A bank or an insurer can put an AI agent in front of a credit file, a claim or an underwriting memo only if it can answer four questions afterwards: what data did the model see, what checks did the output pass, who approved the action, and where is the record. ATX Agent Hub answers them with a fixed procedure — the AI Guardian checkpoints — and with trace logs written as a by-product of every run. ATX AnalyticsTable adds the piece that keeps personal data out of the model in the first place: native PII masking that pseudonymises customer data on your own machine before anything is sent.
No software makes an organisation compliant on its own. This page maps product controls to the obligations of the EU AI Act (Regulation (EU) 2024/1689), so your compliance, risk and model-validation teams can make their own assessment. It is not legal advice. After the Digital Omnibus on AI, the transparency duties of Article 50 apply from 2 August 2026 and the high-risk obligations for Annex III use cases apply from 2 December 2027.
Most AI agents in finance are not high-risk systems: summarising a policy, drafting a reply, reconciling a report. Two use cases are named in Annex III and matter directly to lenders and insurers: evaluating the creditworthiness or credit score of natural persons, and risk assessment and pricing for life and health insurance. For those, and as good practice for everything else, the Act expects the same few things.
An AI Guardian in ATX Agent Hub is not a single switch. It is a procedure built into the workflow on the canvas, from nodes the platform ships today, so that no model output reaches a customer, a core system or a decision without passing the same four checkpoints. Because the procedure lives in a versioned workflow, it is reviewed, tested and changed like any other production code.
Every workflow run leaves a step-by-step trace: for each node, its status, start and end time, the branch taken, any error, and its output (each output capped at 64 KiB). Reviewers read the trace in the run observer; access is scoped by row-level security, so a person sees only the runs of projects they belong to.
Alongside the trace, platform actions — sign-ins, credential use, configuration changes, approval decisions — are written to an audit log with the actor, action, resource, outcome, request ID and IP address. The log is append-only at the database level: UPDATE, DELETE and TRUNCATE are refused. Retention removes whole time partitions and records each removal, so even the deletion of old records leaves evidence. An approval decision is logged with the step and the decision, never with the reviewer's correction text, which may contain the reviewed content.
On the analytics side, ATX AnalyticsTable records every AI call, proposal, refusal and review — including the model the provider reports — and mirrors it into a hash-chained audit trail kept 400 days by default, beyond the six-month minimum in Article 26(6).
Underwriting is where AI saves the most time and where a leak costs the most. ATX AnalyticsTable's native PII masking is pseudonymisation, not deletion: 'Anna Kowalska' becomes a token such as <<PERSON_7f3a>>, and the same name becomes the same token everywhere in the payload. Joins, group-bys and counts still work, so the model can still say which applicant or broker drives a variance. The token map exists in memory for one session and is never written to disk; tokens in the model's answer are restored locally, so the underwriter reads real names and the model never saw them.
Column rules are the control that holds: a column whose name matches iban, pesel, nip, customer_id, card, address and similar patterns is masked whatever it contains. Value detectors are the second net, with checksums — IBAN by mod-97, card numbers by Luhn, PESEL by its check digit — so a long order reference is not destroyed and a real IBAN is not missed.
The two products are connected by configuration, not by a built-in connector: a scheduled, masked export from ATX AnalyticsTable lands in a database table or object storage that an ATX Agent Hub workflow reads on its own schedule.
Automating the preparation of an underwriting file is not the same as letting AI decide it. If AI output is used to score the creditworthiness of a natural person, or to price life or health insurance for one, that use is high-risk under Annex III — and an organisation that puts a general system to such a use can take on provider obligations itself (Article 25). As shipped, ATX products are not high-risk AI systems, and their terms exclude that use.
ATX AnalyticsTable detects per-person AI answers — when a masked column or a personal-data token is in what is sent — and applies a policy of warn (the default), refuse or allow. For underwriting projects set it to refuse: the AI works on portfolios, segments and documents, and the decision about each applicant stays with a person.
Pseudonymised data is still personal data under the GDPR, so masking is a data-protection-by-design measure, not an exemption; your records of processing and DPAs with model providers still apply. Under DORA, both products are self-hosted: ATX Agent Hub on your servers or Kubernetes, ATX AnalyticsTable on Windows or a Linux server, with models reached through Azure AI Foundry, Google Vertex AI or a model you host. That keeps AI agents inside your own ICT risk framework and register of information.
No software does that on its own. The platform provides controls — the AI Guardian checkpoints, human-in-the-loop approval, trace logs and an append-only audit log — that support the obligations and produce the evidence your compliance team needs for its own assessment.
A procedure of four checkpoints built into the workflow: a data gate (masked data only), a policy check that routes the model's draft, a human approval that parks the run, and a record in the trace and audit log. It is assembled from standard nodes, so it is versioned and tested with the workflow.
Using AI output to evaluate the creditworthiness of natural persons, or to price life and health insurance for them, is listed in Annex III and is high-risk from 2 December 2027. Using AI to prepare files, analyse portfolios or draft memos that a person decides on is a different use; set the per-person policy to refuse to keep it that way.
Not with masking on, which is the default. Columns are masked by name, values are caught by detectors with checksums, and blocked columns are never sent. Free text naming someone the rules do not recognise is the weak case, which is why per-person detection and egress ceilings sit behind it.
Yes. Under the GDPR, pseudonymised data remains personal data. Masking reduces what a model provider could learn and supports data protection by design; it does not take the processing outside the GDPR.
Retention is configurable to your regulatory need. ATX AnalyticsTable keeps its AI audit trail 400 days by default; the AI Act asks deployers of high-risk systems to keep logs for at least six months.
This site needs JavaScript for the full experience. Every page above is reachable without it.