AI is useful to an accountant in three places: reading a ledger and explaining what it says, doing the repeatable preparation work that eats a morning, and drafting the documents a human then reviews. It is not useful for arithmetic, and it cannot act on its own. Every run starts because a person asked. Once you know that boundary, the question stops being whether to use AI and becomes which jobs to write down first.
Key takeaways
AI reads and drafts. It does not enforce. Nothing fires when an invoice goes overdue. Neither the Xero nor the QuickBooks connection can trigger on an event, so an assistant only ever works when you open it.
The model must never do the arithmetic. Language models total twenty numbers well and a hundred badly, and the wrong answer looks reasonable. Every calculation belongs in a script the AI runs.
Write the workflow down once, at five clients, not fifty. A procedure captured while the client list is small becomes the firm's method. One improvised at fifty stays improvised.
Most of the work is not technical. It is writing down how you already do the job. The setup is an afternoon per workflow.
Interest in this went from nothing to serious inside a year. Searches for AI in accounting now run in the thousands a month in every market we sell into, and the question behind almost all of them is the same one: what does it actually do, in a real practice, on a real client file.
This article answers that. What an AI assistant can and cannot reach in your ledger, the six workflows firms build first, how to tell a trustworthy setup from a dangerous one, and where the assistant has to hand the job over to something that runs without being asked.
An AI assistant reaches your ledger one of two ways. Either through a live connection, which Xero and Intuit both publish, or through a CSV export you hand it. Both work. The connection is faster for one client and the export is the only route that scales across a book, because every connection holds a single organisation and a bookkeeper with twenty clients would otherwise reconnect twenty times.
What it can do once it is there divides cleanly.
| Job | Can AI do it? | What actually happens |
|---|---|---|
| Read and explain the ledger | Yes, well | Aged receivables, movement since last week, which accounts got worse and why |
| Calculate | Only through a script | The assistant runs bundled code. The model itself is unreliable past a few dozen rows |
| Draft documents | Yes, well | Chase emails, review notes, management commentary, engagement letters, all as drafts |
| Send anything | Mostly no | No email tool exists in the Xero connection. QuickBooks can send invoices and overdue reminders, currently for United States organisations only |
| Watch for an event | No | Nothing triggers when an invoice passes its due date. A person has to ask |
| Work across a whole client book | One at a time | Each connection is a single organisation. CSV is the route that scales |
That table is the whole boundary, and most disappointment with AI in accounting comes from expecting the right-hand column to say something else. An assistant is analysis and drafting on demand. It is a very good junior who only works while you are in the room.
The six workflows accounting firms build first are client onboarding, the monthly query chase, management pack commentary, the year-end review pass, fee scoping, and the client email desk. Each one is a recurring job that eats a morning and produces a document somebody reviews. None of them are exotic, which is the point: the work that suits an AI workflow is the work you already do to a written standard, or should.
Client onboarding. One intake conversation produces the engagement letter, the information request, and the AML checklist together. The value is not speed on any one of them. It is that the thirtieth client gets the same three documents as the first, instead of thirty slightly different versions accumulating in a folder.
The monthly query chase. The assistant reads the transactions it cannot classify and drafts the "what was this payment" email per client. This is the job that quietly consumes a bookkeeper's week and it is almost entirely templated once written down.
Management pack commentary. Three reports out of Xero or QuickBooks go in, a one-page commentary comes out. Variance against last month, against budget, and the two or three things worth a sentence. A human edits it. The draft takes minutes instead of an afternoon.
The year-end review pass. A second set of eyes on a file that does not have to be yours. The assistant runs the firm's review checklist and lists what it could not tie, what looks unusual, and what is missing. It finds fewer things than a senior would. It finds them on every file, at the same standard, which a senior does not.
Fee scoping. Every new client priced from their actual data rather than from a guess about how messy they will be. Transaction volume, bank account count, the state of the last filed accounts. Underpricing the messy ones is where practice margin goes.
The client email desk. First drafts of the technical replies that eat an afternoon. Not sent by the AI. Drafted, so the partner edits rather than composes.
If you build only one, build the monthly query chase. It repeats most often, it is the least interesting work in the building, and it is the one where a written procedure most obviously beats whoever happens to be doing it that week.
One rule decides whether an AI output is safe to put in front of a client: the model must never do the arithmetic. It matters more than every tool choice put together, and it is the single thing to check before trusting any setup with a client file.
The model must never do the arithmetic.
Language models handle twenty numbers well and a hundred badly, and the failure mode is quiet. The total comes back looking entirely reasonable and wrong by a few hundred pounds. You find out in front of the client.
The fix is not a better prompt. It is moving every calculation into a script the AI runs. The model reads the request, runs the code, and explains the result. The arithmetic happens in software, so it is right, and it is identical every time you ask.
How to tell a trustworthy setup from a dangerous one
The numbers come from a script, not from the model. If the written procedure says "calculate the total", the model is doing it. If it says "run this script", the code is doing it.
Every output shows its workings. Source file, as-at date, row count, and a control total the reviewer can check. A figure with no provenance cannot be reviewed.
Exceptions get surfaced, not swallowed. Three invoices with no email address and one with no due date are part of the answer, not a detail to drop quietly.
The last one is the one people miss. AI mistakes are hard to spot because they are not the mistakes a human would make, and a clean-looking report with two rows silently excluded is worse than no report at all.
An MCP server is the connection to your data, a skill is the written procedure the AI follows, and an agent is the assistant running that procedure against the data. The three words get used interchangeably and they are not the same, and the distinction is what you need the moment something does not work.
| Thing | What it is | Kitchen analogy |
|---|---|---|
| MCP server | The connection to your data. Xero and Intuit both publish one. | The ingredients and the equipment |
| Skill | The written procedure the AI follows. | The recipe |
| Agent | The assistant running the procedure against the data. | The cook |
You need all three. A connection with no written procedure gives you an assistant that can reach the ledger but improvises differently every time you ask. A procedure with no connection still works fine against a CSV export.
A Claude Skill is the second row: a folder holding a file of instructions in plain English, plus any scripts it needs. There is no model training, no data upload, no new platform to learn. Skills run in Claude, Claude Code, Codex and Cursor. If you want the detail, we wrote it up in Claude Skills for accountants, and the connection steps are in connecting Claude to Xero.
We published our accounts receivable skill as open source under the MIT license, because the fastest way to show what a well-built one looks like is to hand over the code.
Aged receivables, DSO, late fee schedules, customer statements, a ranked call sheet and verified chase briefs. Every figure is computed by a bundled script, never by the model.
Get the skill on GitHubWorks in Claude, Claude Code, Codex and Cursor. Setup for each platform is in the repository.
The calculations are covered by 59 tests against a fixture ledger with hand-computed answers, so you can check the arithmetic yourself in about two seconds. That is the point of publishing it. A skill you cannot audit is a skill you should not put a client file through.
AI changes how the preparation gets done and changes nothing about judgment, the client relationship, or the fact that nothing happens unless a person asks. Being straight about those three saves a lot of wasted setup.
Judgment is still yours. An assistant produces a review note listing what looks wrong. Deciding whether it is wrong, and what to do about it, is the job you are paid for. Nothing in the last two years has moved that.
The client relationship is still yours. A drafted email in your tone is still an email you send, having read it. Firms that let drafts go out unread find out why that was a bad idea within about a month.
Nothing happens unless someone asks. This is the one people underestimate. The value of a workflow is capped by how often a human remembers to run it. A monthly pack is fine. Chasing an overdue invoice is not, because the moment it needs to happen is a Saturday and nobody is in the ledger.
Ask an AI assistant who to chase and it will tell you, with the workings. Ask it to draft eleven chase emails and it will, in your tone, with the right escalation on each. That is genuinely useful and we build for it.
What it will not do is notice on Saturday that an invoice has gone twenty days past terms, send the reminder, apply the fee the contract allows, and escalate when that reminder is ignored. That is not a limitation of the skill or of Claude. It is a limitation of the connection underneath, and it applies to every AI accounting integration available today.
💡 Paidnice insight
We built the skill and we still run the chasing through automation, because the two jobs are different. Judgment on demand is what an AI assistant is good at. Doing the same thing to every invoice, on time, without being asked, is what software is good at. Customers using Paidnice cut their average wait for payment in half within 30 days, and none of that came from someone remembering to run a report.
The division that works in practice is simple. Use the assistant for the thinking: the review, the analysis, the first draft, the explanation of what the numbers say. Use accounts receivable automation for the doing: reminders on schedule, late fees that apply themselves under your rules, statements that arrive without a person triggering them, and escalation that happens whether or not anyone opened the ledger that morning.
Firms that get value from AI are the ones that make this split deliberately, rather than trying to make an assistant do a job that needs to run at two in the morning.
An afternoon per workflow is a fair estimate once the procedure is written down. The writing is the part firms skip, and skipping it is why the second attempt always goes better than the first.
If you would rather start from the software than the procedure, we listed the AI tools accountants are actually installing, with published pricing and an honest limitation on each.
No. A written procedure is plain English, and the connection to Xero or QuickBooks is an authorization screen. The only part that touches code is the calculation scripts, and for the common jobs those already exist and are free.
The connection uses the same authorization your other apps use, and you control which organisation it reaches. The real risk is not access, it is unreviewed output: a figure the model invented, presented confidently. That is what the arithmetic rule above exists to prevent.
It replaces preparation, not judgment. The jobs that disappear first are the ones with a written right answer and no decision in them, which is most of what a practice currently charges juniors out for. What it does not touch is the part clients are actually paying for, which is somebody competent telling them what the numbers mean and what to do next. We went through the split in detail in will AI replace accountants, including the accounts receivable roles.
It can tell you who to chase and draft the emails. It cannot watch the ledger, and with a Xero connection it cannot send at all. Anything that has to happen on a schedule, without a person present, belongs in automation rather than an assistant.
Yes, and the sending rules differ. QuickBooks can send invoices and overdue reminders through its connector, currently for United States organisations only. Xero has no email tool in the connection at all.
Paidnice is accounts receivable automation that enforces your payment terms, trusted by thousands of businesses on Xero and QuickBooks. Credit control and debtor management, run for you.
Try it Now - It's FreeNo card required.
Optimize your cash flow with our suite of financial tools designed for AR professionals. Calculate DSO, aging analysis, late fees, and more.