Supervisory technology now gives two very different kinds of answer. A deterministic check gives the same result every time for the same input: a return either passes the rule or it does not. A probabilistic model gives a likelihood: this return looks unusual, this paragraph probably matters. Both are useful. The skill is knowing which one is allowed to decide.
Two kinds of answer
Deterministic means the logic is written down and fixed. Total assets must equal liabilities plus equity. A capital ratio is calculated the same way for every bank. An XBRL formula assertion either holds or fails. Anyone can rerun it and get the same result, and explain it to the firm, a board or a court.
Probabilistic means the answer is a judgement from patterns. A model ranks a return as an outlier against its peers, or reads a risk report and suggests which paragraphs a supervisor should see. It can find things no rule was written for, but the same input can produce a slightly different answer, and the reasons are harder to show.
Quantitative and qualitative data need both
Quantitative data is the numbers in regulatory returns. Rules check them first: completeness, arithmetic, thresholds. Probabilistic models then look across firms and periods for what the rules cannot see, such as a ratio that is legal but moving unusually fast.
Qualitative data is everything written: board papers, capital and liquidity assessments, risk narratives, complaints. Here AI does the reading at a scale no team can, and suggests what matters. What it finds is then checked against the figures and the rules, and a supervisor decides what it means.
How Line4 puts them together
Line4 uses both, in a fixed order, from the moment a document arrives.
Extract. Supervisors receive everything from scanned letters and PDFs to Excel workbooks and XBRL files. Line4 uses AI, OCR, decision APIs and several models together to pull the data out of whatever arrives, so it can be checked the same way whatever its shape.
Check with rules. The data then goes through two kinds of rule. Calculation rules work like XBRL: the taxonomy's own formulas where a standard exists, or Line4's rule engine where it does not. Skills-based rules go further: they ask questions of the data in plain language, such as whether a narrative supports the figures beside it. Several models answer, and judge models from different providers check each answer, to catch hallucinations and keep the AI honest.
No guesses. If the data is not there, nothing is invented: the question goes back to a person. Even then, average handling time is a fraction of doing the work by hand, because everything that could be read and checked already has been.
People decide. By the time a supervisor reviews a document, it has been read, checked and reviewed by more than ten processes. The supervisor makes the call, and each finding links to its data, its rule and the person who made it. AI takes on more of the preparation only as evaluation sets show its suggestions matching your experts. The decision itself stays with people.
The BIS Financial Stability Institute makes a related point about AI explainability: there may be trade-offs between explainability and model performance, “so long as risks are properly assessed and effectively managed” (FSI, Sep 2025). A deterministic final step is one way to manage that risk. In February 2026 the ECB told banks that AI governance needs “human accountability for AI-driven decisions” (ECB). Supervisors can hold their own tools to the same standard.
Outside Europe: Canada's OSFI makes the same point to the firms it supervises. Its Guideline E-23, which takes effect on 1 May 2027, counts AI and machine learning methods as models, and says the rationale for autonomous or “black box” models should consider the “need for alternative controls”. A deterministic rule is one such control.
Why it matters for real-time supervision
As supervision moves from quarterly returns towards data that arrives continuously, both halves have to keep up. Rules must run in seconds so firms fix errors before they file, and AI must read at the same pace so supervisors see what changed. Read more in what real-time supervision means.
- 1If the decision must be repeatable, write it as a rule, not a model.
- 2Make every AI suggestion point to the data it read, and have a model from another provider check it.
- 3When data is missing, send the question to a person instead of letting a model guess.
- 4Measure AI against your experts on an evaluation set before widening its role.
- 5Put a deterministic check between any AI output and anything a firm sees.
- 6Name the person accountable for each finding, and record their decision.
Sources: ECB, Encouraging innovation, managing risks, Patrick Montagner, 3 February 2026; BIS Financial Stability Institute, Managing explanations: how regulators can address AI explainability, 8 September 2025; Central Banking, Supervision Benchmarks 2026; BIS, IFC Report 19. See how Line4 applies this on the SupTech platform page.
For supervisors
New guides and Insights, by email
One short email when we publish something worth your time, from Jess Luper. No spam, unsubscribe any time.
