Hard Problems Hide in Boring Places: Turning Accounting Workflows into AI Products

Accounting and B2B payments are often seen as boring, solved problems — until you try to apply AI to them. The moment a system can misread an invoice, suggest the wrong action, or leak sensitive financial data, “cool AI demos” turn into serious engineering challenges.

In this talk, I’ll share how we at Pliant build AI features in one of the most constrained domains possible, where correctness, trust, auditability, and permissions are non-negotiable. Instead of treating LLMs as smart oracles, we design them as untrusted components that propose actions, operate on structured data, and are constrained by strict policies and approval flows.

We’ll walk through concrete patterns for turning existing accounting workflows into real AI products: grounding models in financial data, using schemas instead of free text, enforcing authorization at the system level, and designing human-in-the-loop interactions that users actually trust. Along the way, I’ll share failure modes we hit in production and how we fixed them.

This talk is about where innovation really happens: not in flashy demos, but in making AI work reliably in the places where mistakes are expensive.