IMY probes Arbetsförmedlingen's AI: your five-question audit

On 5 November 2026, Integritetsskyddsmyndigheten (IMY) opened a supervisory inspection of Arbetsförmedlingen. The target is the statistical assessment tool that caseworkers use when deciding whether a job seeker should be referred to Rusta och matcha, the outsourced matching service for people who are, or risk becoming, long-term unemployed. IMY says it will examine how the system works, whether the processing is lawful and necessary, which personal data it actually needs, whether it amounts to automated decision-making, and how job seekers are told that it exists. The inspection follows tips to the authority and a run of media reporting since September. It is also the first visible case under IMY's stated 2026 priority: AI in the public sector.

Most Swedish organisations do not run an employment service. Plenty of them do run a model that scores people: credit applicants, insurance customers, job candidates, tenants, patients, students. If that model lives on Azure, IMY's five questions are a ready-made audit. This guide explains what the Arbetsförmedlingen tool actually is, which legal rules sit behind each question, and what evidence your Azure estate can produce for each one before a regulator asks.

What the tool is, according to Arbetsförmedlingen itself

The best description of the system comes from the agency's own analysis report, Arbetsförmedlingen Analys 2023:11, and from IFAU's 2024 evaluation of Rusta och matcha. Neither document is new, and both are public. Together they describe a tool that is far more ordinary than the "black box" label used in the press coverage, which is precisely why the case is useful to everyone else.

Fact	Detail	Source
In use since	2020 for Rusta och matcha; statistical profiling at Arbetsförmedlingen in some form since 2012	IFAU 2024:9
What it predicts	Probability of lasting work or studies within a year, modelled as a piece-wise exponential survival function	Analys 2023:11
Input variables	Sought occupations and experience, education level and orientation, gender, age, country of birth, municipality, disability codes, labour supply, and since April 2023 registration time	Analys 2023:11
Output	A recommendation and a placement in level A, B or C by estimated job chances	IFAU 2024:9
Money attached	Provider fees follow the level: 1,210 kr a month and 20,300 kr results payment in A, versus 1,980 kr and 38,000 kr in C	IFAU 2024:9
Who decides	A caseworker must make the referral, but in practice cannot refer unless the tool has produced that outcome; the caseworker can decline to refer	IFAU 2024:9
Scale	Nearly 60,000 participants in Rusta och matcha in December 2023	IFAU 2024:9

Two details matter more than the rest. First, the feature list includes country of birth, gender, age and disability codes. Those are the attributes any fairness review would treat as sensitive, and two of them are special categories under GDPR Article 9 when they reveal health or ethnic origin. Second, the human in the loop is asymmetric. The caseworker can say no to a referral the tool recommends, but cannot say yes to one it does not. That gate design is exactly the pattern Article 22 of GDPR is concerned with, because a decision that a human can only confirm or veto in one direction may still be "based solely on automated processing" for the people it excludes.

Note what IMY has not done. It has not found a breach. The inspection is at the stage of gathering facts, and the authority will report deficiencies and required measures only once it has them. Kollega reported that the on-site review with Arbetsförmedlingen staff is scheduled for 24 November. Everything that follows is about what a regulator asks and how you would answer, not about the outcome of this case.

IMY's five questions and the law behind each

IMY's press release lists what the inspection covers. Rearranged as questions, with the rule each one tests:

1. Is the processing lawful? GDPR Article 6 needs a legal basis. For a public authority that is normally Article 6(1)(e), a task in the public interest, backed by Swedish law. For special-category data such as disability codes, Article 9 needs a separate exception. The Swedish supplementary provisions in dataskyddslagen and sector-specific registry law fill in the national part.
2. Is it necessary, and which data does it actually need? Article 5(1)(c) data minimisation. The question is not whether country of birth improves the model's accuracy. It is whether the purpose can be achieved without it, and whether anyone documented that analysis.
3. How does the system work? Article 5(2) accountability and Article 35 impact assessment. A regulator expects a description of the model, the training data, the validation, the error rates by group, and the change history. "The vendor knows" is not an answer.
4. Does it involve automated decision-making? Article 22 gives people the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, unless an exception applies and safeguards exist. Swedish administrative law allows automated decisions by authorities in principle (förvaltningslagen 28 §), but that does not remove the GDPR analysis.
5. Were the people informed? Articles 13 and 14 require information at collection, and Article 15(1)(h) gives data subjects the right to meaningful information about the logic involved in automated decision-making and profiling. Aftonbladet's reporting, as summarised by Arbetet, centres on job seekers who did not know the tool had assessed them.

The AI Act adds a second layer that IMY has not invoked in this case but will be entitled to from late 2027. Annex III point 5(a) classifies as high-risk any AI system used by or on behalf of public authorities to evaluate eligibility for essential public assistance benefits and services. Point 4 covers recruitment, selection and worker management in the private sector. Under Regulation (EU) 2026/1744, the Digital Omnibus on AI, the Annex III obligations apply from 2 December 2027. From that date a public-sector deployer must also complete a fundamental rights impact assessment under Article 27 before use and notify the market surveillance authority, which in Sweden for this domain is IMY, as we covered in June when the government designated the national authorities.

Turning the five questions into an Azure audit

The Arbetsförmedlingen model is a classical tabular survival model, not a language model. That describes most scoring systems in Swedish banks, insurers, utilities and HR departments too. On Azure those models typically sit in Azure Machine Learning, with data in Azure SQL, Fabric or Databricks, and the services around them can already produce most of the evidence IMY is asking for. The gap is usually that nobody has assembled it.

1. Lawfulness: one document per model, not one per department

Create a record for every model that scores a natural person: legal basis, the national law or contract it rests on, the Article 9 exception if any sensitive attribute is an input, and the data protection impact assessment reference. Microsoft Purview Compliance Manager has assessment templates for AI regulations and for GDPR, and the useful part is not the score but the evidence attachments and owner assignments, which give you a dated trail. Keep the DPIA itself where lawyers can read it. Link it from the model's registration in Azure Machine Learning using tags, so the technical and legal records point at each other.

2. Necessity: a feature inventory with a reason column

Data minimisation fails audits because the feature list exists only in a training notebook. Export it. For each input, record the source table, the classification, whether it is a protected attribute or a proxy for one (postcode and municipality often are), and the documented justification for keeping it. Purview's data map and classification give you the first two columns for data in Azure SQL, Fabric and Databricks. The justification column is a human decision, and it is the one a regulator reads.

Then run the ablation nobody wants to run: retrain without each sensitive feature and record the accuracy change. If removing country of birth costs you very little predictive power, you have a necessity problem whether or not anyone complains. If it costs a lot, you have a fairness question instead, and you need the next section.

3. How it works: the Responsible AI dashboard is built for this

Azure Machine Learning's Responsible AI dashboard was designed for exactly the tabular classification and regression models at issue here. It combines fairness assessment from Fairlearn, error analysis, interpretability from InterpretML, counterfactual what-if analysis and data exploration in one view, and it exports a PDF scorecard intended for compliance stakeholders. The fairness component lets you define cohorts by sensitive attributes and compare selection rates and error rates across them. The interpretability component gives you both global feature importance and per-person explanations, which is the raw material for Article 15(1)(h) answers.

Two constraints to plan around, from Microsoft's own documentation: the dashboard supports MLflow-registered scikit-learn models (or models wrapped to expose predict and predict_proba), and it visualises at most 5,000 data points, so you sample. If your model is a survival model built elsewhere, as Arbetsförmedlingen's appears to be, the wrapper route works. Even without the dashboard, the underlying Fairlearn library runs anywhere Python does:

from fairlearn.metrics import MetricFrame, selection_rate, false_negative_rate
from sklearn.metrics import accuracy_score

mf = MetricFrame(
    metrics={
        "accuracy": accuracy_score,
        "selection_rate": selection_rate,     # share recommended for referral
        "false_negative_rate": false_negative_rate,
    },
    y_true=y_test,
    y_pred=model.predict(X_test),
    sensitive_features=X_test[["gender", "country_of_birth_group", "disability_code"]],
)

print(mf.by_group)          # metric per cohort
print(mf.difference())      # largest gap between cohorts, per metric

Store the output of that run with the model version, every time you retrain. A regulator who asks "how does it work" is satisfied by a dated record of what the model does to each group, far more than by a methodology slide.

4. Automated decision-making: log the human, not just the model

The Arbetsförmedlingen design, where the tool's output is a precondition for referral, shows why "a human makes the final decision" is not enough on its own. To show that human involvement is meaningful, you need evidence that humans actually change outcomes. Log every scored decision with the model version, the recommendation, the human's final decision, and whether the two differ. Send those events to a Log Analytics workspace alongside the Azure Machine Learning diagnostics, and report override rates monthly. If the override rate is close to zero in one direction, your process is automated in practice regardless of what the policy says, and you should either redesign the gate or satisfy the Article 22 exception and safeguards explicitly: a legal basis in law, the right to obtain human intervention, to express a point of view and to contest the decision.

For the AI Act, this same log is your Article 12 record-keeping and your Article 14 human-oversight evidence once Annex III applies in December 2027. Build it once.

5. Information: write the notice before the model goes live

The most repeated complaint in the Swedish coverage is that job seekers did not know they had been profiled. The fix is not technical. Your privacy notice, and the letter or screen the person actually sees, must say that a statistical model is used, what it is for, what it means for them, and how to request a human review. Article 15(1)(h) means you also need a process that produces a plain-language explanation for an individual on request, which is where the per-person explanations from the interpretability work earn their keep. Test that process with an actual request before a journalist does.

Who else this applies to: a decision guide

The Arbetsförmedlingen case is public-sector, but the same five questions attach to private models with very little translation. The table maps common Swedish scoring systems to the rules that will be asked about.

System	GDPR pressure point	AI Act status from 2 Dec 2027	Swedish supervisor
Public benefit or service eligibility	Art 22, Art 6(1)(e) basis, Art 9 for health data	High-risk, Annex III 5(a); Art 27 FRIA required	IMY
CV screening and candidate ranking	Art 22, Art 13 notice to applicants	High-risk, Annex III 4(a)	IMY (data), Arbetsmiljöverket and DO for the employment side
Consumer credit scoring	Art 22 with Art 22(2)(a) contract exception, Art 15(1)(h)	High-risk, Annex III 5(b); Art 27 FRIA required	IMY and Finansinspektionen
Life and health insurance pricing	Art 9 health data, Art 22	High-risk, Annex III 5(c); Art 27 FRIA required	IMY and Finansinspektionen
Marketing propensity scores	Art 21 right to object to profiling	Not high-risk	IMY

If your system sits in one of the first four rows, the five-question audit above is not optional preparation for 2027. It is the GDPR work IMY can ask about today, as it is now asking Arbetsförmedlingen.

The Swedish and EU angle

IMY has said what it is looking at this year. The press release states that AI in the public sector is a prioritised area for IMY's supervision and guidance in 2026. Municipalities, regions and agencies running scoring or triage models should read this inspection as a template for the questions they will receive. The authority asked for documents from Arbetsförmedlingen before opening the case, so the first contact is a records request, not a site visit.

Procurement spreads the exposure. Rusta och matcha is delivered by private providers paid according to the level the model assigns. A supplier whose revenue depends on a public authority's model has a direct interest in how that model is documented, and a procurement team that buys a scoring service from a vendor inherits the deployer obligations under Article 26 of the AI Act. Write the five questions into your upphandling requirements and your data processing agreements now, while contracts are being renewed for 2027.

The timeline is set, and it is shorter than it looks. Annex III obligations apply from 2 December 2027. EN 18286 is already in the Official Journal for the Article 17 quality management system, as we covered in October. Harmonised standards for data governance, logging and human oversight are still pending, so the GDPR evidence you build now is the only concrete material you can show a Swedish supervisor for the next thirteen months, and most of it carries straight over into the AI Act file.

What to do this quarter

• Inventory every model that scores a person. Include vendor-supplied scores and rules engines with statistical thresholds, not just models your data scientists built.
• Produce the feature inventory with justifications for each model, flag protected attributes and proxies, and run the ablation on the sensitive ones.
• Register models in Azure Machine Learning with MLflow and attach a Responsible AI dashboard run and scorecard to each production version.
• Log recommendation versus final decision for every scored case into Log Analytics and publish the override rate to the model owner and the DPO.
• Rewrite the data-subject notice and the Article 15(1)(h) response procedure, then test it with a real request.
• Classify each model against Annex III and, for public bodies, draft the Article 27 fundamental rights impact assessment skeleton now, so it is a revision rather than a project in 2027.

Conclusion

The Arbetsförmedlingen inspection is not about an exotic algorithm. It is about a survival model with a documented feature list, a caseworker gate that only opens one way, and people who were not told. IMY's five questions test whether an organisation can account for a model that affects lives, and Azure Machine Learning, Purview and Log Analytics can produce most of the answers if someone wires them together before the records request arrives. Do that work for your own scoring models this quarter. The regulator has told you what it will ask.