Agentic finance is bigger than expense management — it's the whole money chain

Two documents landed this year that, read together, tell you where agentic finance is actually going. The IMF’s How Agentic AI Will Reshape Payments maps the risks and layers from a policy seat. Hui Gong’s Agent-to-Agent Finance maps the infrastructure from a research seat. They come at it from opposite directions and converge on the same two conclusions: agents in finance are real and valuable, and their autonomy has to be bounded and governed, not maximised.

What neither is really about — and what most of the product market is stuck on — is scope. So let me make the argument I’ve been making inside Fynex since we started: the interesting thing about agentic finance isn’t automating expenses. It’s running the whole money chain.

Most “agentic finance” is agentic spend

Look closely at what today’s flagship agents actually touch, and it’s almost always the expense side. Auto-approve a low-risk expense. Audit spend against policy. Code the transaction into the GL. That’s real and useful work — I said as much in the honest Brex breakdown — but notice what it has in common: it’s all money that already left, on the vendor’s own card.

Expenses are the easiest slice to make agentic precisely because they’re the most contained. The card sets the rails, the policy sets the bounds, the money’s already gone. It’s a good place to start. It’s a small place to stop.

The whole money chain

Here’s the shape of the thing an operator actually lives inside, in order:

• Money in. Invoicing and collections — raising invoices, chasing the ones that go quiet, collecting from clients across borders and currencies.
• Money out. Payouts — paying suppliers, contractors and partners across borders, routed and priced, on whatever rail fits.
• Money tied together. Reconciliation — matching every payment to every invoice and PO, into your ledger, catching the mismatches and duplicates before they cost you.
• Money looking forward. Working capital and cash forecasting — how much cash you’ll have and when, and what that means for the next payout run.

And because the agents run all four stages on one reconciled record, the whole chain becomes something you can actually read — the real state of the business at a glance, continuously, instead of a month-end scramble to assemble it. That cross-cutting view is only possible when one system reasons over the entire chain rather than a slice of it.

Every one of these is heavy on exactly the work agents are good at: reading messy documents, matching records, chasing, forecasting, drafting. And every one of them is reversible reasoning right up until the single moment money moves. That’s the prize the IMF’s three-layer model points at and that the A2A paper’s six requirements are really about — not a smarter expense report, but an agent that reasons across the entire chain and prepares every decision in it.

And notice: the official framing already points here, not at expenses. The IMF’s own use-case list runs to agents that orchestrate “the entire cross-border payment chain” — initiation, routing across correspondent banks and rails, compliance checks, settlement monitoring — plus autonomous liquidity and treasury management and real-time agentic compliance. That’s the whole money chain, described by a central-bank note, as the place agentic AI actually reshapes payments. McKinsey’s phrase, quoted approvingly in the note, is “digital factories” of AI agents that handle entire tasks, with humans needed only for exceptions and oversight. Entire tasks — not the expense corner of them.

That’s what Fynex builds. Agents that run the full chain, on top of the accounts, cards and PSPs you already hold — not a replacement for your bank, a reasoning layer over it.

The two design questions that actually matter

Strip away the marketing and every agentic-finance product answers two questions. Both papers, in their own vocabulary, tell you these are the ones to ask.

1. Which part of the money chain do the agents run? Expense agents optimise money already spent — the contained corner. The larger value for most operators is the whole chain: invoices raised and chased, payouts routed cheaply, everything reconciled, cash forecast forward. When you evaluate a tool — us included — ask how much of the chain its agents actually reach, and be suspicious of “agentic finance” that turns out to mean “agentic expenses.” (Our comparisons exist to make that scope difference concrete.)

2. Where does the human sit relative to the money? This is the question both papers spend their pages on. The IMF frames it as the tension between probabilistic agents and deterministic settlement, and warns about herding, opacity and accountability gaps. Gong frames it as bounded autonomy — agency that doesn’t make markets “more opaque, fragile or unaccountable.” Same answer, two accents: the agent should reason freely and its authority over irreversible money should be bounded. Fynex draws that bound at settlement — agents prepare, the human approves what moves — because instant money is irrevocable and the approval line is where finality begins.

What I actually think happens next

I’ve spent seventeen years in payments, and I built a payments business to €2B across eight European markets before this one. So I’ll say plainly what I think the next few years look like, and where I’d bet.

Agent reasoning over the whole money chain is here and it’s going to become table stakes. The finance function stops being data entry and chasing and manual matching, because agents do that continuously and better. That part isn’t speculative — it’s shipping.

Fully autonomous money movement is going to arrive much more slowly and much more selectively than the demos suggest, because the papers are right: the risks are systemic and the liability questions are unresolved, and no one sensible wants to be the case study that settles them. The winners won’t be whoever removes the human fastest. They’ll be whoever gives agents the most reach over the reasoning and keeps the firmest, best-designed hand on the movement.

That’s the version of agentic finance we’re building — the whole money chain, run by agents, with a human on the one boundary that’s final. Not agentic expenses. Not autonomy for its own sake. Agents everywhere they help, bounded exactly where money stops being reversible. If you run a business where money comes in from clients and goes out to many parties — an agency, a marketplace, a studio — that whole chain is the thing worth making agentic. The corner isn’t.