A life insurer ran regulatory monitoring as a manual craft: a compliance team of two to three people watching national and EU legal sources, filtering hundreds of publications a month down to the five or ten that actually touch the business, by eye. FlowX.AI built a regulatory and compliance insight agent that scans the sources daily, decides what is relevant, classifies it into the insurer’s own legal categories, and reasons about its impact on the insurer’s written procedures, with a confidence score on every match and a human validating every judgement.
Live in the pilot
What if the hard part of compliance was not finding the law, but knowing which five of several hundred publications actually matter?
The old way ran on reading and recollection. A compliance team of two to three people watched the national legal information system, an EU legal database, the national draft-legislation portal, the national insurance supervisor, EIOPA, the tax and anti-money-laundering authorities, the data-protection authority, and ministry and government pages. They collected the publications, decided by eye what was relevant, assessed the impact, and once a month formatted a compliance report for the management board and the board of directors. Gathering the laws was never the problem. Filtering them was.
Regulatory monitoring, screening and impact assessment ran almost entirely by hand across a team of two to three. Four things defined the starting state.
Collecting the laws was never the problem. Out of hundreds of national and EU publications a month, perhaps five to ten actually touched the insurer, and that call was made by reading a title or a summary and drawing on know-how, sometimes on intuition.
Sources were checked by hand, relevant developments spotted by eye, and impact judged from experience. There were no formal legal categories to classify against; the taxonomy simply did not exist.
Checking a new regulation against internal procedures, and updating those procedures, was a different workflow entirely. It sat outside monitoring and screening, and ran only when a matter was important enough to convene a working group.
Up to 30 days could pass between a regulation being published and the board becoming aware of it, slowing strategic responses and budget approvals. A team of three could not widen its coverage without simply working more hours.
The team no longer collects and formats. Each pressure now has a concrete answer, built to work with the sources, the procedures and the reporting rhythm that already existed.
Two to three people manually check a long list of national and EU sources, filtering hundreds of publications down to the five to ten that matter, by eye.
A daily automatic scan captures everything published since the last run. The agent decides what is relevant to the life-insurance domain, classifies it, and extracts the metadata: publication date, number, type, status and stage, and the several deadlines a single law can carry.
No formal legal categories existed, so classification depended on individual know-how and, in the team’s own phrase, intuition.
The agent classifies into categories the insurer defined, with descriptions in its national language and English written to guide it, and derives the impacted department from the owners of the impacted procedures.
Gap analysis against around 100 relevant internal procedures (of around 300) ran as a separate, working-group-triggered workflow.
Opening a law shows the impacted procedures ordered by confidence score, each gap labelled missing, partial, leverageable or compliant, with an impact description, proposed next steps, and a department-level roll-up so each team sees only what concerns it.
Up to a 30-day lag from publication to board awareness, with no deadline tracking attached to anything.
Monitoring now flows straight into deadline tracking. Impacted procedures carry a status (not started, in progress, updated) and an active-deadlines view aggregates them against thresholds: urgent under 45 days, upcoming 45 to 90, and beyond, with an overdue card for anything past due.
The monthly board report was a manual assembly ritual, and laws that never appear online, draft acts received by email and agency guidelines, had no route into any system.
One button reproduces the report: choose the period and the sources, export to Excel for reuse in Word and PowerPoint, so the report becomes a by-product of the dashboard. A manual company-uploads channel feeds offline laws to the agent, which analyses them for gaps exactly like crawled sources.
It scans the sources daily, decides what is relevant, scores its own confidence against the insurer’s procedures, and hands a human a ranked queue to validate. Nobody re-reads a title to guess relevance any more.
An automatic daily scan captures everything published since the last run, with a guarded manual trigger as a once-a-day fallback. The agent decides which of the day’s publications are relevant to the life-insurance domain. One run serves all users.
It classifies each relevant publication into the insurer’s own legal categories and extracts the metadata: number, type, status and stage, plus adoption, publication and entry-into-force dates. It reads natively in the insurer’s national language and English, with no translation round-trip.
Opening a law shows a summary, a timeline of dates, and the impacted procedures ordered by confidence score, each gap labelled four ways (missing, partial, leverageable, compliant) with an impact description, proposed next steps, and a department-level roll-up.
Deadline tracking runs against 45- and 90-day thresholds with an overdue card; one button exports the board report to Excel; and a per-procedure impacted / not-impacted control turns the team’s judgement into a training signal that calibrates the agent over time.
The value was never in gathering the laws. It was in deciding which ones matter, reading them in their own language, and getting better at both over time, on three moves.
From grounding the agent in the insurer’s own procedures to a board-ready export, with the agent proposing and the team validating every judgement.
Internal procedures
The team uploads and maintains its ~100 relevant procedures, each defined by name, type and owner department, updated by re-uploading a new version so the latest text always wins. Fully filterable and sortable, with view, edit and delete on every entry.
External legal sources
For the pilot, three sources feed the agent, an EU legal database and two national sources, accessed through web crawling, plus a manual company-uploads channel for laws that never appear online. The wider source set is reserved for phase 2.
The daily run
An automatic daily scan captures everything published between the last run and now, with a manual scan-sources trigger guarded to once a day as a fallback. One run serves all users.
Identify, classify, extract
The agent decides which of the day’s publications are relevant to the life-insurance domain, classifies each into the insurer’s legal categories, and extracts number, type, status and stage, plus adoption, publication and entry-into-force dates.
The dashboard, in three streams
Results land in one shared dashboard split into adopted laws, draft laws and company uploads, separated because they behave differently: drafts and uploads get a gap read but no procedure-status monitoring.
Impact & gap analysis
Opening a law shows a summary sized small, medium or large, the timeline of dates, the impacted procedures ordered by confidence score with four-way gap labels, an impact description, proposed next steps, and a department-level roll-up.
Feedback loop
Every impacted procedure carries an impacted / not-impacted control. That feedback calibrates the agent, turning the team’s judgement into a training signal instead of throwaway effort.
Deadlines & status
Statuses (not started, in progress, updated) and an active-deadlines view against 45- and 90-day thresholds, plus an overdue card.
Report export
One button, with the period and sources selected, exported to Excel rather than PDF because the team reuses the data in Word and PowerPoint.
Human in the loop
The agent proposes, the team validates and feeds back. Data lives on FlowX.AI SaaS, EU-hosted.
The process and UI build landed on plan. The extra time went where it mattered: calibrating the agent’s judgement, and a mid-project decision to widen the tool from a document-diff into a regulatory radar.
Kick-off
Scope, roles, success criteria and governance agreed, with weekly status and bi-weekly demos. On plan.
Discovery
Two discovery sessions plus a prototype walkthrough; the end-to-end journey and business requirements agreed and signed off. On plan.
Building
The process and UI backbone delivered on plan: Sprint 1 on mock data, around 80% of screens, then the AI agent, its calibration, the national-language localization and the UAT environment.
Testing & UAT
The stretch that earned its time. AI calibration and validation took a few more iterations than planned, and a mid-project scope expansion, from gapping single procedures to reading the whole domain, widened the tool from a document-diff into a regulatory radar. Go-live followed shortly after UAT was completed.
The plan at kick-off was compact: two weeks of discovery, four weeks of build, two weeks of testing, with go-live targeted for early summer. The process and UI backbone landed on plan; the additional time went into AI calibration and validation, which took a few more iterations than expected.
A mid-project scope expansion did more than move a date. The moment the requirement grew from only gapping single procedures to reading the whole domain and then gapping, the tool stopped being a document-diff and became a regulatory radar.
Three lenses on the same result: the coverage and speed it gained, the quality and control it now enforces, and the reporting and adoption it reshaped.
A monthly ritual became a daily radar, and awareness that used to lag by up to 30 days now arrives the day a publication lands.
Relevance and impact are reasoned, scored and explained, in an organization that had no formal legal taxonomy before the project.
The board report becomes a by-product of the dashboard, and every judgement the team makes teaches the agent.
On the surface, this automates regulatory monitoring and impact assessment, useful and measurable in compliance-team hours saved and in closing the publication-to-board lag. Under the surface, it is the insurer’s first flow where an AI reads primary legal sources natively, decides what is relevant to the business, classifies it, and reasons about its impact on the insurer’s own written procedures, with a confidence score, four-way gap labels, an audit trail, and a human-in-the-loop feedback loop that calibrates it over time.
The scope pivot is the tell. The moment the requirement moved from only gapping a procedure to the whole domain plus gaps, the tool stopped being a document-diff and became a regulatory radar. It changes the compliance team’s day: a manual monthly screening ritual becomes a living radar, the board report becomes a by-product, and the team shifts from collecting and formatting to judging, with every judgement teaching the agent.
Phase 2 widens the radar. The remaining monitored sources, the national insurance supervisor, EIOPA, the tax and anti-money-laundering authorities, the data-protection authority, and ministry and government pages, come into the crawled set. Beyond that, the same pattern of native legal reading plus internal procedures, producing classified, confidence-scored, explainable impact with a full feedback trail, extends into media sentiment and law-version diffing. The pilot is proof that the radar can be widened.
We are the enterprise orchestration platform that enables regulated institutions to deploy deterministic, zero-hallucination agentic workflows on top of legacy infrastructure. Insurers run critical customer and operations journeys on FlowX.AI, including:
Bring a monitored domain and we will show you a path to production: native legal reading, relevance decided, impact gap-scored against your own procedures, the human in the loop and the audit trail intact.