Notes

what's actually working in agentic finance

8 operators. 90 minutes. Off the record.

By Samer Azar, Fractional CFO · 2026-05-23 · 2 min read

Hi Reader,

For the last couple of months, my team and I have been heads-down building this. An autonomous AI brain for a CFO's workflow. The kind of system that ingests, categorizes, drafts, flags, and learns.

I'm at a point where I want to share what we've found with peers who are building in the same direction.

Operators, not commentators.

People shipping AI into real financial workflows, comparing notes on what's actually working and what's hype.

The questions I want to put on the table:

- LLM choice. Which models hold up under finance workloads, and which collapse the moment context gets dense or numbers get specific.

- Knowledge ingestion. What works for getting messy financial data, multi-entity GLs, prior-period commentary, board decks into an LLM in a way it can reason over.

- Data structuring for parsing. How to shape inputs so the model doesn't drift, doesn't pick the wrong subsidiary, doesn't invent a number to fill a gap.

- Anti-hallucination architecture. How to build the chatbot structurally so it never hallucinates. The hardest one. The one most teams are still solving by hand.

Behind all four sits the bigger question: where is the value actually being created in agentic finance, and how do we move forward with the wave that's clearly unfolding.

These are conversations I felt belonged among peers building this directly, not on LinkedIn, not in a webinar.

So I'm launching a series of small roundtables. Eight people per room. Founders, finance leads, builders who've actually shipped AI into a financial workflow in the last six months.

The first one is in Montreal, this Thursday May 28. 2 seats remaining.

Reply if you're there.

Paris is next: June 9, Station F, 18:30 to 20:00. Off the record. 90 minutes. Other cities follow.

If you're in Paris or commute-distance, and you're building agentic finance for real, this is the room.