On July 28, Juniper Square launched Fay, an Admin Oversight Agent that runs 150+ accuracy and consistency checks across a fund's close pack — financial statements, PCAPs, workbooks — and surfaces errors before they reach an LP or an auditor.

It's a good product announcement. Not because of the AI, but because of three things they said about the AI.

The three sentences that matter

"The reasoning comes from AI. The math runs in code." The LLM decides what to look at; deterministic code computes whether it ties. The answer never drifts between runs. This is the correct architecture for financial work, and it's worth saying plainly: any vendor whose numbers come out of the language model, rather than out of code the language model orchestrates, is selling you a liability.

A countable check library. "150+ checks" is a claim you can audit. "AI reviews your financials" is not. Enumerable checks mean you can ask which ones ran, which failed, and why — the same standard you'd hold a junior accountant to.

A published eval program. They re-test their checks against documents with verified-correct answers whenever the underlying models change. Model drift is real; an eval program is the difference between a controlled process and a vibe.

If you're evaluating any AI for your back office — including ours — those are the three questions to ask. Show me the check library. Show me where the math executes. Show me the eval program. A vendor who can't answer all three hasn't built for finance.

We hold ourselves to the same bar. Fundry's agents use the model for extraction and drafting; every reconciliation — K-1 ties, capital account rollforwards, fee tie-outs, waterfall tiers — runs as deterministic code. Our extractors are tested against golden document sets with verified answers, re-run when models change. This isn't a differentiator anymore. As of July 28, it's table stakes, and that's good for everyone.

The question Fay doesn't answer

Fay checks work that someone else produced. That's genuinely valuable if you're a GP with a large third-party administrator generating close packs at scale — which is Juniper Square's world: enterprise GPs, billions in administered assets, a fund operating system underneath.

Under $1B AUM, the situation is different. There often is no fund admin producing a polished close pack to oversee — or there's one admin, one controller, and you. The DDQ answers, the LP emails, the K-1 reconciliation, the management company budget: the checking problem and the doing problem are the same person's Tuesday.

An oversight agent supervising work you didn't have capacity to produce in the first place solves the second half of a problem you never got to the first half of.

That's the wedge we build in. Fundry's eight agents draft the DDQ from your own documents with citations, triage the LP inbox, reconcile the K-1 batch, and build the budget-versus-actual — and run the deterministic checks on their own output before a human reviews any of it. Do the work, then check the work, then queue it for your approval. For a lean fund, production and oversight can't be separate purchases.

One more structural note: an oversight layer is most useful when it's independent of the thing it oversees. Fay reviewing close packs produced by Juniper Square's own fund administration business is a reasonable convenience. It's also a reminder that the checker and the maker sitting in different vendors is a feature, not a bug.

The agentic era arrived in fund operations this year — Juniper Square just confirmed it at enterprise scale. If you run a fund below the scale their proof points assume, the same architecture is available to you. It just has to do the work, not only grade it.