

For team planning, distinguish missing evidence from a confirmed development need. A capability map can show areas with reviewed examples, areas with only reported or inferred coverage and areas that remain unknown. Avoid calculating a precise “skills gap” percentage from those mixed states without explaining the denominator and evidence rules.

Use approved sources and access arrangements for the pilot. More data is not automatically better evidence: a large collection of casual messages can still leave responsibility and performance unclear. Keep personal records accessible only to the people who need them for the agreed development purpose.

## Pilot the evidence transitions before expanding

Start with one service and a few clearly described tasks. Use approved sample material to check whether the assistant preserves attribution and uncertainty. Include a course-only record, a team achievement without individual responsibility and an occupation-based suggestion. Check that none is automatically promoted to demonstrated proficiency.

Also test corrections. If a person explains that a project summary describes a colleague’s contribution, the revised inventory and any team summary should reflect that correction. Retaining an accurate source link is insufficient if the downstream capability map still counts the original inference as established coverage.

Our guide to [learning after AI assistance](https://skillsforbusiness.ai/blog/ai-workplace-training-measure-learning-after-assistance/) addresses how to design a related task that reveals retained understanding. That can inform a development exercise where appropriate; it does not validate a broad rating inferred from a document.

Agree who reviews the pilot, which task criteria they use and what evidence remains unresolved. Expand only when the inventory helps people identify a concrete next learning or coverage action. A useful skills inventory makes the boundary between a suggestion and demonstrated work easy to inspect.

## Sources & further reading

Sources checked for this article. Research findings and vendor claims are distinguished in the text.

1.  [SkillSpan: Hard and Soft Skill Extraction from English Job Postings](https://aclanthology.org/2022.naacl-main.366/)Zhang and colleagues; NAACL 2022 / ACL · Accessed 1 November 2026
2.  [Essential](https://esco.ec.europa.eu/en/about-esco/escopedia/escopedia/essential)European Commission, ESCO · Accessed 1 November 2026
3.  [Optional](https://esco.ec.europa.eu/en/about-esco/escopedia/escopedia/optional)European Commission, ESCO · Accessed 1 November 2026
4.  [Addressing common skills-measurement challenges](https://www.mercer.com/en-jo/insights/talent-and-transformation/skill-based-talent-management/addressing-common-skills-measurement-challenges/)Mercer · Accessed 1 November 2026

Produced with AI-assisted research and writing. Examples and recommendations are our editorial synthesis unless attributed. This article does not claim independent product testing or human expert review. [Read our editorial policy.](https://skillsforbusiness.ai/blog/editorial/)

## Keep reading

[AI-generated training quizzes: verify the answer key before scoring people](https://skillsforbusiness.ai/blog/ai-training-quizzes-answer-key-review/)[AI-drafted job descriptions: justify each requirement with the work](https://skillsforbusiness.ai/blog/ai-job-descriptions-justify-requirements/)