Why Your Financial AI Needs a Query Classification Layer Before It Touches Any Data
The Question Before the Question
Before Stralevo processes any financial question, it answers a prior question: what kind of question is this?
A supplier price search routes to Needle Finder, the cross-document query engine built for that specific task. A bank reconciliation routes to Smatched, the proprietary engine designed for 1:N, N:1, and N:M transaction matching. A quarterly board report routes to VibeFlow, which formats it to the organization's reporting standards automatically. A payroll sensitivity query applies additional data protection controls before any AI processing begins.
The routing takes milliseconds. The alternative — sending every financial query to the same AI endpoint regardless of type — is what every ChatGPT wrapper does. And it is why their answers are slower, less accurate, and structurally unable to apply different controls to different kinds of data.
ChatGPT treats your salary data and your supplier price query identically. Stralevo treats them very differently — before either one touches a model.
---
Query Classification Is a Performance Feature
The most counterintuitive thing about query classification in financial AI: it is primarily a performance feature, not a security feature. Security is a benefit. Speed and accuracy are the reasons.
A reconciliation query routed to a general-purpose language model takes 5 to 10 times longer than the same query routed to Smatched, which was purpose-built for transaction matching and runs against pre-indexed data. A document search routed through a general AI produces lower recall than the same search routed to Needle Finder, which queries a structured Knowledge Graph rather than attempting free-form retrieval.
When a CFO asks "which of our top 10 suppliers changed payment terms in the last 12 months?", the correct routing is to Needle Finder, which crosses 47 contracts and 847 invoices against a persistent entity graph. The same question sent to a general LLM requires the CFO to manually supply the relevant documents, iterate on prompts, and verify the output. Classification is what makes one-question, one-answer financial intelligence possible — not a better model, but better routing to the right engine for each question type.
Every financial AI query falls into one of four types, each handled differently. Retrieval queries — "show me all invoices from this vendor in Q3" — route to the document search layer against the Knowledge Graph. Reconciliation queries — "match these 200 purchase orders against the corresponding invoices" — route to Smatched, where the matching algorithm handles complex N:M relationships that general models cannot process reliably. Analytical queries — "which suppliers raised prices more than 10% over the last 18 months?" — route through Stralevo's Orchestrator, the coordinating engine that assembles context from multiple document layers. Generative queries — "prepare the quarterly board report on supplier costs" — route to VibeFlow, which formats the output to the organization's reporting standards without requiring prompt iteration.
---
What Happens Without Classification
Samsung pasted semiconductor source code into ChatGPT in three separate incidents in April 2023, triggering an immediate internal ban across the company (Bloomberg). Apple, JPMorgan, Goldman Sachs, and Deutsche Bank all restricted or banned similar tools the same year. The underlying problem in each case was identical: a general-purpose AI tool with no classification layer cannot distinguish between a query containing publicly available information and a query containing trade secrets, financial strategies, or employee data. Everything goes through the same infrastructure. Everything receives the same handling.
82% of employees who use AI tools do so from personal accounts, with no organizational classification controls applied (LayerX, 2025). When a finance team member pastes a salary spreadsheet into a personal ChatGPT session, that query contains payroll data — personal financial data about identifiable individuals — that belongs in a restricted sensitivity tier. The tool has no mechanism to recognize this. The data processes through US-based infrastructure with no audit trail and no applicable data processing agreement covering what was actually sent.
The GDPR fine for unauthorized cross-border transfer of personal data reaches 4% of global annual turnover. For a €200 million company, that is €8 million per incident. The exposure is not theoretical: the Irish Data Protection Commission fined TikTok €530 million in May 2025 for cross-border data transfers — the largest data protection fine of that year.
89% of enterprise AI usage is invisible to IT departments: no logs, no SSO, no oversight (LayerX, 2025). In finance specifically, where the queries routinely touch salary data, tax strategies, margin information, and client financials, invisible AI usage means invisible sensitivity exposure. The classification layer makes it visible.
---
How Stralevo's Classification Layer Works
Every query that enters Stralevo passes through three classification steps before any AI processing begins.
Query type classification identifies what kind of work the query requires: retrieval, reconciliation, analysis, or generation. This determines which engine handles the response and which data sets are consulted. A misclassified query — a reconciliation routed to a general LLM, for example — produces a plausible-looking result that may be structurally wrong. The matching logic in a general LLM is inferential; Smatched's matching logic is deterministic and auditable.
Sensitivity classification identifies what data the query will touch. Stralevo's classification tiers range from SL0 (non-sensitive queries where US-hosted models are acceptable) to SL3 (fully on-premise processing for the most sensitive financial data). A supplier price query might route at SL0. A payroll analysis routes at SL2 or SL3 regardless of the user's request, because the data itself triggers the higher tier. This classification happens before the query reaches any model — the system applies the appropriate controls automatically, not contingently on the user remembering to set them.
Audit log creation records every query at the point of classification: the query text, the sensitivity tier applied, the routing decision made, the model used, and the source documents cited in the response. This log answers the compliance question that most organizations cannot currently answer: "What data has our AI processed, under what controls, and who asked about it?"
During their first Stralevo compliance review, finance teams consistently report the same surprise: the classification logs reveal patterns nobody had been tracking. Which query types are most frequent. Which users ask the most sensitive queries. Which routing decisions were made and why. This is not surveillance — it is the operational record of a system where AI decisions are documented rather than invisible.
---
The Design Problem It Solves
Between the CFO's need to ask any financial question freely and the compliance requirement to handle sensitive financial data with appropriate controls is a genuine design problem. The naive solution is to restrict what questions can be asked. Restrict salary queries. Restrict contract analysis. Build approval workflows for sensitive data requests.
The correct solution is to classify what data the question touches and apply appropriate controls per query — automatically, without restricting the question. The CFO asks "how does our headcount cost compare to budget?" and the classification layer routes payroll data through SL2-compliant infrastructure, logs the query, cites the source documents, and delivers the answer. No approval request. No manual sensitivity selection. No reminder that sensitive data requires different handling.
In practice, a single natural-language question like "prepare the quarterly board pack comparing supplier costs to the same period last year, flagging any contract compliance issues" triggers multiple classification steps — retrieval from the Knowledge Graph for cost data, cross-reference against contracted terms, generation of a formatted report via VibeFlow, sensitivity assessment for any data including restricted commercial terms. All of this routing and classification happens in the background. One question. One board pack. Appropriately handled throughout.
---
The Compliance Record That Builds Automatically
Which of your employees asked your AI about your tax strategy last month, and where was that query processed? For most organizations running general-purpose AI tools, that question has no answer. There are no classification logs, no routing records, no audit trail for AI queries by type or sensitivity.
Stralevo's classification layer creates that record automatically for every query, every day. The compliance team can produce the full query classification history for any time period, for any user, showing sensitivity tiers applied and routing decisions made. When an EU AI Act audit requires documentation of how high-risk AI decisions were made and what oversight controls were in place, the classification logs are the audit evidence.
The finance function where every question is answered appropriately — a supplier price search in seconds through the right engine, a payroll analysis through compliant EU infrastructure with a full audit trail, a board report formatted to presentation standards without prompt engineering — is not a future architecture. It is what the classification layer enables today, for every CFO who asks a financial question and wants to know the answer was handled the right way.