An applicant tracking system: what the audit found, and how we fixed it
We ran an EU AI Act audit on an applicant tracking system for hiring. The engineering guardrails were strong. The documentation, the conformity route and the retention evidence were not in place. This case study sets out what the audit found, and how we fixed it.
- Obligations checked 28 10 do not apply
- Applicable 18 3 met, 8 partial, 7 missing
- Findings 8 High 4 Medium
- Tier High-risk Annex III, point 4
At a glance
- Sector
- HR technology, recruitment and applicant screening
- System
- An applicant tracking system for hiring. It parses resumes, applies hard filters, scores candidates with model personas and writes a synthesis for recruiters.
- EU AI Act scope*
- Annex III, point 4: employment. The audit places the use in the high-risk tier. The role assessed is provider.
- Assessment date
- 21 October 2026
- What was audited
- The screening pipeline, guardrails, storage and interface, read from source. The README, configuration and tests. Runtime behaviour in a test environment, with test resumes only.
- What the team has now
- An obligation map covering 28 obligations, a signed AI bill of materials, runtime checks, and a written report with the remediation for each open item.
- Stack detected
- Python, SQLite or Postgres, Agno (optional), Anthropic, Ollama or OpenAI-compatible model endpoints, Docker
Context
A hiring tool shapes who gets seen.
The EU AI Act treats systems that screen or rank job applicants as high-risk under Annex III, point 4. That brings duties on data, records, human oversight, technical documentation and conformity.
Teams building or hosting a tool like this ask the same questions:
- Which duties apply to a screening tool?
- What evidence exists for each duty?
- What is missing, and who closes it?
What we audited
Our audit reads the code and the documents.
It then runs the system and checks what it does. For this system, the method had five parts. The audit assumed EU jurisdiction.
- Static scan. A scan of the codebase that builds a signed AI bill of materials, listing the AI components found.
- Source review. The screening pipeline, guardrails, persistence and user interface.
- Document review. The README, configuration and tests.
- Runtime test. The system run in an offline stub mode, with no API keys and test resumes only.
- Real-model pass. A run with one model, on test resumes only, with each case repeated three times.
What the audit found
We checked 28 obligations. 18 apply to this system.
Of those, 3 are met with evidence, 8 are partial and 7 are missing. The other 10 do not apply. Severity follows the audit's priority rating.
-
HIGHArticles 43, 47 to 49
No conformity route in place. There is no conformity assessment, EU declaration of conformity, CE marking or registration.
Why it mattersBefore a high-risk system is placed on the EU market, the provider must complete a conformity assessment and register it (Articles 43, 47 to 49).
-
HIGHArticle 5
No documented prohibited-practice screen.
Why it mattersArticle 5 bans certain AI practices. Without a written screen, the check cannot be shown.
-
HIGHArticle 11 and Annex IV
No Annex IV technical documentation set.
Why it mattersArticle 11 and Annex IV require a technical file. The codebase has architecture notes, not the Annex IV sections.
-
HIGHArticle 9
No risk management system with residual ratings.
Why it mattersArticle 9 asks for a risk management process across the life of the system. The guardrails exist. There is no risk register with owners, residual ratings or a review cadence.
-
HIGHArticle 12
Security events cannot be traced to a job.
Why it mattersArticle 12 asks for logs that allow traceability. An injection flag is stored with an empty job ID and does not appear in the per-job audit view.
-
HIGHArticle 15
Years of experience not established.
Why it mattersArticle 15 asks for appropriate accuracy. In the test runs, experience parsing did not establish years from an eight-year resume, so a five-year requirement could not be checked.
-
HIGHArticle 73
No serious incident reporting process.
Why it mattersArticle 73 asks providers to report serious incidents to the authorities. No procedure was found.
-
HIGHArticle 10
No bias measurement on outputs and no retention policy.
Why it mattersArticle 10 asks for data governance. Personal details are stripped before model evaluation, but output bias is not measured and candidate data has no retention rule.
-
MEDIUMArticle 13
Instructions for use do not cover limitations or oversight.
Why it mattersArticle 13 asks for instructions on limits and human oversight. Deployment and configuration are documented. These duties are not.
-
MEDIUMArticle 17
No quality management system.
Why it mattersArticle 17 asks for a documented quality management system. None was found.
Two more items are Medium: audit log retention and export are not defined (Articles 12 and 19), and there is no post-market monitoring plan (Article 72).
Runtime checks
- Personal details were absent from the model-facing output.
- An unqualified resume with an injection payload was rejected by a hard filter before any model call.
- A qualified resume with an injection payload was flagged at ingest.
- Across three runs per case, the injection payload did not change the recommendation. All six runs returned "maybe". The average fit score was lower in the injection cases, 38.3 against 50.7. The model's scores varied between runs, so this is not a clean measure of the payload's effect.
- Six evaluation probes were run. Three passed and three failed. The failures were the audit-trace check, the experience-parsing check and the test-suite check.
- The test suite passed 118 of 121 tests. The three failures concern serving static design-system assets.
- Recruiters can override hard filters, bias flags require review, and the system does not act on a candidate by itself.
How the system works
The system takes in PDF, DOCX and text resumes. It flags exact and near-duplicate resumes. Its hard filters are deterministic. A candidate who fails a must-have is rejected before any model call. Untrusted text passes a guardrail that fences the prompt and checks for injection. Personal details are stripped before a model sees the profile. Persona models then evaluate the candidate. A synthesis step writes the recommendation, and borderline cases go to a recruiter for review.
Human oversight
Recruiters review the output for each candidate. They can override the hard filters. Bias flags require a review. Borderline cases are routed to a person. The system does not act on a candidate on its own.
What we did next
After the report, our forward-deployed engineers either fixed the issues hands-on or advised and worked alongside the team to close them.
Triage ran by severity first. Then came the Article 5 screen and the technical file, because the provider route depends on both. The team kept the decisions that belong to it, such as the legal basis for processing candidate data and the wording of any candidate notice.
-
Fixed hands-on
Our engineers attached a job ID and a candidate ID to every security event, so an injection flag now appears on the job audit view. They set up bias measurement on model outputs, and a risk register with owners and residual ratings that is reviewed after each model change.
-
Advised and worked alongside the team
Our engineers completed and signed the Article 5 screen with the team before the next release. They wrote the Annex IV technical file with the team, linking each section to the evidence behind it, and planned the provider route: conformity assessment, EU declaration of conformity and registration. Together with the team they defined serious-incident and corrective-action procedures, with timelines, and a retention and erasure rule for candidate data. Teams that host the system themselves carry deployer duties as well; the engineers set out a plan for those too.
What the team has now
An obligation map.
Each of the 28 obligations has a status, a confidence rating, evidence references and a remediation.
A signed AI bill of materials.
It lists the AI components found in the codebase.
Runtime checks.
They cover prompt-injection handling, the personal-data check, the audit-trace check and the experience-parsing check.
Repeated runs.
Three runs per case, with fit scores and recommendations.
A written report.
It sets out the remediation for each open item.
Open items
-
OPEN ITEM
The legal basis for processing candidate data and the wording of the candidate notice stay with the team and its legal advisers.
-
OPEN ITEM
The conformity assessment, the EU declaration of conformity and registration are planned but not yet completed; they gate placing the system on the EU market.
-
OPEN ITEM
Post-market monitoring needs a standing plan with an owner and a review cadence.
-
OPEN ITEM
Teams that host the system themselves carry deployer duties alongside the provider's, and must evidence them in their own right.
Questions teams ask
Talk through a system like this
Building or hosting an AI hiring tool? Bring the codebase and your review process. We will run the same audit and walk through the findings with you.
Read on: Credit decision support: what the audit found, and how we fixed it · Healthcare AI toolkit: what the audit found, and how we fixed it
* EU AI Act scope: Annex III, point 4 (employment). The tier is the one the audit returned for this system.
