Skills intelligence platforms explained: how AI infers skills from HR data
A skills intelligence platform uses AI to build a skill profile for each employee from data the company already holds, such as job profiles, CVs, learning records and activity in work systems, and maps every skill to a taxonomy like ESCO, O*NET or the vendor's own. Companies use these profiles for internal mobility, workforce planning and learning. The inferred skills are estimates rather than verified facts, so buyers should ask how each skill was inferred and which decisions it is allowed to drive.
- The AI reads job profiles, CVs, learning records and work systems, extracts skill mentions and maps them to a taxonomy.
- Public taxonomies differ in size and design: ESCO lists 13,939 skills in 28 languages, Lightcast more than 34,000, and O*NET collects skill ratings for 923 occupations.
- In a peer-reviewed benchmark, the best model extracted skills from English job ads with a span-F1 of 56.64 out of 100.
- Using skills data for promotion decisions is regulated in New York City today, in California from January 1, 2027, and as high-risk AI in the EU from December 2, 2027.
- Most skills tools in our directory publish no prices: 85 percent quote on request.
A skills intelligence platform tries to answer a question most HR systems leave open: what each employee can actually do today, beyond the job title on file. It reads data the company already holds, from job profiles and CVs to learning records and activity in work systems, turns it into a skill profile per person and feeds that profile into internal mobility, workforce planning and learning.
This guide explains how the inference works, which taxonomies the platforms map to, how accurate the underlying methods are and which privacy and employment rules apply. We used vendor pages, the official ESCO and O*NET documentation, published research on skill extraction and the law texts, all checked in September 2026, plus our own directory of 185 AI tools for HR. ai-toolfinder is independent: no vendor pays for placement or for articles.
How a skills intelligence platform turns HR data into skill profiles
Most platforms run the same pipeline. They collect text and activity data, extract skill mentions with language models, map each mention to an entry in a skills taxonomy and combine the results into a profile per person that updates as new data arrives.
The inputs differ by vendor, and the input decides what a skill claim is worth. iMocha lists resumes, certifications, work data, assessments and AI interviews as sources for its skills inference, and retrain.ai extracts skills from CVs and job posts. TechWolf, from Belgium, says its purpose-built models ingest raw work signals, and it feeds the result into the skills engines of the large suites: Workday Skills Cloud, SAP SuccessFactors and Oracle HCM Dynamic Skills.
Some platforms keep the evidence apart. 365Talents describes its skill map as validated, declared and inferred from real work, labels that tell a manager whether anyone has checked a skill. Talent Alpha, from Krakow, indexes more than 3,500 technical skills with machine learning but measures soft skills with questionnaires, while Kaonavi, the Japanese talent management system, offers more than 200 templates for skill maps that people fill in, a declared approach that is only as current as the last update.
The difference shows in a simple case. A support engineer closes tickets tagged Kubernetes in a service desk tool: a model that reads work signals can credit the skill to that person, a CV-based model sees it only if the CV mentions it, and a model that works from job profiles assigns it to everyone in the role, whether they use it or not.
An estimate from indirect signals: a job title, a line in a CV, a ticket closed in a work system. Cheap at scale, unchecked for the individual.
Evidence that someone checked: a proctored assessment, a manager's confirmation, a certificate. Slower to collect, safer to use in a promotion decision.
365Talents separates validated, declared and inferred skills. Ask every vendor whether its profiles make the same distinction.
Skills taxonomies and ontologies: ESCO, O*NET, Lightcast and vendor graphs
Extracted skills only become comparable once they share a vocabulary. A taxonomy is that vocabulary, a list of skills arranged in a hierarchy. An ontology adds relationships: which skills a role needs, which skills sit close to each other and which people hold them.
ESCO, the European Commission's classification, describes 3,039 occupations and 13,939 skills in 28 languages. It maps every occupation to a single code in ISCO-08, the international occupation standard, so it works well beyond Europe. Public taxonomies change slowly: the May 2024 update added 35 occupations, 42 skills and 196 knowledge concepts, and version 1.2.1 followed in December 2025.
O*NET, run for the US Department of Labor, lists 1,016 occupational titles and collects data for 923 of them from job incumbents, occupation experts and analysts. It rates 10 essential and 25 transferable skills and links 8,753 software examples to occupations. Updates arrive quarterly, and version 31.0 came out in August 2026 under a CC BY 4.0 license. Lightcast, a labor market data company, publishes a taxonomy of more than 34,000 skills drawn from job postings and profiles, free to browse.
Vendors build their own layers on top. Eightfold cites 1.6 million skills and 1.6 billion career trajectories behind its matching, and Gloat stores people, skills and roles as connected entities in a knowledge graph. iMocha says its taxonomies connect to O*NET, Lightcast and Singapore's SkillsFuture framework. If a vendor's ontology is proprietary, ask for a mapping to ESCO or O*NET, or your skills data stays tied to that vendor.
Where skills data gets used: internal mobility, workforce planning, learning
Skill profiles are rarely the product on their own. Buyers use them to move people internally, to plan the future workforce and to point employees to learning, and most programs are young. In a Deloitte survey of 1,021 workers and 225 business and HR executives, published in September 2022, fewer than 1 in 5 organizations were adopting skills-based approaches to a significant extent.
Internal mobility platforms match profiles against open roles, projects and short assignments. Fuel50, with offices in New Zealand, Australia, London and North America, adds a Career Advisor Agent and a Leader Coach Agent to an expert-curated skills framework. Gloat's career agent flags internal openings that fit a person's inferred skills, and 365Talents, now part of Docebo, says 2 million employees in more than 60 countries use its platform. Neobrain offers Lexi, a free AI agent that audits an existing skills framework and promises a first result in 10 minutes.
Workforce planning uses the same data in aggregate. TalentNeuron adds external demand data from 3 million job postings a day in 33 languages, and TechWolf scores how automatable tasks are, checked against the Stanford Human Agency Scale. According to TechWolf, HSBC mapped more than 250,000 employees into a single skills taxonomy in 5 months, with granular profiles for more than 30,000 technology staff.
Learning platforms close the loop. Cornerstone lists a Skills Inference Model among its data sources and a Skills Architect agent that shows skills gaps for a person, a team or the whole organization. Workera takes the opposite route and measures skills directly, with more than 100 signature assessments and proctoring, which is slower than inference but gives a verified signal.
Share of the skills intelligence tools in the ai-toolfinder directory, computed from our live dataset.
Accuracy: inferred skills are estimates with measurable error
Skill extraction is a hard language task, and the research shows it. For SkillSpan, a benchmark presented at the NAACL 2022 conference by researchers from the IT University of Copenhagen and LMU Munich, domain experts annotated more than 12,500 skill and knowledge spans in 14,500 sentences from English job postings. The best models matched those annotations with a span-F1 of 56.64 out of 100 for skills and 63.88 for knowledge, such as programming languages.
The experts did not agree perfectly either: after refining their guidelines, the 3 annotators reached a Fleiss' kappa of 0.70 to 0.75, which the authors rate as substantial agreement. Linking a sentence to the right taxonomy entry is harder still. In a 2023 paper, researchers from Ghent University and TechWolf mapped job ad sentences to ESCO skills and reached an R-Precision@5, a measure of how many of the top 5 suggested skills are correct, of 45.74 to 54.62 out of 100 with their best method, up from 29.27 to 32.12 for an earlier approach.
Language models have moved on since then, but the numbers show the size of the problem: a skill named in a text is often ambiguous, and the text rarely says how well someone has it. TechWolf's homepage backs its automation scores with the Stanford Human Agency Scale but gives no accuracy figure for skills inference. The European Commission combined AI with human experts to update ESCO in version 1.2, and that split of machine suggestion and human confirmation is a sensible standard for any buyer.
Ask vendors how they handle stale evidence, shared work credited to a single person and proficiency levels. A CV line that says Python tells you nothing about whether someone can maintain production code.
Privacy and employment rules once skills data drives decisions
A skill profile built from work systems is personal data about how someone works, and the rules tighten once it feeds a decision about that person. Under the GDPR, people have the right not to be subject to a decision based solely on automated processing that significantly affects them (Article 22). A systematic and extensive evaluation based on profiling that feeds such decisions requires a data protection impact assessment before processing starts (Article 35).
The EU AI Act classes AI systems as high-risk when they are used for promotion or termination decisions, to allocate tasks based on personal traits, or to monitor and evaluate performance (Annex III, point 4). Employers must inform workers' representatives and affected workers before putting such a system into use (Article 26), and the obligations apply from December 2, 2027, after the 2026 Digital Omnibus amendment. Our guide to the EU AI Act for HR covers the full checklist.
Other jurisdictions draw the same line at decisions about individuals. New York City's Local Law 144 requires a bias audit within one year before an automated tool substantially assists screening at any point in a hiring or promotion process. California's rules on automated decisionmaking technology took effect on January 1, 2026, cover hiring, allocation of work, compensation and promotion, and require compliance by January 1, 2027. Brazil's LGPD lets people request a review of decisions taken solely by automated processing, including decisions that define their professional profile (Article 20).
The practical line runs between aggregate and individual use. A skills heatmap for workforce planning carries the least exposure, a ranking of named employees for promotion the most, so show employees their inferred profile, let them correct it and keep a person accountable for every decision.
Our reading of the GDPR, the EU AI Act, New York City Local Law 144 and California's ADMT regulations, September 2026.
Dates from New York City's DCWP, the California Privacy Protection Agency and the EU AI Act as amended, checked in September 2026.
What to check before you buy
Skills intelligence is mostly bought by large employers, and prices come late in the process. Our directory lists 13 tools with skills intelligence as a main or secondary focus, and 11 of them quote prices only on request.
Run the checks below on a pilot group with real data, not on a demo tenant. Each tool page in our skills intelligence category records data hosting, certifications and sources as we found them.
If you already run Workday, SAP SuccessFactors or Oracle HCM, check whether the vendor writes skills into the suite's own skills engine, as TechWolf says it does, or keeps a separate profile that employees have to visit. The first keeps a single skills record; the second creates a parallel one that someone has to reconcile.
Settle the exit before you sign. Ask for an export of all profiles with the taxonomy codes attached, such as ESCO or O*NET-SOC identifiers, and ask whether corrections your employees make are kept by the vendor to train its models.
- Evidence per skillAsk which source produced each skill and whether profiles separate validated, declared and inferred skills.
- Taxonomy mappingConfirm which taxonomy the platform uses and whether it maps to ESCO, O*NET or Lightcast, so your data can move with you.
- Measured accuracyRequest precision and recall on a sample of your own employees, and the share of skills employees reject.
- Employee controlEmployees should see and correct their profile before it feeds matching, reviews or succession plans.
- Decision scopeWrite down which decisions the data may drive. Promotion and task allocation trigger high-risk duties in the EU and audit or notice duties in New York City and California.
- Connected systemsList every system the platform reads, from the HRIS to ticketing tools, and run a data protection impact assessment before connecting them.
Tools in this article
Frequently asked questions
What is a skills intelligence platform?
Software that uses AI to infer employees' skills from HR and work data, map them to a taxonomy and keep the profiles current. Companies use the results for internal mobility, workforce planning and learning recommendations. Examples in our directory include TechWolf, iMocha, Workera and Gloat.
What is the difference between a skills taxonomy and a skills ontology?
A taxonomy is a structured list of skills, such as the 13,939 skills in ESCO or the more than 34,000 in Lightcast's taxonomy. An ontology also stores relationships, for example which skills a role needs, which skills are adjacent and who holds them. Most platforms combine a public taxonomy with their own ontology.
How accurate is AI skills inference?
Research gives the best guide, since accuracy depends on the input data. In the SkillSpan benchmark, the best model reached a span-F1 of 56.64 out of 100 when extracting skills from English job postings, and a 2023 study that linked job ad sentences to ESCO skills scored 45.74 to 54.62 out of 100 on its main metric. Treat inferred skills as estimates until an employee, a manager or an assessment confirms them.
Is a skills intelligence platform high-risk under the EU AI Act?
It depends on the use. Annex III classes AI used for promotion or termination decisions, for task allocation based on personal traits or for performance monitoring as high-risk, with obligations from December 2, 2027. A platform used only for aggregate workforce planning is less exposed, but the same product moves into high-risk use once it ranks individuals for those decisions.
How much does a skills intelligence platform cost?
Most vendors do not publish prices. In our directory, 85 percent of skills intelligence tools quote prices only on request. Ask whether the quote is per employee or per module, and which connected data sources it includes.
Sources
- iMocha, Skills intelligence platform (homepage)
- retrain.ai, Talent intelligence platform (homepage)
- TechWolf, Workforce intelligence platform (homepage)
- Decorte and others, Extreme Multi-Label Skill Extraction Training using Large Language Models (Ghent University and TechWolf, 2023)
- 365Talents, Skills intelligence and internal mobility (homepage)
- Talent Alpha, Talent intelligence and the Talent Genome
- Talent Alpha, Privacy policy (registered office in Kraków)
- Kaonavi, Talent management system (homepage, Japanese)
- ESCO, What is ESCO
- ESCO, Occupations pillar and its link to ISCO-08
- ESCO, ESCO v1.2.1 is live (December 2025)
- ESCO, ESCO v1.2 (ESCOpedia)
- O*NET Resource Center, O*NET 31.0 Database
- O*NET Resource Center, Database releases archive
- National Center for O*NET Development, Updating the O*NET-SOC Taxonomy
- O*NET Resource Center, The O*NET Content Model
- Lightcast, Lightcast Skills taxonomy
- Eightfold AI, Talent intelligence platform (homepage)
- Gloat, Workforce knowledge graph (homepage)
- Deloitte Insights, The skills-based organization (September 2022)
- Fuel50, Skills intelligence and talent marketplace (homepage)
- Fuel50, Contact and office locations
- Neobrain, Lexi skills framework agent (homepage)
- TalentNeuron, Labor market intelligence (homepage)
- TechWolf, HSBC customer story
- Cornerstone, Workforce AI and readiness agents (homepage)
- Workera, Skills Intelligence Engine (homepage)
- ai-toolfinder.com dataset, September 2026
- Zhang, Jensen, Sonniks and Plank, SkillSpan: Hard and Soft Skill Extraction from English Job Postings, NAACL 2022
- GDPR, Article 22 (consolidated text)
- GDPR, Article 35 (consolidated text)
- EU Artificial Intelligence Act, Annex III
- EU Artificial Intelligence Act, Article 26
- EU Artificial Intelligence Act, Implementation timeline
- Gibson Dunn, EU AI Act Omnibus Agreement, Postponed High-Risk Deadlines
- NYC DCWP, Automated Employment Decision Tools
- NYC DCWP, Automated Employment Decision Tools: Frequently Asked Questions
- California Privacy Protection Agency, ADMT and cybersecurity regulations approved (September 2025)
- California Privacy Protection Agency, Text of regulations on ADMT
- Brazil, General Data Protection Law (LGPD), Article 20 (English text)