AI use case

AI job matching and guidance for public employment services

AI job matching for public employment services matches job seekers registered with the service to vacancies, training and support from their skills and work history, and suggests next steps to the counselor. The counselor keeps every decision on benefits, sanctions and the support a person receives.

Last verified 18 November 2026 · 4 public deployments

At least 1 million
Users served
VDAB (organization claim).
EUR 108,000 to EUR 1.4 million
Indicative value per year
A regional employment service with 150,000 registered job seekers and 400 counselors. Worked example, see how it is calculated.

What problem does it solve?

A public employment service holds hundreds of thousands of job seeker profiles and a large stock of vacancies, and its counselors have a few minutes per person. Matching by job title or keywords misses people whose skills fit a job they never searched for, and matching by hand does not scale. Services also need to know which job seekers will find work quickly on their own and which need help sooner.

Services use AI in three ways: skills extraction and vacancy matching for the job seeker, ranking and screening candidates for a vacancy for the counselor, and predicting a person's chance of finding work to prioritize contact. All three touch decisions about people's access to help, so the evidence on this page includes the audit findings and the failures: an audit office that found a new matching model less effective than the older system, a gender difference in recommended vacancies and a profiling system stopped by a data protection authority.

How does it work?

  1. Read the profile. A language model reads a CV or the registered profile and extracts skills, which the person can check and edit.
  2. Match to vacancies. The system ranks vacancies by fit, using skills, past hirings and applications. For a vacancy it can instead rank job seekers.
  3. Suggest, never impose. The job seeker sees suggestions and chooses whether to act. The counselor sees the ranked list or a predicted chance of finding work.
  4. Screen by message. Some services use a messaging assistant to ask the candidates about interest, availability, hours and pay, and write the answers into the file for the counselor.
  5. Check for bias. The service compares what different groups are offered, such as men and women, and takes differences to an ethics body before they are accepted.
Audience
Customer facing
Autonomy
Assist
Adoption
Early adopters
Channels
Email, SMS, Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI job matching and guidance for public employment services
KPIMedianReported rangeData pointsClaimed by
Users servedNot pooled
60,000 to 1 million
21 organization, 1 regulator or auditor

Value drivers: Speed and cycle time, Employee productivity, Customer experience, Inclusion and access.

Indicative value

A regional employment service with 150,000 registered job seekers and 400 counselors

EUR 108,000 to EUR 1.4 million

Counselor time released per year

How this is calculated

Formula: contacts * shareUsingMatching * minutesSaved * costPerMinute. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Counselor contacts with job seekers per year contacts, contacts per year600,000600,000The reference service, about four contacts per registered person.
Share of contacts where the matching list replaces manual searching shareUsingMatching, fraction of contacts0.10.3Editorial assumption, replace with your own usage data. VDAB reports that measuring who found work with AI is still its main difficulty.
Counselor minutes saved per such contact minutesSaved, minutes per contact38Editorial assumption, replace with a measured time study.
Fully loaded cost of a counselor minute costPerMinute, EUR per minute0.61Editorial assumption, replace with your own cost.

What it leaves out: Counselor time only. It leaves out license and integration cost, the effect on how fast people find work (no deployment on this page publishes one) and the cost of bias monitoring and ethics review.

Which organizations already do this?

4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.

France Travail

France · Government and public sector · 2025

PilotGrade A

Vadore (Valorisation des données de recherches d'emploi) ranks job offers for each registered job seeker with machine learning and deep learning, using past hirings and applications, so that the ten best offers can be sent to the person by email. It has been developed since October 2018 with a research team and differs from the older rules based matching system the agency has used for about ten years.

A 2019 to 2022 experiment in the Auvergne-Rhône-Alpes region measured fairness and found gender differences in the offers distributed. A larger experiment with about 300,000 job seekers started in February 2025. The Cour des comptes noted the development had taken seven years, and the tool was still in testing when it audited the agency.

  • Users served: about 60,000, respondents in two phases of an experiment from 2019 to 2022
    "Afin d’évaluer l’équité et les biais de cet algorithme, une expérimentation a été mise en œuvre au sein de la région Auvergne-Rhône-Alpes (ARA) entre 2019 et 2022 auprès de panels limités (environ 60 000 demandeurs d'emploi répondants en deux phases)."
    Claimed by: regulator or auditor

France Travail

France · Government and public sector · 2024

ProductionGrade A

MatchFT helps France Travail advisers fill a job offer. A profile to offer matching system ranks registered job seekers by their presumed fit with the offer, and the adviser can have a generative AI assistant text up to 50 of them to check interest, availability, hours and salary. The answers tell the adviser how well each person matches and update the person's file. The Cour des comptes reports that a new statistical AI matching tool tested against the existing system proved less effective, so the agency kept the existing system and added only the conversational part. At the end of 2024 the tool was deployed in 87 of 800 agencies.

No outcome disclosed.

VDAB

Belgium · Government and public sector · 2024

ScaledGrade C

VDAB, the public employment service of Flanders, uses AI to read a job seeker's CV, extract skills and link them to vacancies the person might not have considered (the Jobbereik tool), and shows job suggestions and similar vacancies from its Talent API to logged in users. It also runs a model that predicts whether a job seeker will find work within six months and uses it to decide who the call center phones first, with an adviser checking the estimate. VDAB's head of AI says that measuring who found work with AI remains a challenge. The AI found that women were offered more part time vacancies, which VDAB took to its ethics council.

  • Users served: at least 1 million, over the past two years, users of the Jobbereik and Orient tools
    "Al meer dan 1 miljoen mensen gebruikten AI-tools van de VDAB om een job te zoeken"
    Claimed by: organization

Arbeitsmarktservice Österreich (AMS)

Austria · Government and public sector · 2020

Paused or rolled backGrade C

The Austrian public employment service AMS developed AMAS, the Arbeitsmarktchancen-Assistenzsystem, a logistic regression model estimated on past case data by Synthesis Forschung that sorted unemployed people into groups with high, medium and low labour market chances so that support such as training could be directed more efficiently, with advisers keeping the final decision. It was due for national use in early 2021, but the Data Protection Authority stopped it in August 2020, citing the missing legal basis and prohibited individual decisions by profiling. The Federal Administrative Court later annulled that decision, but all AMAS data had been deleted and the project is finished.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Job seeker profiles, CVs and vacancies, with a skills taxonomy
  • Outcome data (applications, placements and time to work) linked to what was suggested
  • A legal basis for profiling and for the use of the score in counseling

Systems to integrate

  • Public job portal and the vacancy database
  • Counselor desktop and case management system
  • Messaging channel (email, SMS) for suggestions and candidate screening

Complexity: High

The build needs the job seeker and vacancy databases with clean skills data, a matching service connected to the public job portal and the counselor desktop, a measurement plan that follows people into employment, and a legal basis and ethics review for any scoring of individuals. France Travail's Vadore took seven years from start to a large scale experiment.

  1. 1

    Start with suggestions the person can ignore

    Show skill based vacancy suggestions on the job portal first. The person chooses, and no consequence for benefits follows from a suggestion.

  2. 2

    Test against the system you already have

    Run a new model against the current matching on the same job seekers before replacing it. France Travail found that a new statistical matching tool performed worse than the system it had used for a decade and kept the old one.

  3. 3

    Measure group differences from the start

    Compare what men and women, age groups and people with a migration background are offered, and send differences to an ethics body.

  4. 4

    Keep the counselor in charge of consequences

    Use scores to decide who to call first or what to propose, never to decide benefits, sanctions or access to support.

Guardrails

  • No benefit, sanction or access decision is made from a score or a ranking
  • A person can see and correct the skills the system extracted
  • A documented legal basis and data protection impact assessment before any scoring of individuals
  • Regular comparison of offers across gender and other groups

KPIs to instrument

  • Share of suggestions that lead to a view, an application or a placement
  • Time to work for people who used the tool and a comparison group
  • Difference in vacancies offered between groups
  • Counselor time per contact

Human in the loop

Counselors decide what to propose and judge whether a model estimate fits the person. They own every decision on benefits, support and sanctions. An ethics body reviews group differences in the suggestions.

Common failure modes

Bias reproduces stereotypes
VDAB found that the system offered women more part time vacancies than men, and France Travail's experiment found gender differences in offers. Measure differences before launch and decide with an ethics body what to do about them.
A profiling system stopped before launch
Austria's AMAS, which sorted job seekers into groups of labour market chances, was stopped in August 2020 by the Data Protection Authority, citing the missing legal basis, and was never rolled out. Secure the legal basis for profiling before the build.

What are the risks and rules?

EU AI Act

High risk

Annex III point 4(a) lists AI systems used for recruitment or selection, in particular to place targeted job advertisements, to filter applications and to evaluate candidates, and Annex III point 5(a) lists AI used by public authorities to evaluate eligibility for essential public benefits and services. A system that recommends vacancies to job seekers or ranks candidates for them, or scores people to decide on support, falls under one or both, with the duties for risk management, data governance, human oversight and logging.

Rules that apply

Guidance

Controls to put in place

  • Data protection impact assessment and a documented legal basis for profiling
  • Bias testing by gender and other groups before and after launch
  • Ethics body review of differences found
  • Log of suggestions, scores and counselor decisions

When it went wrong elsewhere

Frequently asked questions

Does AI job matching help people find work faster?

No deployment on this page publishes that result. VDAB reports more than a million users of its AI tools and says the challenge is to measure who found work with AI. The Cour des comptes reports that France Travail's new statistical matching tool performed worse than its decade old system, so the agency kept the old one.

Is AI job matching biased?

It can be. VDAB found that its system offered women more part time vacancies than men and took the question to its ethics council. France Travail's experiment from 2019 to 2022, with about 60,000 job seekers, found gender differences in the offers distributed, part of which came from preferences the candidates expressed.

Why was Austria's AMS algorithm stopped?

The Data Protection Authority stopped AMAS in August 2020, citing a missing legal basis and prohibited individual decisions by profiling. The Federal Administrative Court annulled that decision more than five years later, but the data had been deleted and AMS regards the project as finished.

How to cite this page

Blits.ai AI Use Case Library, "AI job matching and guidance for public employment services", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/public-employment-job-matching. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 18 November 2026: First published

Related use cases

Government and public sector

AI for benefit fraud and error detection in social security

Benefit fraud and error detection uses risk models that help a social security or benefits agency decide which claims, payments and recipients to check for fraud or error. Caseworkers then verify the riskiest cases first, while every decision on entitlement stays with a person and the model is tested for fairness before and during use.

Deployments
6 public, best grade A
Reported detection improvement
2.5x
Department for Work and Pensions, organization claim
Government and public sector

AI assistant for benefits eligibility questions and applications

A benefits eligibility and application assistant is an AI assistant that helps people see which public benefits and grants may apply to them, explains the rules and the required documents in plain language, answers questions on claim status and payments, and guides the person through the application and checks it for completeness, while the eligibility decision stays with the agency's rules and caseworkers. General questions about government services are the citizen information assistant's job.

Deployments
8 public, best grade B
Reported accuracy
97%
Department for Work and Pensions, organization claim
Government and public sector

AI drafting copilot for civil servants for correspondence, briefings and ministerial replies

A civil servant drafting copilot is a generative AI assistant that drafts replies to correspondence from the public and elected representatives, briefings, submissions and summaries for civil servants, grounded in the department's approved lines, policy documents and case data, with the official editing and approving every word before it is sent or cleared; it includes the general text assistants governments build for their own staff. Bills, explanatory notes and answers to parliamentary questions belong to legislative drafting and parliamentary question support, and an office suite assistant rollout such as Microsoft 365 Copilot to workplace AI assistant rollout.

Deployments
17 public, best grade A
Reported employee adoption
at least 50%
Government Technology Agency of Singapore, organization claim
Government and public sector

AI for freedom of information request processing

AI for freedom of information request processing is AI that helps a public body handle freedom of information and open government requests. It logs and clarifies requests, spots duplicates, searches and deduplicates the records in scope, proposes redactions with the applicable exemption, and drafts the response letter, with an FOI officer deciding what is released.

Deployments
8 public, best grade B
Reported accuracy
98%
U.S. Food and Drug Administration, organization claim