Foresight - AI Financial Intelligence

Source of truth

Excel → queryable tables

Answers

Grounded, traceable

Client data training

None

Records leaving region

None

Finance for a multi-entity client lived across Excel workbooks with no single source of truth. Scheduling, invoices, P&L, and receivables sat in separate files with columns that did not agree, so every board question meant somebody rebuilding a spreadsheet by hand. The off-the-shelf tools meant to help were quietly picking the wrong line items and getting the arithmetic wrong - and a CFO cannot act on a number nobody can trace back to its source.

Rather than put AI in front of the mess, we rebuilt the foundation first. Raw workbooks land untouched so the original is always recoverable, an ETL pass normalises them into consistent tables, and two further layers hold progressively cleaner, modelled data with personal fields stripped on the way through. Only those clean layers can be queried. When a CFO asks a question, the model writes a query against known schemas, runs it, and answers from the rows that come back - so the answer is grounded in the data rather than recalled from training, and every figure traces to a table. Nothing is trained on client data, the model runs inside the client's own region so records never leave it, and a person reviews before any number is treated as final.

• Fix the data before adding the AI

  The instinct is to point a model at the spreadsheets and demo it. That produces answers nobody can check. Building the layered tables first meant every later answer had something solid underneath it.

• Strip personal data on the way in, not at question time

  Removing personal fields before they reach the queryable layers makes the AI structurally incapable of reading them. Filtering at question time would have left those columns one crafted prompt away from exposure.

• Let the model query, not remember

  The model writes a query against known schemas and answers from the rows returned. Nothing is trained on client data, so there is no secondary-use problem - and every number stays traceable to the table it came from.

• Keep AI advisory, with a person before anything material

  A wrong number that reaches a board paper is worse than no number at all. Human review and the ability to challenge an answer were designed in from the start rather than retrofitted after the first audit.

• Four layers, each one cleaner than the last, so raw personal data and analytics-ready data never share a blast radius
• Personal fields stripped before data reaches the queryable layers - the AI cannot read what was never carried forward
• Queries are scoped to the clean layers only; the raw landing zone is never reachable from the question path
• The model runs inside the client's own region, so financial records never leave their account
• Connector-based access over OAuth (Xero, MYOB, Zoho, Salesforce) - fetch on demand rather than copy wholesale
• Encryption at rest and in transit, credentials in a managed secret store, and an audit trail on every data access
• Wrote the platform's data-privacy, security, and AI-usage policy - the control set the build is measured against

• Scattered workbooks replaced by consistent tables anyone on the team can query
• Finance questions answered in plain language, without hand-written SQL
• Every AI-surfaced figure traceable to an underlying table and reviewable by a person
• Privacy, security, and AI-usage controls documented and mapped to the stack before pilot go-live

• Designed a layered finance data platform on AWS, not a notebook prototype
• Grounded AI answers where accuracy is a compliance obligation, not a nice-to-have
• Authored the data-privacy and AI-usage policy the whole team builds against
• Least-privilege access, encryption, and audit logging treated as product requirements