Many small accountancy practices can produce the numbers for monthly or quarterly management accounts fairly quickly. The bookkeeping software can generate the profit and loss report, balance sheet and budget or prior-period comparison. The real time is often spent afterwards.
Someone still has to explain what the numbers mean.
Why is revenue up? Why has gross margin moved? Are staff costs unusually high, or just in line with a known pay rise? Is the increase in software subscriptions a genuine trend, a one-off annual renewal, or a coding issue? Those comments are where the client gets much of the value from the management accounts.
Under pressure, that commentary can become rushed. In some cases it is skipped entirely, leaving the client with a pack of numbers but little practical explanation. This is a useful place for AI, provided it is used carefully.
An example workflow: variance commentary assistance
Imagine a small accountancy practice preparing quarterly management accounts for several retained clients. The accounts are exported from the bookkeeping platform, including actual figures, budget comparisons and prior-period comparisons.
An AI-assisted workflow could take that export, apply a materiality threshold agreed by the practice, and identify the variances that are actually worth commenting on. For each material variance, it could draft a plain-English paragraph explaining the movement and suggesting a likely driver where the data supports one.
It could also do something equally important: clearly list the movements it cannot explain from the data alone.
The outputs might include:
- A variance list showing size, direction and category
- Draft commentary for each material movement
- A list of unexplained variances to ask the client about
- A draft covering email to accompany the management accounts
- A note of any assumptions or source data limitations
This is not a real case study or a claim about a specific client. It is an example of the sort of workflow that can be built around management accounts where the figures already exist, but the explanatory work is taking too long.
The important guardrail: no invented explanations
Financial commentary has a different risk profile from ordinary admin. If the AI produces confident but unsupported explanations, the accountant could end up misleading the client.
That is why the workflow should not be designed to “write the answer” and send it. It should be designed to support the accountant’s review.
For example, if revenue is 18% above budget and the export shows a new sales category or a known one-off event already entered in the client notes, the AI can draft a cautious explanation. If gross margin has fallen but the data does not show whether the cause is pricing, supplier costs, stock movements or coding, the AI should not guess. It should raise a question for the client.
A useful line in the draft might be:
“Gross margin was lower than budget this quarter. The data shows the movement clearly, but does not identify the underlying cause. We should ask whether supplier costs, discounting or job mix changed during the period.”
That is much better than a made-up explanation dressed up as analysis.
What the accountant still does
The accountant’s judgement remains central. The AI can calculate, sort and draft. It cannot know the client relationship, the context of a difficult quarter, or whether a variance should be framed carefully because of a board discussion, lender covenant or tax planning issue.
A sensible approval process would include:
- The managing accountant reviewing every drafted paragraph
- Checking any suggested driver against source evidence
- Removing or softening anything that goes beyond the data
- Turning unexplained variances into client questions rather than guesses
- Approving the final covering email before it is sent
Used this way, AI does the repetitive groundwork and first-draft wording. The accountant supplies the judgement, edits the message and decides what is appropriate for the client.
What data would be needed?
The basic inputs are usually straightforward:
- A management accounts export from the bookkeeping platform
- Actual versus budget and/or actual versus prior-period figures
- A chart-of-accounts mapping so the system understands the categories
- Variance threshold rules, such as percentage and monetary limits
- Any known one-off events the client has already flagged
The thresholds matter. Without them, the system may produce too much commentary on small movements that do not matter. With sensible thresholds, it can focus attention on the items most likely to affect decision-making.
It is also worth keeping the language deliberately plain. Most SME clients do not need a technical essay. They need a clear explanation of what moved, why it may have moved, and what needs to be checked next.
Where the time saving comes from
For a small practice, this type of workflow could plausibly save 1-3 hours per client per reporting period. The saving is not usually in producing the accounts themselves. It is in reviewing the variances, deciding what needs a comment and getting from a blank page to a useful first draft.
That time saving can make better commentary more realistic, especially for smaller clients where the budget does not allow for long narrative reports. It can also help standardise the quality of the first draft across the practice, while still leaving final judgement with the accountant.
A practical first step
If you run an accountancy practice, start by reviewing the last three sets of management accounts you produced. Look at where the commentary took time, where explanations were repeated, and where you had to go back to the client because the data alone did not tell the full story.
That will show whether an AI assistant would be useful, and where the review points need to be.
If you would like help designing a practical, low-hype AI workflow for your own reporting process, I help professional firms identify useful opportunities, set sensible guardrails and build implementation plans that fit how the business already works. Book a short call or send through the process you have in mind, and we can look at whether it is a good candidate for AI support.
