For the past several months, we have been running our own month-end close in a similar way we advise our clients to run theirs.
Not with better prompts. Not with a Claude tab open next to a spreadsheet. With a purpose-built agent embedded in the actual workflow, pulling data from live systems, running calculations, staging outputs, and flagging the handful of things that require a human to make a real decision.
Most finance leaders we talk to can picture AI helping with finance. They can't picture what an agentic close actually looks like end-to-end. Here is what it looks like on ours.
---
The data layer
Every close starts the same way it always has: the data. Time logs, bank activity, collections, the general ledger. The difference is that no one has to go collect it. The agent pulls from each source system at the start of the cycle, cross-references for completeness, and surfaces anything that looks inconsistent before a human ever touches it.
The first two hours of old-close, logging into four systems, exporting files, reconciling formats, praying nothing changed since last month, are gone. That is not a small thing. That time, repeated across twelve cycles, is the reason closes feel like a grind even when nothing goes wrong.
---
The memory layer
This is the part most implementations miss, and the reason most agentic close pilots produce output no one trusts.
Before the agent calculates anything, it consults what we call a standing memory layer: a structured, version-controlled record of billing rates, contractor pay rates, client-specific rules, and every exception that has ever been documented. Rate changes, special agreements, one-time adjustments. None of it lives in someone's head or in a Slack message from three months ago that may or may not be findable.
The agent never guesses. It never carries a stale rate into a calculation. If something in the memory layer needs to be updated, that is a human decision made explicitly, before the calculation runs, and logged. Not discovered after the close when the numbers are already wrong.
This is what makes the output auditable. It happens to look like a document. What it actually is, is a governance layer with a clear chain of custody.
---
The calculation and staging layer
Once the data is pulled and the memory layer is locked, the agent runs the work. Billing calculations. AR reconciliation. AP summary. Journal entry drafts.
What comes out is a fully staged close workbook, every figure sourced, every journal entry drafted and labeled before a human reviews it. The mechanical work that used to consume a week of a controller's calendar lives in the workbook. The controller's job becomes reviewing it, not building it.
We want to be precise about what this is and what it is not. The agent is not making judgment calls. It is not deciding whether to invoice a client, write off a balance, or classify an unusual expense. Those decisions stay with humans, at four defined gates where the workflow stops and waits for explicit sign-off before it continues.
Rate lock. Data review. Billing decisions. Final approval.
The agent runs between the gates. Humans own the gates. The gates are what make this trustworthy, not just fast.
---
The output
What comes out the other end: financial statements ready for review, contractor confirmations, posted journal entries, and a margin summary by client. Every figure has a source. Every exception was flagged, documented, and resolved before the books closed.
\When something is wrong, and occasionally something is wrong, it surfaces at a gate, not after the fact. That is the design. The agent is not a black box that hands you finished financials and asks you to trust it. It is a structured workflow with defined handoff points, and the handoffs are where the judgment lives.
That is not a demo. That is our actual close.
---
What this means for a mid-market CFO
The systems involved, accounting software, a bank account, a time-tracking tool, a collections platform, are the same systems sitting in most mid-market finance stacks right now. Nothing we built required custom infrastructure or a seven-figure implementation. It required doing the harder, less glamorous work: mapping the workflow precisely, building the memory layer so the agent has rules to run against, and deciding exactly where human judgment belongs.
That last part is the real work. Most teams trying to automate their close skip it. They connect the tools, watch the agent produce something, and then spend the next three closes cleaning up the exceptions the agent didn't know how to handle because no one defined how to handle them.
The technology is not the bottleneck. The workflow design is.
If you are 6 to 12 months from wanting a close that runs like this, the work starts with one conversation: your workflow, mapped. Not a pitch. Not a demo. A working session where we look at what you have, where the friction is, and what a version of this looks like on your balance sheet.
That is where every engagement starts. It is also where most of the value gets created.
THE NUMBER
That is what Zapier's finance team did to their monthly journal entry review time after deploying an AI compliance monitor inside their accounting stack. A workflow that used to require a controller's full attention for the better part of two business days now closes in under two hours, with higher accuracy and an automated audit trail.
Their CFO Ryan Roccon shared the full stack live this week: an 8-person team running accounting for a $5B company across a set of AI-native workflows, each one named to the person who built it.
Next issue we are going inside exactly how they did it, workflow by workflow.
THIS WEEK IN AI-NATIVE FINANCE
Anthropic and Blackstone close their first acquisition.
The $1.5B joint venture between Anthropic, Blackstone, and Hellman and Friedman acquired San Francisco-based Fractional AI last week and is now forward-deploying engineers inside PE portfolio companies. The implementation layer for AI in finance has an institutional house brand. It is aimed at the same mid-market and PE-backed businesses that read this newsletter.
Blackstone and Google launch a dedicated TPU cloud.
A joint venture to build AI compute infrastructure using Google's Tensor Processing Units. Blackstone is now invested in the AI software layer, the AI application layer, and the AI infrastructure layer simultaneously. The largest alternative asset manager in the world is making a full-stack bet.
Google AI token volume hits 3.2 quadrillion per month.
Up from 480 trillion in May 2025 and 9.7 trillion in May 2024. Seven times year-over-year growth. The compute demand underlying every AI tool your finance team uses is growing faster than most CFOs have modeled into their technology budgets. The token bill is not a rounding error. It is becoming a P&L line.
SAP names Anthropic Claude the reasoning layer for its ERP.
Claude is now the primary AI model for Joule, SAP's agent framework across S/4HANA, Ariba, and SuccessFactors. It covers close-the-books workflows, AP approval chains, and supplier rerouting for hundreds of thousands of enterprise customers. MCP handles cross-system coordination for non-SAP tools.
WHAT WE'RE READING
The X thread announcing the cut is worth reading for the org design framing, not the number. Evans introduces the "100x org" thesis and is explicit that this is a structural redesign around agents. The $1M salary bands for remaining roles tell the real story about what high-output humans are worth in an agent-native company.
The company that builds QuickBooks and TurboTax cut 17% of its workforce and attributed it directly to AI integration. Read it as a signal about where the tools in your finance stack are heading, not just as a workforce story.
The structural argument for why AI makes expert demand go up, not down. The counterweight every finance leader should read alongside the two memos above before drawing conclusions about their own team.
Christian Sanford and Kenny Jen are the co-founders and managing partners of QuantFi, where they help PE- and VC-backed companies build investor-grade finance functions powered by AI-native infrastructure.
Christian came up through investment banking at Barclays ($20B+ in M&A and capital markets), buy-side investing at a hedge fund, and fractional CFO work for investor-backed companies. BBA and MSA in Accounting from Texas Tech.
Kenny led financial strategy at Pilot CFO, Pure Beauty, and Emil Capital Partners after starting his career at Credit Suisse. He has overseen finance for consumer, SaaS, and manufacturing businesses, and works directly with founders on pricing, working capital, and capital allocation. BSBA in Finance from Georgetown.