An honest map, not a sales pitch
You have probably been told that AI is about to transform accounting. You have probably also noticed that nobody selling that line has ever had to explain a misstatement to a client, a partner or a regulator.
So let us do this properly. There is a real line running through your working day. On one side sit tasks where AI saves you genuine hours and the worst case is a slightly clumsy sentence. On the other side sit tasks where the worst case is a wrong figure in something you signed. This whole course is about knowing which side of that line you are standing on at any given moment.
The single fact that draws the line is this: an AI chat tool produces text that looks like a good answer. It does not add up your columns and report the result. It writes a sentence containing a number that fits the shape of the sentence. Sometimes that number is right. You cannot tell which times by looking.
In plain English
- Large language model:
- The technology behind chat AI tools. It predicts likely text one piece at a time. It is not a calculator and has no ledger inside it.
- Prompt:
- What you type in. The more context and constraint you give, the more useful and the more checkable the reply.
- Hallucination:
- A confident, fluent, entirely invented answer. It reads exactly like a correct one, which is the problem.
- Context window:
- How much material the tool can hold at once. Paste in too much and earlier parts quietly fall out of view.
- Non-determinism:
- Ask the same question twice and you can get two different answers. Neither one is flagged as the uncertain one.
What it is genuinely good at
These are language tasks, and language is the thing it actually does.
Drafting client correspondence. The chase email for the sixth time. The polite note explaining why the figures moved. The onboarding message you have written four hundred times and still resent writing.
Explaining a treatment in plain English. You know why the depreciation policy is what it is. Turning that into something a bakery owner understands without feeling patronised is a separate skill, and AI is good at it.
Summarising long documents. A lease, a loan agreement, a set of board minutes. Ask it to pull out the dates, obligations and clauses that might matter, then read those parts yourself.
Categorising transaction descriptions. Messy bank narratives sorted into your chart of accounts. This is genuinely useful and gets a whole lesson later.
Writing up procedures. Your firm's month-end process exists mostly in your head. AI is very good at turning a rambling verbal description into a clean numbered procedure.
First-draft commentary. Give it a set of verified figures and ask for the narrative. You will rewrite half of it, but starting from half is faster than starting from nothing.
Notice the pattern. Everything on that list is a task where you already know the answer and want help expressing it, or where you will read every word of the output anyway.
What stays with you
The figures themselves. Every number that ends up anywhere official comes from your accounting system, your spreadsheet or your own hand. Never from a chat window.
Judgment on treatment. Whether something is capital or revenue, whether a provision is required, whether that related party transaction needs disclosing. AI can remind you what the considerations are. It cannot weigh them for your client.
Anything filed or signed. If your name goes on it, you own every digit in it.
Anything your professional body or the law reserves for a qualified human. That reservation exists precisely because someone has to be answerable, and software cannot be answerable.
This course is general education, not professional, tax, legal or regulatory advice. It makes no claims about what any particular jurisdiction, regulator or professional body requires. Your obligations are set by your own professional body, your regulator and your own judgment. Check with them.
The confidentiality problem nobody mentions first
Before you paste anything, stop. Consumer AI tools are third party services. Client financial data, personal data, payroll records and anything covered by a confidentiality clause should not go into one unless your firm has done the assessment and told you which tool is approved and on what terms. Do not paste client data into a consumer AI tool.
Rewrite the following so it keeps the structure and the relationships
but replaces every real identifier with a placeholder.
Replace names with Client A, Client B and so on.
Replace account numbers, addresses and reference numbers with placeholders.
Keep the amounts and dates as they are, since I need the shape of the data.
Return only the rewritten version.
Even then, ask yourself whether the amounts and dates alone identify the client. In a small sector, they often do.
Checkpoint
Language tasks: yes. Numbers, judgment, and anything you sign: yours. And nothing leaves your desk into a consumer tool without a confidentiality check first.
Two prompts to start with
You are helping me explain an accounting matter to a client
who runs a small business and has no finance training.
The matter is: [describe it in your own words].
Write about 150 words, in plain English, no jargon,
warm but not chatty, second person.
Do not include any figures. I will add those myself after checking them.
End with one sentence offering to talk it through on a call.
Here is a document. Summarise it for me under these headings:
- Key dates and deadlines
- Obligations on each party
- Anything with a financial consequence
- Anything ambiguous or that I should read in full myself
Quote the exact wording for anything under heading 4,
and give me the section number so I can find it.
Do not draw conclusions about the accounting treatment.
That last line matters more than it looks. You are asking for retrieval, not judgment. Retrieval it can do well. Judgment is the part you are paid for.
โ Weak prompt
Prompt
Should my client capitalise this equipment purchase or expense it?
Output
You should capitalise it and depreciate it over five years on a straight line basis.
Confident, specific, and about a client it knows nothing about. It has no idea of the amount, the policy, the useful life, the jurisdiction or the materiality threshold. It answered because it always answers.
โ Good prompt
Prompt
I am deciding how to treat an equipment purchase for a client. Do not tell me the answer and do not recommend a treatment. List the factors I should be weighing, and for each one tell me what information I need to gather before I can decide. Flag anything that depends on jurisdiction or on my client's stated accounting policy.
Output
A structured list of considerations, each with the information needed, and jurisdiction-dependent points clearly marked as things to check against the client's policy and the applicable framework.
Now it is a prompt for your own thinking rather than a substitute for it. You still make the call, which is the only way this can work.
I am going to describe a process I follow, in a messy order,
the way I would describe it to a colleague.
Turn it into a numbered procedure with a clear owner
and a clear check at each step.
Where I have skipped a step or been vague,
list the gaps at the end as questions rather than filling them in yourself.
Here is my description: [paste it].
Read the gaps list carefully. It is often the most useful part of the reply, because it shows you where your own process is undocumented.
๐ Quiz
Question 1 of 4Why is a chat AI tool unreliable for producing figures?