Free lesson · Module 1 of 7
What AI can and cannot do in an accounting practice
By The AI Ledger editorial team. About 10 minutes to read. General guidance on working practices, not tax or legal advice.
In short
- AI is good at reading documents, suggesting categories, drafting words and summarising.
- It is not a substitute for judgement, current rules from a proper source, or arithmetic you have checked.
- Before anyone uses it with client data, agree five things in a short practice policy (the list is below).
Where AI already saves time
Most of the time AI saves in a practice comes from a few plain jobs, not from anything dramatic. These are the five that pay off most often in the tools we review:
Reading documents
Pulling supplier, date, amount and VAT from receipts, invoices and bank statements. This is the most mature use: the capture tools have done it for years, and newer models read messier documents.
Suggesting categories
Proposing a nominal code or a bank match from past patterns. The suggestion is often right for routine transactions, and a person still accepts it.
Drafting words
First drafts of client emails, records requests, engagement letter wording and meeting notes. Faster than starting from blank, and easy to check because you know what it should say.
Summarising
Turning a long document, a thread or a call recording into a short note with the actions in it.
Spotting the odd one out
Flagging duplicates, unusual amounts or transactions that break a pattern, so a person looks at fewer lines.
Notice what they have in common: each produces something a person can check quickly. That is the test to apply to any new use. If checking the output takes longer than doing the job, the tool is not ready for that task.
Where it does not
Judgement calls
Deciding a tax treatment, whether a cost is allowable or how to present something in the accounts. A model can describe the options; the decision and the responsibility stay with you.
Current rules without a source
General assistants can state thresholds, rates and dates confidently and wrongly, especially for recent changes such as MTD for Income Tax. Check every rule against gov.uk or HMRC's own guidance.
Arithmetic you have not checked
Language models are not calculators. Use the ledger or a spreadsheet for the numbers, and use AI for the words around them.
Anything you cannot review
If nobody in the practice can tell whether the output is right, AI is not saving time; it is moving risk.
The three kinds of AI you will meet
AI inside the ledger. Xero, QuickBooks and Sage now build assistants and automation into the accounting software itself. The advantage is that the data never leaves a system you already trust; the limit is that it only sees what is in that ledger. See our reviews of Xero, QuickBooks and Sage Copilot.
Specialist tools. Capture, reconciliation, close, practice management and tax tools that do one job well and connect to the ledger, such as Dext for document capture. Most of the 300 tools in our directory are in this group, and it is where the supplier checks matter most because your clients' data passes through another company.
General assistants. Chat tools that can draft, summarise and explain almost anything. They are the most flexible and the easiest to misuse: what happens to anything typed into them depends on the product and the plan, so read the terms of the version your staff actually use before any client detail goes in.
The risks, and the controls that handle them
Client data. Personal data about clients and their customers is covered by UK GDPR, so a tool that processes it is acting for you and needs terms that say so. Check where the vendor stores data and whether it uses customer data to train its models. Our tool pages show both, in the vendor's own words, under the supplier check.
Confidentiality. The codes of ethics of the accountancy bodies require you to keep client information confidential. Putting it into a tool whose terms let the vendor reuse it is hard to square with that, whatever the tool's quality.
Accuracy. Models produce fluent text that can be wrong. The control is simple and non-negotiable: a person reviews anything that reaches a client or the books, and the person who reviews it owns it.
Clients' expectations. Some clients will want to know whether AI touches their affairs. Decide what you tell them before they ask.
A five-point AI policy for your practice
A page is enough. Agree these five points, write them down and share them with everyone who works on client files:
- Which tools staff may use, and for which tasks, including whether client data may go into general assistants at all.
- A supplier check for each tool: where data is stored, whether the vendor uses your data to train models, and what security certification it states.
- A review rule: every AI output that reaches a client or the books is checked by a person, and the person who checks it owns it.
- What you tell clients about AI use, for example a line in your engagement letter or privacy notice.
- Who keeps the policy up to date, and when it is next reviewed.
Try this week
Pick one task from the first list that someone in your practice does every week, such as drafting records requests or reading supplier statements. Time it once by hand, then once with a tool you have checked against your policy, including the time spent reviewing the result. Keep the tool only if the second number is smaller.
Next in the course
Module 2: data capture and bookkeeping
How receipt and invoice extraction works, how to check it, and how to choose between the main capture tools. The full course opens in early 2027. See all seven modules.