The shape of the system

Each agent has a name, a scope, a schedule, and a written runbook. None of them ship work directly to the books — every one writes proposed entries to a staging table that a CPA reviews on a fixed cadence. The agent’s job is to do the rote work so the human can spend their time on the part that requires judgement. The arrangement matters: if the agents are ever in a position where their output goes straight to a production ledger without a CPA seeing it, the firm has lost the thing that makes the engagement worth paying for.

The agents are:

• categorize — assigns a GL account and class to every incoming transaction, tuned to your chart of accounts. On a new chart, the first weeks are a warm-up while the model learns it.
• reconcile — three-way match across the bank statement, the processor settlement file, and the GL. Surfaces mismatches as a daily review queue.
• cashflow — pulls AR, AP, payroll, recurring revenue, and seasonal patterns to produce a 13-week rolling forecast. Refreshed weekly with operator notes attached to anything anomalous.
• anomaly — diffs the current period against a rolling baseline at the account level. Flags whatever runs meaningfully off the trend, with the suspected driver attached.
• sales-tax — collects taxable-transaction data from the relevant sources, normalizes against jurisdiction rules, and prepares filings + remittance for the CPA to review and file.

Tuned to your chart, not a generic model

Categorization accuracy is one of the things every automation vendor claims and nobody benchmarks honestly. The honest version is conditional: the agent gets reliable once it has seen a chart of accounts for a while and has enough operating history to learn its patterns. We measure it — we log every disagreement between the agent and the CPA and track the rate over time, so the claim is something we watch rather than something we assert.

The misses aren’t random. They cluster around three patterns: ambiguous vendor names where the memo could plausibly be one of two accounts (a Stripe payout that could be sales or a refund reversal), transactions where the right answer depends on context only the operator has (a wire labeled “consulting fees” that’s actually a refund), and new vendors with no prior history. The first two require a CPA. The third resolves itself within a billing cycle.

When we onboard a new chart, the first weeks run at a lower confidence threshold — the agent flags more transactions for review, learns the chart, and the review queue contracts week over week until it settles.

CPA-in-the-loop, not as marketing

Every agent the firm runs has a human review step. This isn’t a regulatory hedge or a marketing posture — it’s the operating model. The reason is direct: the financial statements are an attestation about a business. The CPA is the person attesting. If the attestation is wrong, that’s a problem for the CPA, not the model. So the CPA reviews the edge cases, signs off on the close, and owns the answer.

What this means in practice: the agents save the work that’s expensive in person-hours and cheap in judgement — the categorization, the matching, the variance flagging — and leave the work that’s cheap in person-hours and expensive in judgement to the CPA. The work that used to consume a senior accountant’s days happens overnight, and the CPA spends the recovered time on the edge cases the agents flag, plus the work that wasn’t getting done because the close consumed the calendar.

Standalone, or the engine for the whole operation

This line is sellable on its own. If you already have a finance team and the pain is manual work that shouldn’t be manual, we’ll implement AI and automation against their existing workflows — no requirement to hand us the books. It’s also the engine when we run the full operation: the same agents are what let one firm keep a clean close across §3–§6 without a back office full of people.

What we don’t do is ship the agents as standalone software. They stay proprietary to the firm — implemented, tuned, and maintained as part of the engagement, not licensed as a product. Selling them as software would change the model into something neither side wants.

What the operator sees

A few things, on a few cadences. Daily: a one-line digest of what the agents did overnight, what they flagged, and what’s in the review queue. Weekly: the cashflow refresh and a brief written commentary on what changed. Monthly: anomaly summary attached to the close packet, plus the agent-accuracy log for the period. Quarterly: a review of which agents are pulling weight, which ones we’d retire if we could, and whether there’s a new pattern in the workflow that warrants a new agent.

What the operator doesn’t see: the prompts, the embedding tuning, the staging-table schemas, the failure-handling code. That’s the firm’s problem, not the operator’s, and the engagement is set up so the operator never has to learn any of it to get the benefit.