Case Study

An Agentic AI Finance System Built for Confidential Client Financials

AIM's fractional CFO practice needed an AI system that can answer questions across several clients' ledgers, filings and models without ever mixing one client's data with another's, and without producing a number it cannot prove. We are building it to run offline, with verification and access control in the architecture rather than in the prompt.

Agentic AI · In progress · By Nathaniel Rub, Founder · Published September 19, 2026

The context

AIM (AIManagement Inc.) serves more than one finance client at a time. Each client’s ledger, contracts and models are confidential, and every figure that reaches a board has to trace back to a source. Off the shelf AI assistants failed both tests.

The challenge

Three problems stall most finance AI projects. Language models are unreliable at arithmetic and at writing raw SQL against messy ledgers. Client data has to be isolated at the database layer, not by asking the model nicely. And a model that answers confidently without showing its source is useless for advice a client will act on.

What we designed and are building

A gap analysis came first. It ranked every missing piece of a production finance system and named the pieces not worth building at all. The build follows a 30, 60 and 90 day plan.

Where AI fit

Agents plan and run the queries, gather the evidence and draft the analysis. A separate verification layer decides what is allowed to be shown.

Acceptance targets

MetricTarget
Answers with a traceable citation or formula100%
Cross client data exposure in testing0
Ledger writes without human approval0
Works with no internet connectionYes

This is AIM’s own build and it is still in progress. These are the acceptance targets the pilot will be measured against, not results.

Why it matters for finance leaders

The question is no longer whether AI can read your financials. It is whether it can prove where every number came from and keep one client's data away from another's. That is an engineering problem, and it has an answer.

FAQ

Can AI be trusted with financial data?

Only when isolation, verification and access control are built into the system instead of requested in a prompt. Those three layers are the core of this build.

Does it connect to QuickBooks?

Yes, through Intuit's official MCP server, read only by default. Other ledgers can be added behind the same semantic layer.

Why local models?

Client financials never leave the practice's hardware, and the system keeps working on a client site with no connectivity.

Related: Finance Transformation · FP&A and Excel · Finance transformation for a global managed services provider. See the rest of the work on the case studies index.

Work with AIM

Want AI analysis your auditors would sign off on?

Bring one workflow, its owner, and what an incorrect result would cost. That is enough to scope the first engagement.

Request a Consultation Agentic AI Solutions