Clerical work
Gathering, formatting, routing, matching, chasing status. No judgment is involved and nothing is being decided. This is where automation pays for itself, and it is usually most of the volume.
AI and automation
I build the tooling around the accounting work I am responsible for. That means the automation has to survive an audit, which changes what you are willing to automate.
Most automation builders do not understand accounting controls, and most accountants do not build automation. I do both, which means the systems I build respect the controls they run inside.
In practice the difference shows up in where the automation stops. Someone approaching a bank reconciliation purely as an engineering problem will try to automate the whole thing, because from the outside it looks like a matching exercise. Someone who has had to explain that reconciliation to an auditor knows which decisions have to stay with a person, and builds the automation to hand those decisions over cleanly rather than absorb them.
Every process I have automated in finance has been split the same way, into three kinds of work.
Gathering, formatting, routing, matching, chasing status. No judgment is involved and nothing is being decided. This is where automation pays for itself, and it is usually most of the volume.
Work where a proposal can be generated with reasoning attached and a person accepts or rejects it. The automation drafts, the person decides. Useful as long as the reasoning is visible and rejecting is as easy as accepting.
Approvals, exception resolution, anything that creates an obligation or moves money. These stay with an authorized person. Automating a control point does not make a process faster, it makes it undefendable.
Getting that boundary right matters more than the tooling. An automation that quietly closes out exceptions is worse than no automation, because it removes the signal that something is wrong while appearing to improve throughput.
The work below came out of problems in my own close, not from looking for a reason to use a new tool.
An interactive close checklist where every task links to its SOP, so any team member can pick up an unfamiliar close task and complete it correctly.
Open live dashboard: Month-End Close Command CenterRead the case study: Month-End Close Command Center
Dashboards that track accounts payable, accounts receivable, and inventory KPIs so the team can see operational problems between closes rather than after them.
Open live dashboard: AI KPI Dashboards for AP, AR, and InventoryRead the case study: AI KPI Dashboards for AP, AR, and Inventory
Automation that handles the high volume, low judgment portion of bank reconciliations and routes anything ambiguous to a person for review.
Open live dashboard: Bank Reconciliation AutomationRead the case study: Bank Reconciliation Automation
Automation across the accounts payable cycle that speeds up routing and coding while leaving approval authority and segregation of duties intact.
A daily briefing assembled from several sources and delivered through Telegram each morning.
I use Claude, ChatGPT, and Codex daily to build finance tools, n8n for workflow automation on a self-hosted VPS, and Supabase and Next.js when something needs to be a real application rather than a workflow. For anything touching private data I run local models instead.
The same approach applies outside accounting, which is why I started Hawk Eye AI. Small businesses lose work to problems that are nobody’s fault: a call that came in while everyone was busy, a follow-up that did not happen because the next morning got away from them. Those are process problems, and they respond to the same treatment.
Reconciliation looks like a matching problem, which is why it gets over-automated. The work splits into three categories, and only two should run alone.
Automating AP is easy. Doing it without ending up with faster payments and weaker controls takes one discipline: separating clerical steps from controls.
Where Claude, ChatGPT, and n8n genuinely help in accounting work, where they do not belong, and the one rule that reliably tells the difference.
Most finance dashboards fail because they show too much. A small set of metrics that each point to an action beats a comprehensive one nobody opens.
Most automation advice is written by people selling automation. Here is where small business time actually goes, and which of it responds to automating.
The habits that make a month-end close reliable are the same habits that make any process reliable. Accounting is unusually good preparation for this.
Follow-up is the most commonly skipped task in small business. The design question is not how to send more messages, it is where automation should stop.
The questions worth answering before you build anything, and the four mistakes behind most automation projects that end up abandoned.
You cannot automate a process you have not defined. Most failed automation projects are undocumented processes discovering that expensively.
Four years, 20+ small business clients, and a distributed team. What it taught me about small business and the limits of doing it alone.
What it takes to sell and deliver automation to small businesses honestly, including the parts that make it harder than the pitch suggests.
If you are working out where the line should sit between automated and reviewed, that is a conversation I enjoy having.