AI for Finance — Learn AI

By the end of this lesson you will be able to

Apply the one hard rule about AI and arithmetic
Turn a variance table into a narrative in minutes
Know which finance tasks stay entirely with a qualified person

The one hard rule

Do not let AI do arithmetic you will rely on. It is a language model, not a calculator — it produces numbers that look right by pattern, and it is confidently wrong often enough to matter. Your spreadsheet computes; AI explains what the spreadsheet computed.

That rule sounds like it rules AI out of finance entirely. It does the opposite. An enormous share of finance work is not calculation — it is explaining, documenting, and communicating figures that were already calculated correctly elsewhere.

The split

Hand over

Writing the variance narrative from a variance table
Turning a model output into a board paper
Drafting policy and procedure documents
Summarising a long contract's financial terms
Explaining a technical treatment to a non-finance reader
Drafting collections and payment-chase emails

Keep

Any calculation
Judgement on accounting treatment
Anything going to a regulator unreviewed
Forecast assumptions
Deciding what is material
Sign-off. Always.

Variance narratives

The month-end task everybody dreads: turning a table of variances into readable commentary, on a deadline, having already done the hard part. The numbers are settled; the writing is the bottleneck.

Prompt

Here is our variance table for [month]. Figures are final
and already reconciled.

[PASTE TABLE]

Here is context I have on the drivers:

[PASTE YOUR NOTES — e.g. "the marketing overspend was the
trade show we moved forward from Q4"]

Write the management commentary.

Format: one short paragraph per variance over [threshold],
in order of size. State the number, the driver, and whether
it is timing or permanent.

Rules:
- Use only the figures I gave you. Do not calculate anything,
  including percentages or totals.
- If I did not give you a driver for a variance, write
  "driver to be confirmed" — do not infer one.
- Plain language. No "unfavourable variance attributable to".

Note the two explicit prohibitions

Telling it not to calculate — including percentages — matters, because deriving a percentage from two numbers you supplied looks harmless and is exactly where a wrong figure enters a document nobody re-checks. Telling it not to infer drivers matters for the same reason: a plausible-sounding explanation is worse than an admitted gap.

Translating for non-finance readers

One of the highest-value and least-recognised uses: taking something technically correct and making it comprehensible to an owner, a board, or a department head, without losing accuracy.

Prompt

Rewrite the paragraph below for a reader with no accounting
background — a business owner who is smart but has never had to
think about accruals.

[PASTE THE TECHNICAL PARAGRAPH]

Constraints:
- Keep every figure exactly as written
- Do not simplify to the point of being wrong
- If a concept genuinely cannot be simplified, keep the term and
  add one plain-language sentence explaining it
- Four sentences maximum

Policy and procedure drafting

Expense policies, approval matrices, month-end checklists, credit control procedures. These are structure-heavy documents where the content is known and the writing is the chore — a very good fit.

The pattern is the same each time:

Describe the rules roughly, in whatever order they come to you.
Ask it to list every case your description does not cover. This is where it earns its money.
Answer those, then have it produce the document.
Check every threshold and every named role before it circulates.

The awkward emails

Chasing payment is unpleasant, which is why it gets delayed, which is why it gets harder. Having a draft already written removes most of the friction.

Prompt

Draft a payment reminder.

Facts: invoice [number], [amount], issued [date], terms net 15,
now [N] days overdue. This is the [first/second/third] reminder.
Customer is [long-standing / new]. Relationship is [good / strained].

Format: email, 3–4 sentences, no bullet points.

Tone: escalate with the reminder number but never become
aggressive. Firm and matter-of-fact.

Include the invoice number and amount. Do not apologise for
chasing. Do not offer a payment plan unless I asked.

Worth building a set of three

First, second, and final reminders as saved prompts. Collections works far better as a consistent sequence than as three individually-agonised-over emails, and consistency is precisely what this makes cheap.

What about AI features in our accounting software?

Categorisation and anomaly detection in accounting tools are a different technology to a chat model, and generally more trustworthy for their specific job — they are trained on structured transaction data. Still review the categorisations; still check the data terms.

Can it read a set of accounts and tell me what is wrong?

It can produce plausible observations, some of which will be right, and you will not be able to tell which without checking every one. That is a poor trade. Use it to explain a variance you have already identified, not to find variances.

Is it safe to paste our management accounts in?

On a business tier with training disabled and appropriate retention, this is a normal use. On a personal free account, no — and this is one of the most common shadow-AI exposures in small businesses.

Key takeaways

Never let it calculate. Explicitly ban percentages too — that is where wrong figures sneak in.
Your spreadsheet computes; AI writes the narrative around what it computed.
"Driver to be confirmed" beats a plausible invented explanation.
Translating finance for non-finance readers is the most undervalued use in the function.
Build three saved collections prompts. Consistency is the thing that actually improves cash collection.