Before you start
This course gives general information about using AI tools at work. It is not accounting, tax or legal advice. Always apply your firm's policies, your professional body's rules and the primary sources (legislation, standards, HMRC guidance) to your own situation.
In this level you will:
- See how the profession is actually using AI, with dates and sources.
- Learn three limits of general-purpose AI assistants, with a finance example of each.
- Sort your own tasks into three groups.
The examples use ChatGPT, Gemini, Claude and Microsoft 365 Copilot, but the points apply to any large language model (LLM) assistant.
Task 1: Read the evidence and know what it shows
Use figures that are dated and sourced. Two recent examples:
- Intuit 2026 Accountant Technology Survey (published 1 July 2026; 725 US accounting and bookkeeping professionals; surveyed May 2026). 88% had used AI for at least one client service in the past year, but only 30% said AI was embedded as the default in day-to-day work. The largest group (54%) used it only for some tasks.
- ICAEW, Evolution of Mid-Tier Accountancy Firms 2026 (leaders of 35 UK firms, surveyed February–March 2026). 86% had a technology strategy that explicitly included AI adoption.
What these figures show: AI use is now common, and firms are planning for it.
What they do not show:
- They do not show that the output was accurate or saved time overall.
- They do not represent everyone. One survey is US-only. The other covers 35 mid-tier firms, not sole practitioners or finance teams in industry.
- "Used AI for at least one service" can mean a single email draft.
DO: Each time you quote a statistic, give the source, the date, the sample size and the country. Then add one sentence on what it does not prove.
Task 2: Learn the three main limits
An LLM predicts likely text. It does not look facts up in a ledger or a database unless a tool is connected to it, and even then it can make mistakes. Think of it as a fast, well-read trainee who sounds confident and never says "I'm not sure" unless you ask.
Limit 1: Made-up facts ("hallucinations")
The model can produce sources, figures or case names that look real but do not exist.
- Finance example: In Harber v HMRC [2023] UKFTT 1007 (TC), a taxpayer appealing a penalty for failing to notify capital gains tax gave the First-tier Tribunal nine tribunal decisions supporting her case. None of them were real. The tribunal found they had been generated by an AI system. Some looked like real cases in which the taxpayer had actually lost.
- DO: Check every citation, rate and reference against the primary source before you rely on it.
Limit 2: Out-of-date knowledge
Each model is trained on data up to a cut-off date. Unless it searches the web, it may not know about later changes, and it may not tell you that.
- Finance example: From 6 April 2025 the employer National Insurance rate rose to 15% and the secondary threshold fell from £9,100 to £5,000 a year. A model relying on older data could still give you 13.8% and £9,100 when you ask it to calculate payroll costs.
- DO: Ask the tool which tax year its figures apply to, then confirm rates and thresholds on GOV.UK.
Limit 3: Weak arithmetic
LLMs handle numbers as text. Long additions, percentages over many rows and rounding can come out slightly wrong while looking neat.
- Finance example: You paste 40 invoice lines into a chat and ask for the total and the VAT at 20%. The answer may be a few pounds out, or one line may be counted twice.
- DO: Do the calculation in a spreadsheet or your accounting software. Use the AI to explain or draft, then check its totals against your own.
Task 3: Sort your tasks into three groups
Use this framework for any task:
Group A: Suitable for AI drafting. The content is low-risk, easy to check and you finish it yourself.
- First draft of a polite reminder email asking a client for missing receipts
- Rewording a technical paragraph in plain English
- A checklist outline for month-end tasks
Group B: Suitable for AI with close review. The output feeds into work that matters, so you check every line.
- Summarising a technical update, checked against the source document
- Suggesting a spreadsheet formula, tested on known data
- Flagging unusual transactions for you to investigate
- Draft commentary for a management report
Group C: Unsuitable for AI. These need professional judgement, legal responsibility or a decision only a person can make.
- Signing off accounts or a tax return
- Final decisions on tax treatment for a client
- Deciding whether to make a suspicious activity report
- Anything where the AI's answer would be the only evidence
DO: Write down five tasks from your last working week. Put each one in A, B or C. For each B task, note how you would check the output. Later levels cover which tools you may use and what data must never go into them, so do not paste any client data yet.
Task 4: Apply a quick check routine
Before using any AI output:
- Source: Can I trace every fact to a primary source?
- Date: Does this reflect current rules for the right period?
- Numbers: Have I recalculated the figures in a proper tool?
- Judgement: Is a decision being made that should stay with me?
Key takeaways
- This course is general information, not accounting or tax advice.
- AI use in accountancy is now common, but the surveys measure use, not accuracy. Always give the date, source, sample and limits of any statistic you quote.
- The three main limits are made-up facts, out-of-date knowledge and weak arithmetic.
- Sort tasks into: AI drafting, AI with close review, and unsuitable for AI.
- Professional judgement and sign-off stay with you.