What Is Automation in Accounting? Bank Feeds, Capture and AI Agents

Automation in accounting means software doing repetitive bookkeeping steps for you: importing bank transactions through bank feeds, reading receipts and bills with capture tools, applying rules to categorise and match them, and now AI agents that suggest entries for you to review. Rules handle fixed patterns; AI is for messy ones.

The split matters because the two behave differently when something goes wrong. A rule does exactly what you told it, every time, including the wrong thing if you set it up wrongly. An AI suggestion can be right most of the time and occasionally wrong in a way no rule would predict. Knowing which you are looking at tells you where to check.

Follow me on Instagram@sagnikteaches

Four layers, from plain rules to AI agents

Most accounting automation falls into four layers. Each sits on the one below, and each needs a different kind of attention.

Connect on LinkedInSagnik Bhattacharya

The first two layers have existed for years and cost almost nothing to run. Layers three and four are where the recent change lies, and where the pricing and the claims need reading carefully.

Subscribe on YouTube@sagnikteaches

Layer 1 and 2: bank feeds and rules do the fixed work

A bank feed is a link between your bank and your accounting software that imports each transaction as it clears. It removes the keying and the typos. It does not decide what the transaction is. That is where rules come in: "if the description contains CLOUDHOST, code to software subscriptions" is a rule, and it will classify every matching line the same way until you change it.

Rules are powerful because they are predictable. They are also where quiet errors live. Take an illustrative recruitment agency with a rule that codes anything containing "AMAZON" to office supplies. In March, a one-off purchase of a laptop through the same marketplace is coded as supplies rather than equipment, and sits in the wrong place for a year. The fix is not AI; it is to narrow the rule (a supplier name plus a typical amount range) or to leave ambiguous suppliers to manual coding.

Practical advice for the rules layer: write a rule only for a supplier whose transactions always mean the same thing; review the list of rules twice a year; and look at the "uncategorised" or "suspense" total each month, because it should be small and falling.

Layer 3: capture reads documents so nobody retypes them

Capture tools extract data from a photographed receipt or a PDF bill and create a draft bill or expense. Dext is a long-standing specialist: its business pricing page lists $302.50 a year, billed annually, for 250 documents a month and 5 users, and once you hit the monthly limit, new uploads wait unprocessed until the next billing date or an upgrade. Xero now lists Smart Document Capture on all three of its plans with no add-on charge (the area Xero used to call Files is now Documents). QuickBooks bundles its own receipt and bill capture and, like the rest, changes plan details often, so read the current page for your plan.

The difference to understand is that capture is reading, not understanding. It can lift a total and a date from a clear receipt very reliably, and mistake a crumpled one. Tax lines, multi-page invoices and handwritten notes are the weak points. Xero's bills page says its AI extracts the key details to create draft bills ready for approval, which is the right design: the tool drafts, a person confirms.

A before and after for a physiotherapy clinic: before, the practice manager typed 60 supplier bills a month, about three minutes each, so three hours. After, she forwards bills to a capture inbox, checks each draft against the PDF in about 40 seconds, and spends 40 minutes. The saving is real but smaller than the headline, because checking never reaches zero. Take the number of documents your business handles a month and multiply by 40 seconds to see the realistic figure for you.

Layer 4: AI suggestions and agents that propose, you approve

The newest layer is software that learns from your history. Intuit groups its QuickBooks features under the Intuit AI and Intuit Intelligence names, and the one that matters here is Accounting AI, included with QuickBooks Online Essentials, Plus and Advanced but not Simple Start. It suggests the category, customer or vendor for each bank transaction and gathers the confident ones under Ready to post, where Intuit's help pages say you can review and post them all with a single click or change them first. Two settings are worth knowing: an amount threshold above which a transaction is never marked ready and always waits for you, and the guarantee that the AI never overrides your bank rules. It does post some matches by itself, such as paychecks and payroll taxes run through Intuit's payroll and bills paid through QuickBooks Bill Pay, and labels those with an AUTO badge. Xero's auto-reconcile feature is included on Growing and Established and costs an optional $5 a month on Early. In dollar list prices, Xero's plans are Early $27, Growing $59 and Established $97 a month, and QuickBooks Online's are Simple Start $38, Essentials $85, Plus $140 and Advanced $340 a month. Confirm both on the vendors' pages before choosing, since discounts and features shift.

"Agent" is a marketing word that covers a range. At the cautious end, an agent suggests a match and waits. At the ambitious end, it acts across steps, such as chasing an invoice or preparing a reconciliation, with less supervision. For a small business, the sensible rule is that anything touching money movement or filings keeps a human approval step. Intuit's overview of its AI features (updated August 2026) also says there is currently no way to turn AI features off individually, so if your accountant dislikes one you may have to work around it rather than disable it.

Where AI genuinely beats rules is in messy cases: a supplier whose name varies by month, a split payment covering three invoices, or a transaction description that is just a reference number. A rule needs one exact pattern per case; a model can say "this looks like the same supplier as the one you paid in June". It will also be wrong sometimes with the same confidence, which is why a review step is non-negotiable.

A month-end walked through, with rules, capture and AI side by side

Take an illustrative dental practice with one bank account, one card and a mix of lab invoices, supplies and utilities. Here is how the four layers share a month-end.

1. Day 1, bank feeds. The feed has already imported 214 transactions. The bookkeeper checks the closing bank balance in the software against the bank's own statement balance for the last day. They match, so the feed is complete.
2. Day 1, rules. Bank rules have coded 131 of the 214 lines (rent, payroll, software, utilities). The bookkeeper scans the 131 by supplier, not line by line, looking for anything with an unusual amount. Two stand out: a utilities bill almost double the usual, and a software charge that appears twice.
3. Day 1, capture. 38 supplier bills arrived by email into the capture inbox. The tool drafted 36 correctly; two had a wrong total because the PDF contained a delivery charge on a second page. Both are fixed against the originals.
4. Day 2, AI suggestions. The practice has set the software's review threshold at 1,000 dollars, so the three larger lab payments in the feed are never offered as ready to post and wait for a person. For the remaining 83 uncoded lines, the AI suggests categories. Sixty-eight look right at a glance. Fifteen need thought: a card payment described only as "SQ *KIT", a payment to a lab that the AI coded as supplies instead of cost of sales, and a transfer to the owner coded as an expense. The bookkeeper corrects these and adds a rule for the lab so it stops happening.
5. Day 2, reconciliation. Automatic matching has matched 190 bank lines to bills and invoices. The bookkeeper reviews the 24 it left unmatched and the 6 it matched with low confidence.
6. Day 2, review checks. Before closing, the bookkeeper runs four checks described below.

An illustrative time total: before automation this took about 14 hours. With bank feeds, rules, capture and suggestions, it takes about 5 hours of which roughly half is checking. The saving is around 9 hours a month, and it did not come from AI alone. Most came from the feed and the rules; capture added a good part; AI suggestions covered the messiest slice. That is why I would buy these in that order.

Three small cases that show where the line falls

A recurring supplier, handled by a rule. An estate agency pays the same portal subscription on the 3rd of each month, always for the same amount. A bank rule codes it to marketing and attaches the supplier. Nothing is gained by asking AI, and the rule is easy to audit: you can read it in ten seconds. Result: zero review time for that line, and a monthly glance at the rule list.

A varying supplier, handled by AI suggestion. The same agency buys signage and printing from three local firms whose bank descriptions change each time ("SIGN CO", "SIGNS 4 U LTD", a reference number). A rule per variant is a chore. An AI suggestion that has seen your past coding proposes "marketing" for each and gets it right 9 times in 10. You approve the batch in a minute, then correct the one that was actually a deposit for a refit.

A miss, caught by a check. A recruitment agency's card statement contains a charge for annual software that the AI coded as monthly software and the rule then repeated. The outlier check found a charge far larger than the others. The correction takes two minutes, and the lesson is to treat any line over a threshold you choose, say 500 dollars, as a manual review regardless of how it was coded.

A prompt for checking a month of coded transactions

If your software exports a transaction list, a general assistant can act as a second pair of eyes on it. Strip names and account numbers first, or use a business plan whose terms say content is not used for training, and treat the answer as a list of lines to look at, not a verdict.

Here is a month of coded transactions (date, description, amount, category).
List: (1) lines where the category looks inconsistent with the description; (2) amounts that are more than three times the average for the same category; (3) possible duplicates (same amount within 3 days); (4) any line coded to "owner" or "drawings" that could be a business expense.
Return a table with the line number and a one-sentence reason. Do not change any categories.

An illustrative output flags 11 lines out of 214: two probable duplicates, three outliers and six category doubts. Of those, perhaps five turn out to be real errors, and the rest are fine. Five real errors found in ten minutes is a good return, and none of the corrections happened without a person agreeing. What you would fix in the output: the assistant sometimes flags normal lumpy items such as annual insurance as outliers, so tell it your known annual payments in the prompt next time.

The four review checks that stay manual

Automation moves the work from keying to checking, so the checks are the product. Four are worth running every month whatever tool you use.

1. The balance check. Does the bank balance in the software equal the bank's statement balance on the last day? If not, a transaction is missing or duplicated. This catches broken feeds.

2. The suspense and uncategorised check. Is the total in "uncategorised" or "suspense" near zero? A growing balance means rules or AI are guessing and parking the leftovers.

3. The outlier check. Sort the month's expenses by amount and read the top ten. Capture and AI errors tend to be large (a total read as 1,200 instead of 120) so this is the quickest place to find them.

4. The unusual-pattern check. Compare this month's category totals with last month's. Anything that moved more than, say, 30 percent gets a one-line explanation. A rule that suddenly sends rent to supplies shows up here.

A realistic mistake from a small retailer: a rule coded a payment processor's fees to "bank charges", but one month the processor also netted off a refund in the same line. The category totals looked normal, and the error only appeared when the owner asked why profit was lower than takings suggested. The outlier and pattern checks would have caught it. Ask the same question of any automation: how would I notice if this were wrong?

Where plain rules are enough and AI is a waste

Do not pay for AI where a rule works. Regular payments with fixed descriptions (rent, loan repayments, software subscriptions, insurance) are rule territory, and a rule is cheaper, faster and fully explainable. Standing orders, payroll journals and recurring invoices are the same. A good test: if you can write the pattern in one sentence ("this supplier, this account, always"), it is a rule.

AI earns its keep with variable suppliers, split payments, ambiguous descriptions, and in explaining why two lines do not match. Use the same test for capture: a business with five regular suppliers who send electronic bills may need no capture tool at all, because the bills already arrive as structured data in the software, while a business with a stack of paper receipts does.

The cost angle is covered in how much AI bookkeeping software costs per month, and the comparison of the two big platforms is in Xero versus QuickBooks AI.

Rolling it out without losing control

Add one layer at a time and watch a month before adding the next. Month one: confirm the bank feed and clean up rules, deleting ones that no one remembers writing. Month two: add capture for the supplier bills you key most often, and measure the check time. Month three: switch on suggestions for uncoded lines, with approval required, and track how many you change. If you change more than one in five, the AI is not yet learning your books, and it may need cleaner history to learn from.

Write down who approves what. In a three-person firm, for instance: the bookkeeper approves coding and capture drafts; the owner approves any payment run and any journal over a set amount; nobody approves their own entries. For a practice running this for several clients, rolling out AI in a bookkeeping practice without losing control goes further, and automating bank reconciliation and checking the matches covers the matching step in detail.

If you are an accountant automating your own practice rather than a client's books, the workflow side (intake, reminders, document requests) is a different set of tools; TaxDome automations for a small practice covers what to automate first.

What automation does not cover

Judgement calls stay with people: whether a purchase is a repair or an improvement, how to treat a refund from a supplier, whether a payment to the owner is a loan or a drawing, what to accrue at year end. Automation can flag the oddity and propose an answer, but the accounting treatment depends on your facts and the rules that apply to you, so a qualified accountant should decide. The same applies to tax filings: software can prepare figures, and a person responsible for them signs them off.

A good first question for any new tool is "what does it do when it is not sure?" The best answers are "it leaves the line unmatched" or "it flags low confidence and asks". The worst is "it picks the most likely one and posts". Choose tools and settings that fail by asking. For the wider question of what AI can and cannot do for the books, can AI do my bookkeeping goes further, and if you want to map your own steps with someone, that is what my AI implementation consultation covers.

Questions small businesses ask about automated accounting

Is a bank feed the same as AI?

No. A bank feed is a secure connection that imports transactions from your bank into your accounting software. Rules that then categorise them are fixed instructions you set. AI comes in when the software suggests a category or match from patterns, rather than following a rule you wrote.

Will automation replace my bookkeeper?

It changes the work rather than removing it. Data entry shrinks, while review, judgement on unusual items and conversations with the owner remain. Most small businesses end up paying for fewer hours of keying and more hours of checking, and sometimes the same hours spent on better questions.

What is the riskiest part to automate?

Anything that moves money or posts without a person seeing it. Auto-approving payments or auto-posting uncertain categories lets one wrong rule repeat every month. Keep a human approval on payments and on any transaction the software marks as low confidence.

Do I need a paid add-on for each tool?

Often not. Bank feeds are standard, and capture and suggestions are increasingly bundled in the main plans. Check what your current plan already includes before paying for a separate capture or AI product, and compare the cost against the hours it would really save.

Further reads

• How to Speed Up Month-End Close in a Small Business With AI — Shorten the close itself, step by step.
• AI Bookkeeping Software vs a Human Bookkeeper: Accuracy and Cost — Accuracy and cost against a human bookkeeper.
• Accounting Software for CPA Firms: Automation Features Compared — Automation features compared for accounting firms.
• Is AI Worth It for a One-Person Bookkeeping Business? — Whether this pays off for a one-person business.
• Zapier for Accountants: Five Automations for Intake and Reminders — Intake and reminder automations for a practice.
• AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — A small CPA firm's first four agent jobs, scored and sequenced, with a review gate, pass rule and stop condition for each, plus a 90-day start.
• Sage Copilot for Accountants: What It Does in Your Sage Product — A product-by-product map of Sage Copilot for practices, with a subscription check and examples of the work accountants must still review.
• Which AML Client Checks Can an Accountant Automate With AI? — A step-by-step map of client due diligence in an accounting practice: what to automate, what AI may only draft, and what must stay with a person.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Xero pricing and bills pages, QuickBooks pricing page, Dext business pricing, and Intuit help articles on Accounting AI, Ready to post and Intuit AI (checked November 2026). Plans and features change, so check them before you buy.