No Idea If Your AI Is Right? KNOWIDEA's New Research Lets Finance Check Every Answer.

New York, Oct 12, 2026 — Two AI-written queries ask the same database the same question and return the same answer. One of them is wrong, and nobody can tell. Then a single missing value appears in the data, and the wrong one quietly drops accounts from a reconciliation report, with no error message and no warning.

That is the failure KNOWIDEA, the Toronto-based decision intelligence company, set out to catch. Today it published "VeriPIE: Checkable Query Plans and Evidence-Guided Selection for Text-to-SQL," research co-authored with Georgia Tech and developed with researchers at one of the world's largest asset managers. The paper shows how to make AI-generated database queries more accurate and, more importantly, checkable: every answer comes with evidence a risk team can examine and challenge before it informs a decision. The paper is available at knowidea.dev/papers/veripie-text2sql.pdf.

Banks, asset managers and corporate finance teams are adopting AI tools that turn plain-English questions into database queries, a task known as text-to-SQL. Ask for "accounts with no matching payment," and the AI writes the code that pulls them.

The danger is not the query that crashes. It is the query that runs, returns a reasonable number and is wrong. The paper's own example: two common ways to write "no matching payment" return identical results on today's data. Add one payment record with a missing account ID, and one version silently excludes every unmatched account while the other still returns them. The report looks complete. It isn't.

In finance, that kind of error does not stay in a spreadsheet. A wrong query, the paper notes, "can silently omit exposures, double-count transactions, or miscalculate financial aggregates while returning plausible results."

• In risk, an omitted exposure understates a portfolio's risk.
• In reporting, double-counted transactions misstate a regulatory filing.
• In operations, a reconciliation that drops unmatched accounts hides the very breaks it was built to find.

Most AI systems settle on the answer that most of their candidate queries agree on. The paper shows why that is not enough: queries can agree on today's data by coincidence and disagree tomorrow.

Financial institutions are expected to validate the models behind their decisions and put them through "effective challenge" by independent reviewers. Generative AI fits that discipline badly. A language model's reasoning cannot be reviewed line by line, and its answers can change from one run to the next.

Regulators have noticed. The Financial Stability Board has identified AI-related model risk, data quality and governance as financial stability vulnerabilities, and the EU AI Act sets risk-based requirements for AI systems. In April, U.S. bank regulators replaced their landmark model risk guidance, SR 11-7, with new guidance that explicitly leaves generative and agentic AI out of scope, leaving firms to build their own controls. In Canada, OSFI's Guideline E-23, which names AI models explicitly, takes effect in May 2027.

As the same AI tools spread across the industry, the same silent errors can repeat across many firms at once. That turns AI model risk from one firm's problem into a market-wide one.

VeriPIE treats an AI's answer as something to check, not something to trust. It works in five steps:

• Generate. The AI writes several candidate queries, some directly in SQL and some as step-by-step plans in QuerySketch, a plan language designed to be checked.
• Check. Each plan is checked against the database's actual tables, columns and data types by software built in Lean, a theorem prover increasingly used to verify AI-generated mathematics and critical software. Every query is also run.
• Repair. When a check fails, it explains why. The AI fixes the plan, and the fix is checked again.
• Group. Queries that return the same results are grouped, and a bounded search looks for data on which they would disagree, splitting apart queries that only agree by coincidence.
• Select. VeriPIE returns its ranked answers with the evidence behind them.

The result is a record a validator can work with: which checks passed, why a revision was needed and where candidate answers disagreed.

The authors tested VeriPIE on the complete development sets of Spider and BIRD, two widely used public benchmarks, using the same language model for every system, in single runs.

• On Spider, VeriPIE's first answer was correct for 745 of 1,034 questions (72.05%), 66 more than DeepEye (65.67%), the system it builds on, and slightly ahead of DIN-SQL (71.86%).
• Its two-answer shortlist covered 811 questions (78.43%), 39 more than DIN-SQL and 97 more than DeepEye.
• On BIRD, its first answer was correct 68.90% of the time, ahead of DeepEye's 68.45%, and its shortlist reached 73.73%.

The checking takes time, about 40 seconds per question on Spider. The authors argue that in financial analysis and review, a basis for accepting an answer that can be inspected matters alongside speed.

The authors are explicit about the limits. No checker can prove that a query matches what a person meant, and a person must still judge the business meaning. The evidence, they write, can "strengthen AI governance without establishing semantic correctness or regulatory compliance." VeriPIE is designed to plug into any text-to-SQL system that uses a language model, and future work will test it on financial reporting and regulatory workloads.

“Model risk management rests on a simple rule: you don't act on a number you can't check. AI quietly broke that rule. This research is a step toward restoring it, with AI that shows its work on every answer.”

“A language model can sound certain and still be wrong, and asking another model to check it only moves the doubt somewhere else. Formal verification is different. A check either passes or it fails, and anyone can run it again. That is the standard finance already holds its numbers to, and it is the one AI should meet.”

To read the paper, visit knowidea.dev/papers/veripie-text2sql.pdf. To see VeriPIE on your own data, bring one decision to a 30-minute session at knowidea.dev.

KNOWIDEA is a Toronto-based decision intelligence company. Its Verified Predictive Intelligence Engine turns company data and market evidence into predictions leaders can verify. Learn more at knowidea.dev.

Media contact: Yatharth Sejpal, info@knowidea.dev