Can an AI agent do eCommerce bookkeeping?
Partly. An AI agent can categorize transactions, reconcile Shopify orders, ad charges and bank activity, and flag anything that looks wrong. It should propose entries for a person to approve, not post them on its own. In our books it caught 384 orders worth about $49,000 that store credit had counted twice, and ad accounts that were never connected to the accounting system.
In April 2026 we had an AI agent read our Mully numbers for the year to date and explain where revenue came from. The first thing it found was not a growth insight. It was that 384 orders worth about $49,000 had been counted twice. Gross Shopify sales looked healthy. Real revenue was smaller.
That is the case for an AI bookkeeping agent in one sentence. It is less about speed and more about reading every transaction, every month, and noticing when two numbers that should match do not.
Can an AI agent actually do the bookkeeping?
It can do most of the prep and almost none of the sign-off. The work an agent handles well is the repetitive, rule-based part:
| Good fit for the agent | Keep with a person |
|---|---|
| Suggest categories for bank and card transactions | Approve entries before they post |
| Match Shopify payouts to bank deposits | Decide how to treat unusual items |
| Compare ad charges to ad platform reports | Tax filings and anything you sign |
| Flag duplicates, gaps and odd spikes | Capital, loan and pricing decisions |
| Draft a month-end list of open questions | Answering lenders and investors |
Our Bookkeeper agent runs at L2: it acts only after a person approves. Its one KPI is zero uncategorized transactions at month end. Clean books come first because every capital conversation starts with them.
What did the agent find in our books?
Two problems that a normal monthly glance would have missed.
1. Store credit counted twice. When a Mully subscription renewed, our subscription app issued store credit. The customer then redeemed it through a "pick your box" checkout, which created a second Shopify order for the same money. The agent traced 384 orders, about $49,000 year to date, through that loop. The recommendation was simple: tag every store credit redemption order and exclude it from revenue reporting.
This is the same logic Shopify uses for gift cards. Selling a gift card is not counted as a sale because it is a liability, and the full value is counted when the card is redeemed (Shopify Help Center). Store credit from a third-party app may not follow that rule for you, so check.
2. Ad spend counted from payments, not reports. This run, we checked Mully Group's books for July to October 2026. Neither the Meta nor the Google Ads report was connected to the accounting system, so ad spend for those months came only from what was paid: $19,108 to Meta and $2,646 to Google. That works for totals, but it means campaign-level spend never reaches the books.
Why do ad charges and ad reports disagree?
Because ad platforms bill on their own schedule, not by calendar month. Meta charges you when your costs reach a payment threshold and again on your monthly bill date, so you can be charged several times a month, or after your ads have stopped (Meta Business Help Center).
That means a month's card charges rarely equal that month's reported spend. A small gap is billing timing. A large and lasting gap usually means an ad account that is not connected, or spend on a card nobody mapped. The agent's job is to tell those apart and only flag the second kind.
Which ROAS number should the books show?
Show two: blended and new-customer. For Mully Group from July to October, blended ROAS, all sales divided by all ad spend, came out at 19.1. New-customer ROAS, sales from first-time buyers divided by ad spend, was 1.24.
Both are true. The first one includes every subscription renewal, which ads did not create that month. The second one is the number that tells you whether ads are paying for new customers. An agent that only reports the first will make your ads look 15 times better than they are.
How do you set up an AI bookkeeping agent safely?
Give it read access first and earn the rest. The order we use:
- Connect the sources. Accounting system, Shopify, bank and cards, and every ad account. Missing connections are the most common reason numbers do not tie.
- Write the checks. Payouts against deposits, ad charges against ad reports, orders against payments, and anything that can create two orders for one payment.
- Run it as flags only. For the first months, the agent writes a list. A person decides each one.
- Measure one thing. Uncategorized transactions at month end. When the agent's suggestions keep matching what your bookkeeper would do, let it apply routine ones after approval.
- Keep the trail. Save the flags and decisions with your records. The IRS generally says to keep records that support a tax return for 3 years, and employment tax records for at least 4 years (IRS). Publication 583 has the full table (IRS Publication 583).
What should you not expect from it?
Do not expect it to make judgment calls or replace your CPA. How to treat a loan, an owner draw or an inventory write-down is a decision with tax consequences. The agent can lay out the options and the numbers behind them. A person who is accountable for the answer should pick one.
Common questions
Will an AI bookkeeping agent replace my accountant?
No. It does the repetitive prep work, like categorizing, matching and flagging, so the books are clean when your accountant or CPA looks at them. Tax filings, judgment calls on how to treat unusual items and anything you sign should stay with a qualified person.
What should the agent be allowed to change in the books?
Start with nothing. Let it read, categorize as suggestions and write a list of flags. Once its suggestions match what your bookkeeper would have done for a few months, let it apply routine categorizations after approval. Ours runs at that level, L2, which means it acts only after a person approves.
How do I know if my store has a double-counting problem?
Look for any flow where one payment creates two orders. Store credit, gift cards redeemed through a second checkout and subscription apps that issue credit on renewal are the usual suspects. Compare payouts and bank deposits to Shopify sales for a month. A gap that grows with your subscriber count is a strong hint.
How long do I need to keep the records the agent works from?
The IRS says to keep records that support your tax return until its period of limitations runs out, generally 3 years, and employment tax records for at least 4 years. Keep the agent's flags and the decisions you made on them with the same records.