How a Finance Team Cut Month-End Close 10 Days to 1
A finance team rebuilt month-end close from scratch every month. A Claude agent now drafts variance commentary and routes exceptions — collapsing close from 10 days to 1.
90%
Cut from month-end close cycle
The Challenge
Month-end close at most finance functions is a manual, people-intensive process with no memory, no self-correction, and no leverage. Each close starts from scratch: pulling the same reports, chasing the same reconciliations, answering the same variance questions. The controller carries the entire process in their head, which makes every close dependent on one person's bandwidth and creates no institutional learning. QuantFi ran this problem on itself before building for clients.
What They Built
QuantFi built a four-layer agentic close system using Claude as the orchestration model. The architecture consists of: (1) a knowledge base layer that encodes the company's chart of accounts, entity structure, accounting policies, and prior-period close logic; (2) an integration layer connected to QuickBooks Online and relevant source systems; (3) an agent execution layer that runs the close workflow autonomously, pulling trial balance data, identifying variances against prior periods, drafting explanations, and flagging anomalies that require human review; and (4) a human gate layer where the controller or CFO reviews, approves, or overrides agent outputs before anything is finalized. The system improves each cycle: approved explanations and decisions are written back into the knowledge base, so the next close starts with more institutional memory than the last. The result is a close process that runs faster, documents its own work, and gets more accurate over time.
QuantFi built a four-layer agentic system on top of its existing QuickBooks Online stack: a knowledge base encoding the chart of accounts, prior-period logic, and company accounting policies; an integration layer pulling live transaction data; an agent execution layer running the close workflow autonomously (journal entry drafting, variance detection, flux commentary); and a human gate layer where the controller reviews and approves before anything is finalized.
What specifically changed: the close no longer depends on a single person remembering why last month's accruals were structured a certain way. Every approved decision writes back to the knowledge base, so the system compounds across cycles. The result is that QuantFi compresses a multi-day, controller-dependent close into a same-day agent-driven process, with the human reviewing outputs rather than producing them from scratch.
QuantFi treated its own month-end close as the first engagement, on the principle that you cannot automate judgment you have not first codified. Phase one built the knowledge base before any agent logic — the company's chart of accounts, entity structure, accounting policies, and prior-period close decisions were encoded into a structured, queryable layer. With that foundation in place, the team wired up an integration layer to QuickBooks Online and mapped the trial balance to the entity structure with logic for detecting period-over-period variances at the line-item level. The agent execution layer was built on Claude as the orchestration model: it pulls the trial balance, identifies variances against prior periods, drafts journal entries and variance explanations, and flags anomalies that require human review. A human gate layer is deliberately retained — for auditability and SOX-adjacent control, the agent never finalises anything without controller or CFO approval. Each approved decision is written back into the knowledge base, so accuracy and autonomy compound across cycles rather than resetting every month. The full build ran in under four weeks, and the month-end close dropped from ten days to one.
AI Role
The model reads the general ledger, prior-period financials, and company accounting policies, then autonomously drafts journal entries, detects and explains variances versus prior periods, generates flux commentary for each account, and routes flagged exceptions to the controller for review. This replaces the produce-from-scratch work that previously required a human to hold all that context in their head.
Infrastructure
QuickBooks Online (source of truth for general ledger and trial balance data) • Internal knowledge base layer encoding chart of accounts, entity structure, accounting policies, and prior-period close logic • Hostinger and Google Cloud (hosting and compute environment) • Hermes (agent orchestration framework running Claude)
Integration Points
QBO trial balance pulled and mapped to the entity structure for line-item variance detection • Period-over-period variance detection across the general ledger with explanations drafted by the agent • Knowledge base read/write loop where each approved close decision persists across cycles • Human-gate handoff routing flagged exceptions to controller/CFO for review and approval
Impact
Variance commentary drafted in minutes, not hours.
The agent produces first-draft explanations for every line-item movement exceeding threshold, eliminating the most time-intensive manual step in a typical close.
Institutional memory compounds each cycle.
Each approved close decision is written back into the knowledge base, so the system's accuracy and autonomy increase over time rather than resetting to zero each month.
Close cycle time cut by 90% — from 10 days to 1 day
Speed of data gathering, human approval, and execution enabled faster decisions and faster close time.
Implementation Complexity
The foundational unlock was building the knowledge base artifact before writing a single line of agent logic. Without encoding the company's accounting policies, chart of accounts, and close decision tree into a structured knowledge layer, the agent would have no context to reason against. Integration with QBO required mapping the trial balance output to the entity structure and building logic to detect period-over-period variances at the line-item level. The human gate architecture was intentional, not a limitation: for auditability and SOX-adjacent control purposes, the agent never finalizes anything without a human approval step.
Best Fit For
PE-backed portcos and middle-market CFOs who run a monthly close with a lean team and want to eliminate the manual, repetitive steps without replacing their controller. Specifically relevant for companies that have adopted cloud accounting software (QBO, Sage Intacct) but have not rebuilt any of the workflows on top of it.
