What AI Should and Should Not Do in Your Bookkeeping
You run a 15-person agency. Receipts arrive as photos in a group chat, supplier bills land in three inboxes, and the bank export sits in a spreadsheet until someone has a free evening. You have heard that AI can "do the books". You also know that one wrong posting in March turns into a painful conversation with your accountant in April. So the real question is not whether to use AI in bookkeeping, but where exactly to let it work and where to stop it.
The split in one sentence
AI prepares, people post. Everything an assistant produces is a draft: a proposed expense, a suggested category, an answer with the source figures attached. A person with the right role reads the draft, fixes what is wrong and posts it. The ledger only ever changes by a human action, and the audit trail records who did it.
That rule sounds restrictive. In practice it removes most of the risk while keeping most of the time savings, because the slow part of bookkeeping was never the click on "post". It was typing 14 fields from a crumpled receipt.
Good fits: work that is tedious, repetitive and easy to check
Reading receipts and supplier bills into drafts
OCR plus a language model reads a photo or PDF and fills in supplier, date, net, VAT, gross, currency and a proposed category. The output is a draft document with the original image attached. Nothing is posted. Someone still looks at it, but they look at a filled form instead of a blank one.
Categorizing bank lines
After a Wise, MT940 or camt.053 import you get 80 to 300 lines a month. Most are obvious: the same SaaS subscriptions, the same payroll transfers, the same client paying the same invoice number. An assistant can propose a match or an account for each line and mark its confidence. You accept the obvious ones in bulk and spend your attention on the 10 that are not obvious.
Answering questions about your own numbers
"What did we spend on contractors for the Nordic client in Q2?" "Which invoices over 30 days are still open?" A finance copilot that reads your ledger can answer in seconds, and a good one shows the figures it used so you can click through to the account card and the source document. This is reading, not writing, so it is the safest place to start.
Spotting anomalies before period close
A supplier bill that is 40 percent higher than the last six from the same supplier. A duplicate invoice number. Hours billed twice on two invoices. An expense dated in a closed period. These checks are boring for humans and cheap for software. Running them a week before close means the corrections are small and the close is calm.
Bad fits: decisions with consequences and no review step
- Posting to the ledger. A posted document in a proper double-entry system cannot be edited; a mistake becomes a credit note or a reversing entry that lives in your books forever.
- Deciding tax treatment. Whether a lunch is fully deductible, whether a service from a US vendor triggers reverse-charge VAT, whether a laptop is an expense or a fixed asset: these depend on jurisdiction, on your accountant's policy and sometimes on facts the receipt does not show. The assistant can flag the question. It should not answer it in the ledger.
- Anything with no human in the loop. "Auto-post everything under EUR 50" sounds efficient until a wrong category is repeated 200 times. If nobody reviews, nobody catches the pattern.
- Anything the model has to guess. A receipt with no VAT line, a bank transfer with the reference "payment", a bill in a currency you do not normally use. The right output is a draft with a clear "needs attention" flag, not a confident-looking entry.
Grounded in your own data, connected through an open standard
What "grounded" means
An assistant grounded in your data answers only from your ledger, your documents, your rate cards and your bank lines. When you ask about Q2 contractor cost, it runs a query against your books and returns the sum with the rows behind it. It does not compose a plausible-sounding number from general knowledge.
This is why a bookkeeping assistant should not browse the internet for your numbers. Your figures are not on the internet, and the web is full of similar-looking figures that are not yours. An answer that cannot be traced to a trial balance line is not an answer you can use in a board deck or a tax filing.
A practical test: ask the assistant a question, then ask "show me where that came from". If the reply is an account card and a list of source documents, it is grounded. If it is a paragraph of reasoning, it is not.
Connecting an assistant through MCP
Model Context Protocol (MCP) is an open standard for connecting AI assistants to external tools and data sources. A system exposes a set of named tools (for example "list open invoices", "create draft expense", "income statement for a period"), and an assistant such as Claude or another MCP-capable client calls those tools with the user's permissions.
Two properties matter for bookkeeping. First, the tools define what the assistant can do; if there is no "post document" tool, the assistant cannot post. Second, a well-built server runs every call as a specific user, so the same roles and permissions apply as in the web interface. TridentERP exposes its ledger this way: an assistant can read reports and create drafts, and posting stays with a person whose name goes in the activity log.
Worked example: a EUR 184.50 restaurant receipt
Your delivery lead takes a client to dinner, pays with the company card and photographs the receipt with a phone. Here is what the assistant proposes.
| Field | Draft value | Source |
|---|---|---|
| Supplier | Trattoria da Marco | Header of the receipt |
| Date | 2026-08-27 | Printed date |
| Gross | EUR 184.50 | Total line |
| VAT | EUR 16.77 (10%) | VAT line on the receipt |
| Net | EUR 167.73 | Calculated |
| Category | Client entertainment | Suggested, confidence 0.82 |
| Project | Nordic retail platform | Suggested from the employee's current allocation |
| Paid by | Company card ending 4417 | Matched to a card statement line |
The 30-second review is a checklist, not a re-entry:
- Does the gross on the draft match the total on the photo? (EUR 184.50, yes.)
- Is the VAT line real or inferred? It is printed on the receipt, so it stays. If the receipt had no VAT breakdown, the reviewer sets VAT to zero or asks the employee for a proper invoice.
- Is the category right for your policy? Client entertainment may be partly non-deductible in your country. That is a tax question, and the reviewer applies the company rule, not the model's guess.
- Is the project right? The lead was working on two projects that week. The note on the receipt ("dinner with the Nordic team") confirms it.
- Is it a duplicate? The assistant already checked the card statement and found one matching line, so the answer is no.
Then the reviewer posts. The activity log records the draft created by the assistant on behalf of the employee, the edits made by the reviewer, and the posting by the reviewer at 09:41.
An audit trail that shows who posted what
The point of the review step is lost if the log only says "posted". A useful trail has four entries per document: who uploaded the source, which tool or assistant produced the draft, who changed which fields, and who posted. When your auditor asks in February why a EUR 184.50 dinner sits in client entertainment, you open the document, see the photo, the draft, the reviewer's correction and the posting, and the conversation takes one minute.
The same trail is what makes it safe to let more people in. A project manager can upload receipts and an assistant can draft them, because the ledger only moves when someone with the accountant role posts.
FAQ
Can I trust the categories the assistant suggests?
Trust them the way you trust a new junior: check every one for the first month, then check by exception. The suggestions improve when they are grounded in your own history of accepted categories.
What if the assistant reads a receipt wrong?
The draft is wrong, the ledger is not. The reviewer corrects the field before posting, and the correction is logged.
Does connecting an AI assistant give it access to all my data?
It should only have the access of the user who connected it, limited to the tools the system exposes. Read the permission list before you connect anything, and revoke the connection when a person leaves.
What to do next
- Write down which documents may be drafted by AI (receipts, supplier bills, bank lines) and which decisions stay with a named person (posting, tax treatment, fixed asset classification).
- Pick one month of receipts and run them through OCR into drafts. Time how long the review takes compared with manual entry.
- Ask the assistant three questions about your numbers and, for each, ask it to show the source. Discard any tool that cannot.
- Turn on anomaly checks a week before your next close and see what they find.
- Open the activity log on a posted document and confirm it shows the draft origin, the edits and the poster by name. If you want to see this flow end to end, the TridentERP demo tour shows receipt OCR producing a draft and the posting step that stays with a person.
