What questions should a bank ask before piloting financial intelligence technology?

Before piloting financial intelligence technology, a bank should be able to get clear answers to questions in five areas: how the forecasts are produced, whether they can be explained, what data is used and under what boundaries, how the pilot itself will be evaluated, and what is actually built today. A provider that welcomes these questions is signalling how it will behave throughout the relationship. A provider that deflects them is also signalling — believe it.

Methodology questions

• What data does the forecast actually use, and what does it deliberately not use?
• How are recurring income and obligations identified? What happens when a pattern is ambiguous?
• How is variable income modelled — as an average, a range, or a distribution?
• Over what horizon does the forecast run, and how does its reliability change across that horizon?
• What does the system do when data is missing or stale — proceed silently, degrade visibly, or refuse?

The answers reveal whether the methodology was designed for real household data or for a demonstration dataset.

Explainability questions

• Pick any output from a live demonstration: can the provider decompose it into the income, obligations, and dates that produced it?
• Are explanations generated from the same reasoning that produced the forecast, or written around it afterward?
• Can a front-line staff member explain this output to a customer in one minute? Ask the provider to demonstrate exactly that.
• Where do assumptions appear — in the product, or only in documentation?

Data and privacy questions

• What is the minimum data scope the capability needs to function? (Smaller is a design virtue.)
• Is personal financial data ever sold, shared for advertising, or used for profiling? Get the answer in writing, in the agreement.
• How is data separated between institutional clients, and what access controls and encryption apply?
• How does the capability operate inside our consent framework and our jurisdiction's regulatory requirements?
• Who reviews this — and will the provider engage our privacy and security teams directly, early, as part of the pilot design?

Evaluation-design questions

• What will the pilot measure, precisely? Forecast accuracy against known outcomes is measurable; "customer impact" needs definition before anyone claims it.
• On what data will the forecast be tested — historical backtesting, a limited live cohort, or both?
• What are the decision points, and what evidence ends the pilot, extends it, or stops it?
• What will the provider not claim during the pilot, however it goes? (A provider with claims discipline will have an answer ready.)

Build-status questions

• What is implemented today, what is in development, and what is conceptual? Ask for the answer in those three buckets, in writing.
• Which parts of the demonstration we saw are production behaviour, and which were illustrative?
• What does the integration actually require from our side — data feeds, engineering time, review cycles?

These questions are not adversarial. They are how a bank protects its customers and its own credibility, and any serious provider has been asked them before.

A final test

Notice how the provider treats uncertainty during the sales process itself. Forecasts are, in the end, a product about honesty regarding the future. A provider that is precise about its own status — what exists, what is being built, what is unknown — is demonstrating the exact quality its product must have. PFSI is in pre-launch development and engages design partners on those terms; this blog's evaluation questions are ones PFSI expects to be asked, and answers in the open.

Key takeaways

• Interrogate methodology: data used, recurrence detection, variable-income modelling, missing-data behaviour.
• Test explainability live: decompose an output, in the room, in staff-and-customer language.
• Pin down privacy boundaries in writing: minimum scope, no sale or advertising use, institutional separation.
• Design the pilot's evaluation before it starts — measures, data, decision points, and prohibited claims.
• Require build status in three buckets: implemented, in development, conceptual.

Banks preparing a pilot evaluation can put these questions to PFSI directly through the [Partnerships page](/partnerships).

---

About PFSI: PFSI (Personal Financial Stability Intelligence) is the financial intelligence program in development at Monimate Technologies Inc. New to PFSI? Start with [What Is PFSI?](/blog/what-is-pfsi).