Skip to content
Fintech

Regulated data arrives inside the prompt

A question to your support assistant carries a national ID, a credit flow sees an IBAN, fraud analysis looks at a card number. That data is inside the prompt — not in your application code. Embed compliance in code and every change in regulatory interpretation becomes a release.

  • National IDs, IBANs and card patterns detected with validation
  • Fallback is disabled while a residency rule is in force
  • The decision trace keeps the rules that matched and lost

0requests leaving the region while a residency rule is in force

What is hard in this sector

The shared problem for teams using LLMs in financial services is where the decision is made. Once model selection, redaction and refusal logic is spread through application code, two things break together: changing a rule gets slow, and the question an auditor asks — why did this answer come out this way — has no single answer.

What is hard in this sector

  • When a regulatory interpretation changes, the rule change enters the release pipeline — typically two weeks.
  • Developers are not privacy specialists; redaction logic gets written wrong in good faith.
  • The audit trail keeps the model name and the rule name, but not the instruction text that produced the answer.
  • Provider keys sit with the team; rotation and leak exposure are engineering overhead.
  • With no safe way to try a new model, migrations keep being deferred.

The first three rules you'd write

These are not illustrative examples; they are the rules we discuss in the first session in this sector. Each is authored as YAML in the console, measured in shadow mode, then enforced.

  1. 1national-id-pin-resident
    When
    signals.pii.types contains a national identity number
    Then
    pin to a model in the local region · mask the prompt · switch the cache off · disable the fallback chain
    Why it is shaped this wayOn a compliance path, failover is a vulnerability. Leave fallback on and the moment the first candidate goes down the request quietly moves to another region — exactly what you were preventing. With no candidate left, the gateway returns an error.
  2. 2card-data-deny
    When
    signals.pii.types contains credit_card
    Then
    refuse the request · 403 · write an audit tag · notify the compliance team
    Why it is shaped this wayRefusing rather than masking and continuing is the cleanest way to keep the system out of PCI scope. Masking is a processing activity; refusal is not.
  3. 3credit-decision-audit
    When
    request.headers.dataClass = credit_decision
    Then
    pin the prompt version · retain the conversation log · audit tag · exclude from canary
    Why it is shaped this wayIn a flow that affects a credit decision, which instruction text ran is the first question an audit asks. Canary must be off here — evaluating the same kind of application against two different prompt versions is not defensible.

What you show in an audit

Decision trace
The winning rule, the rules that matched and lost, the obligations enforced, the model and region selected
Prompt traceability
The identity and version of the prompt that produced the answer
Residency evidence
Which region processed the request, and why the foreign candidate was dropped
Redaction evidence
Which data types were masked; that unmasked content never reached a provider
Change history
Which version of the rule set was in force, and when
Key custody
Provider credentials are not yours; rotation and leak surface are outside your scope

Questions from this sector

When a redaction obligation applies, what reaches the provider is the masked text — [TCKN_TR] in place of the identifier. And if you wrote a residency rule, the request never falls to a foreign candidate in the first place. The two work together: redaction bounds the content, residency bounds the geography.

No. The scanner applies the checksum, so an order number or a random figure produces no signal. That matters: a scanner that treats every eleven-digit number as PII over-constrains production traffic, and the team ends up switching the rule off.

A signal carries not only detected but available. If the scan could not run, policy sees that and can choose to be conservative. "No PII was detected" and "PII could not be scanned for" are different statements — a rule that mistakes a missing signal for a clean prompt is the most expensive error an audit can find.

Modelion makes no certification claim. What it does is make your compliance obligation writable as policy and auditable: move the decision to one place, and produce the evidence automatically. The scope itself remains yours and your legal team's responsibility.

The regulatory references on this page exist to explain which product behaviour they drive; they are not legal advice. The scope of your own compliance obligations is something to establish with your legal team.

Let's start with your own rule set

The first session covers your traffic and your regulatory boundaries; you leave with a divergence report measured in shadow mode and three rules ready to write.