The short answer: bring six questions to any revenue recognition, billing, or "AI finance" demo. Provenance, determinism, agreement across systems, ASC 606 readiness, governance, and openness. If a tool can't trace it, repeat it, and defend it, an agent acting on its numbers still needs significant human oversight.

Most vendors in finance now (finally) use the same three words: deterministic, provenance, governance.

Less than a year ago the pitch was speed. Close faster, automate more, save hours. Now the pitch is control. This is a good shift for CFOs and for finance in general.

It also makes it harder to tell which vendor has built for this type of infrastructure and who simply updated their homepage and vibe-coded a new interface meant to show off governance features.

I've spent the last year in rooms with CFOs, controllers, and revenue accounting leaders asking one question: as agents start acting on finance data, what becomes the constraint, the models, the data, or the controls?

It's never the model.

So here's the checklist I'd bring to any demo right now. Six questions.

1. Provenance: Where did this number come from?

Not which system. Which document, which clause, which field. Guideflow's guide to revenue recognition software puts the test well: "verify that every recognized dollar traces back to a contract, an approval, and a journal entry." If a tool can't take you from a recognized revenue figure back to the exact place in the contract it came from, it likely wasn't built to deliver on provenance.

2. Determinism: Will I get the same answer tomorrow, or after lunch?

Ask where the model is allowed to guess and where it isn't. Reading a contract can be probabilistic. The check that decides whether the contract matches the invoice should not be. Same inputs, same answer, every time. And no, creating a skills.md file won't solve this on its own.

3. Agreement: Does it check across systems, or only inside one?

Most revenue tools automate inside their own walls. The leak lives between them: where systems and teams hand off work to each other. The CRM says one start date, the contract says another, billing runs on a third, and the GL records who knows what.

Ask whether the tool catches that before it hits billing or revenue recognition, or after close. And ask whether maintaining agreement requires manual system-to-system mapping.

4. ASC 606 readiness: Does it catch problems upstream?

Mapping performance obligations and allocating revenue is hard enough on its own; Alguna's roundup of SaaS rev rec tools is a good primer on why. But the harder problem sits before it. Rev rec accuracy starts as a data problem, not an accounting one.

Feed a flawless recognition engine mismatched contract and billing data and you get numbers that are fully traceable and still wrong.

5. Governance: Who decides what AI does and what stays human?

Ask where the policy lives, who sets it, and whether the system enforces it or a slide does. Wise's guide notes that audit trails exist to "show what changed, when it changed, and who made the change." Now add a fourth question: was it a person or an agent? (More on why that question is still open in Who's Testing the AI in Your Audit?)

6. Openness: Does it work with the agents you already have?

Your team will likely want to bring their own agents rather than rebuild from scratch. The question is:

  • A) can the system handle agents built outside of it, and
  • B) if so, can the agents you bring along read verified data, or only raw data?

A tool that only trusts its own agent isn't a foundation. It's another silo.

If you can't trace it, repeat it, and defend it, an agent still needs significant human oversight.

Where Meshly fits

These are the questions we've been building Meshly around from the beginning:

  • Meshly Validate extracts and checks contract terms before they hit billing or revenue recognition, with ASC 606 and policy checks.
  • Every value we extract links back to its exact place in the source.
  • Meshly Intelligence continuously identifies, explains, and evidences where business meaning changes or disagrees across systems.
  • Bring your own agent. It reads Meshly-verified data.

Where we're going: you set the policy for what AI can do and what needs a human. The guardrails are deterministic. On top of them, agents do the work. That's the trust harness.

If you're evaluating revenue software this quarter, bring these six questions to the demo. I'd love to hear what you learn: sandra@meshly.ai.