Kayley Perry, a CPA who runs Firm Five, posted on LinkedIn this month about a milestone in her practice. The initial month-end close analysis for one of her clients, a $20 million manufacturing business running NetSuite, would be completed primarily by Claude.
She then listed exactly what Claude would be doing.
What Claude is doing
The tasks Kayley has handed to Claude are concrete. They are the first-pass jobs every senior bookkeeper or accountant does at the start of a close, the ones that are mostly review and pattern-matching against a known ledger.
Claude is reviewing expense categorisation, looking at the supplier and GL account level for transactions that have been miscoded. It is reviewing the list of open purchase orders and drafting follow-up emails to the purchasing team and to vendors. It is doing the same on open sales orders, drafting emails to the sales team to chase what is still outstanding. It is updating the depreciation schedule for additions and disposals, and posting depreciation. It is assessing accrued purchases to identify the impact on accounts payable and inventory, flagging where item fulfilments need updating. It is identifying unapplied customer deposits. It is reviewing accounts receivable and drafting an email to the sales team summarising new and overdue receivables.
That is the work that fills the first stretch of a close, before any judgement-heavy review can start. Now it is being prepared before the accountant sits down.
The next step in Kayley's pilot, as she described it, is moving from execution to automated execution. Same workflows, with the manual extract-paste-prompt loop removed.
Why this works for manufacturing on NetSuite
A $20 million manufacturer running NetSuite is the kind of file where the close is genuinely complex but where the data is also structured enough for AI to handle the first pass usefully.
The GR/IR account, the reconciliation between received goods and supplier invoices, is one of the more painful parts of a manufacturing close. Accrued purchases, item fulfilments not yet billed, and the matching of purchase orders to receipts and invoices generate the kind of cross-reference work that takes time and concentration but follows a predictable pattern. That pattern-matching is what current AI does well, provided it has access to the right data.
Depreciation schedules are similar. The mechanics are repetitive. The judgement calls are limited. The cost is mostly time. Handing the first pass to an AI agent and reviewing the output is a sensible split.
AR and PO email drafting is the third clear fit. Drafting twenty follow-up emails individually is a low-value use of a controller's time. Reviewing twenty drafts and sending them is faster.
Where this breaks, according to people doing it elsewhere
One of the most useful threads under Kayley's post came from a practitioner running similar Claude workflows in QuickBooks Online and Xero on smaller clients. He had specific failure modes to flag.
Expense categorisation looks like it works on day one. The trap is ambiguous vendors. Amazon, Costco, generic merchant feeds where the vendor name itself is not the signal. Claude pattern-matches to the most common historical code for that vendor instead of the right code for that specific transaction. The fix is feeding it the last twelve months of how the vendor was coded alongside the bank feed description. That gives it the context to pick up the nuance.
AR follow-up emails draft well in isolation. They read generic when the AI does not know that customer X always pays at day 47, or that customer Y is on a payment plan. Same fix. Feed it the customer payment history before it drafts, and the tone calibrates.
The bigger structural point came from a NetSuite specialist. Manual extract-paste-prompt workflows hit a ceiling fast on scale and consistency. For a manufacturing company on NetSuite, the accrued purchases step is where that ceiling shows up. Reviewing open receipts against POs across item fulfilments is painful to export cleanly and easy to miss in a paste workflow. NetSuite shipped pre-configured MCP roles for AP, AR, and Controller workflows last month as part of the AI Connector update. That gives Claude direct access to transaction records rather than depending on manual data pulls.
A more philosophical pushback ran underneath the thread. LLMs are probabilistic by design. They produce errors. Wrapping them in deterministic automation, where the AI handles the language and structured logic handles the calculations, is the version of this work most likely to hold up at scale. A fully autonomous AI controller, where an LLM makes the call on every transaction unsupervised, is beyond what current technology supports reliably.
What this means for firms
The use case is already specific. The work being handed to AI is not "the close." It is the first-pass review work inside the close. Expense classification review, accrued purchase reconciliation, depreciation posting, AR and PO follow-up email drafting. Firms thinking about where to start should look at these specific tasks rather than at the close as a whole.
Data access is the next constraint. Manual extract-paste workflows hold up well enough to prove a concept on one client. They break down across a portfolio. Firms that want to move past pilot need to look at how AI agents connect directly to the ledger, whether through MCP integrations like the NetSuite AI Connector update or the equivalent in other accounting platforms. The infrastructure question matters more than the prompt question.
The human review step stays in place. Kayley's post is careful to describe Claude as completing the initial analysis. The accountant still reviews, catches the ambiguous-vendor traps, calibrates the AR emails against customer history, and signs off. That review work is where the experienced practitioner earns their margin. The first-pass execution is what gets cheaper. The judgement layer above it is what gets more valuable.
For firms running ANZ manufacturing clients on NetSuite or similar mid-market platforms, the workflow Kayley has described is replicable. The known failure modes have been documented by people already doing it. The infrastructure is improving. The question is not whether to start. The question is which client file to start with and which two or three tasks to pilot first.
Kayley Perry's LinkedIn post sets out the full list of tasks she has handed to Claude for the close, and the comment thread underneath contains substantive detail from practitioners running similar workflows.
