Start with drafting client emails and letters, then meeting notes, then chasing missing records, and leave anything that touches the ledger until last. Effort rises with how many systems a job touches and how costly a wrong answer is. The ranking below goes from an afternoon's work to a project needing a trial period.
"Effort" here means three things added together: set-up time, how many tools or permissions must change, and how much checking the output needs afterwards. The checking cost is the one people forget, and it is what decides whether a use case saves anything. The examples use a bookkeeping practice of five people, with an occasional detour into a payroll or tax job; every figure is illustrative.
How the ranking works
Each use case gets two ratings. Set-up effort runs from 1 (an afternoon, no new software) to 5 (weeks, a trial and permissions in the accounting system). Risk is low, medium or high, judged by what happens if the output is wrong and nobody notices. A wrong sentence in a draft email is caught by the person who sends it. A wrong match in a bank reconciliation can sit in the ledger for months.
| Rank | Use case | Set-up effort | Risk | Needs access to the ledger? |
|---|---|---|---|---|
| 1 | Draft and tidy client emails and letters | 1 | Low | No |
| 2 | Meeting and call notes | 1-2 | Medium (recording consent, storage) | No |
| 3 | Chase missing records | 2 | Low | No |
| 4 | Explain figures in plain English | 2 | Medium (wrong reading of numbers) | No, pasted extracts only |
| 5 | Engagement letters and scope wording | 2 | Medium (contract wording) | No |
| 6 | Sort and triage the shared inbox | 3 | Medium | No |
| 7 | Month-end commentary and variance notes | 3 | Medium | Exports only |
| 8 | Client onboarding workflow | 4 | Medium | Sometimes |
| 9 | Receipt and bill capture | 4 | Medium to high | Yes |
| 10 | Bank reconciliation suggestions | 5 | High | Yes |
Ties are broken by risk: of two jobs with the same set-up, the safer one goes first. Treat the order as a sequence for a practice that has done none of these, not a law. A practice already on a cloud ledger with built-in suggestions may find rank 10 is half an hour of switching on, though the risk rating does not change.
1. Draft and tidy client emails and letters
This is where a practice learns what an assistant is good and bad at, because the person who sends the email is the checking step. Paste the email you received (remove names and account numbers if your plan or policy requires), tell the tool what you want to say, and ask for a reply in your firm's tone.
Reply to the client email below. Tone: warm, brief, plain English.
Say: we have the bank statements for April and May but not June;
we need June by Friday to file on time; no other action needed.
Do not promise a filing date. Do not mention fees.
Keep it under 90 words. Leave [first name] and [your name] as placeholders.
[paste email]A typical result: a polite five-line reply, correct on the facts you gave, with a closing line you did not ask for such as "Please don't hesitate to get in touch if you have any questions". The fix is a one-time instruction: add "no stock sign-offs" to the saved instructions. First step: pick the tool you already have (a business chat plan, or Microsoft Copilot Chat if the practice is on Microsoft 365), write one saved prompt for your three most common replies, and use it for a week.
Watch for: invented commitments. An assistant told to "reassure the client" may promise a deadline you never gave. The prompt above forbids it; keep that line in every template.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now2. Meeting and call notes
An AI note taker records or transcribes a call and produces a summary with action items. Set-up is small, but two decisions are not technical. First, consent: tell every participant, and for clients ask at the start of the call, not by assuming. Second, storage: where does the recording and transcript live, for how long, and who can open it? Check the tool's retention terms, and delete transcripts on a schedule you can defend.
Say a client call about a mid-year tax review runs 35 minutes. The tool's summary lists three actions: "client to send director loan statement", "accountant to confirm payment dates", "client considering vehicle purchase". The third is a trap: the client said they were thinking about it, and the summary has turned a remark into an action. Read every action list against your own memory within the day. For a comparison of tools by fit, Best AI Note Taker for Accountants: Which Fits a Small Practice? goes into cost and retention, and Otter's Pro plan lists at $8.33 per user a month billed annually ($16.99 monthly) as an example of the price range.
A second example from the same practice: the notes tool summarises a call with a payroll client and writes "employee to be paid" with no amount at all, because the speaker said the figure aloud and the transcript garbled it. Summaries drop or distort numbers more often than they distort sentences. House rule: any figure, date or name in a summary is checked against the source before it leaves the building, and the summary is stored as a draft until someone initials it.
3. Chase missing records
Every practice has a list of clients whose records are late. Drafting the reminders is quick to set up and low in risk, because the content is a polite request. Take your tracker (client, what is missing, deadline, last chased) and have the assistant write a first, second and third reminder at rising firmness, ending with the date after which you cannot guarantee the filing.
Before/after: the old way is one hand-typed email per client, 4-5 minutes each. The new way is one template with placeholders, a merge from the tracker, and a read-through of each message before it goes. Whether that saves time depends on the length of your list; for 8 clients it may not, for 60 it probably will, and you should measure it. The routine is covered in How to Chase Missing Client Records With AI Before Deadlines.
Watch for: chasing someone who has already sent the records, because the tracker was not updated. Add one line to the process: update the tracker before generating the reminders, not after.
4. Explain figures in plain English
Clients ask why profit is lower than last year or what a tax bill is made of. The task is to turn an extract into three sentences a non-accountant can follow. Paste a small, anonymised extract, never a whole file: income, cost of sales, overheads, and the two comparative figures.
Illustration: revenue is up from $182,000 to $196,000, but profit falls from $41,000 to $36,500. Ask for an explanation and the assistant will say costs rose faster than revenue. Fine. Then it will add "likely due to increased supplier prices", which nobody told it. The rule: the assistant may describe what the numbers show, but may not suggest why unless you supply the why. Add "do not speculate about causes" to the prompt and spot-check by recalculating one figure by hand.
With that line added, an illustrative answer reads: "Your sales rose by $14,000 to $196,000 this year. Profit fell by $4,500 to $36,500: total costs rose by $18,500, more than the rise in sales. The list attached shows which costs rose." Check it before it goes: $196,000 minus $182,000 is $14,000; costs went from $141,000 to $159,500, a rise of $18,500. Every sentence now says what moved and none claims why. If the client asks why, the answer comes from your file notes or a question to the client, never from the assistant.
5. Engagement letters and scope wording
Engagement letters are partly boilerplate, partly scope, and the scope is where trouble starts. An assistant can draft the scope paragraph from a short list of services ("monthly bookkeeping, quarterly management accounts, year-end accounts, not payroll"), and can generate a plain-English summary box at the top. It cannot decide what your terms should say, and it can quietly drop a limitation clause when it rewrites.
The check: take the old approved letter and the new draft, and compare clause by clause, not by reading the new one on its own. Anything that disappeared has to be a deliberate decision. Have a qualified adviser review any change to liability or termination wording. More on this in AI Engagement Letters for Accountants: Faster Drafts, Clear Scope.
6. Sort and triage the shared inbox
This one rises to effort 3 because it needs rules, not just a prompt. The goal is a first-pass label on each incoming email: needs records, tax question, invoice query, new enquiry, spam. Many mail systems can do this with built-in rules, and the AI part adds labels based on content rather than keywords.
Start by sorting 50 old emails by hand into your categories, and keep the result as a test set. Run the assistant over the same 50 and count agreement. If it matches your labels 45 times out of 50, it is useful for a first pass, with a person confirming; if it manages 35, fix the category definitions before anything else, since vague categories cause most of the disagreement. Never let the triage step reply automatically at this stage. A client's "can I just check, is this deductible?" should reach a person.
A concrete case: a Monday inbox of 140 messages after a long weekend. Sorted by hand into five labels it takes a staff member about 40 minutes. With a trial run over the previous week's 50 emails, the assistant disagrees on 7. Looking at the 7, five are "tax question" versus "invoice query" where the client asked about the tax shown on an invoice; the category definitions overlap. Rewriting the two definitions ("tax question: asks what to do or what is allowed; invoice query: asks about an amount on a document we issued") fixes four of the five. That is the pattern: disagreement is mostly your categories, not the tool.
7. Month-end commentary and variance notes
For clients on management accounts, the commentary page is a time sink: this month against last month, against budget, with a line on each big movement. Export the two trial balances (or the profit and loss report), strip the client name, and ask the assistant to list every line that moved by more than a threshold you set, say 10% and $500, with the movement in dollars and percent.
Below are two months of profit and loss lines as a table.
List only lines where the change is more than 10% AND more than $500.
For each give: line name, last month, this month, change in $ and %.
Do not explain causes. Mark anything you are unsure of with "CHECK".
Then list questions I should ask the client, one per line.A realistic output is a clean table of six lines plus questions such as "Why did subcontractor costs rise $2,300?", which is exactly the useful part: a list of questions, not explanations. Recalculate two of the percentages by hand, because arithmetic on pasted tables is where assistants slip. A line that was zero last month and is suddenly large, or a movement on an account the client never uses, is more often a miscoding than a real change; send it to whoever posted the entries before it reaches the client's commentary.
8. Client onboarding workflow
Onboarding touches several systems: the enquiry form, identity checks, the engagement letter, the ledger access request, the practice software. AI is only one piece; most of the effort is wiring those steps together so that a signed letter creates the client record and a checklist. Practices on a workflow tool such as TaxDome, which offers workflow templates with automated status updates and client reminders, can build much of it without extra software.
The AI contribution is narrow: read a returned questionnaire and flag blanks, or draft the welcome email using the services selected. Keep the identity and compliance checks fully human, and do not let an assistant mark a check as passed. Effort is 4 because a mistake in the wiring affects every new client; test with three dummy clients before using it live. The full process is in How to Automate Client Onboarding in a Small Accounting Practice.
Another example, from the quieter end of the same job: after onboarding, the practice sends a welcome pack with the list of records the client should send each month. An assistant can write that list from the service tiers the client bought: a monthly bookkeeping client gets bank statements, sales invoices and expense receipts by the 5th; a year-end-only client gets a single list due after the year ends. The risk is a pack listing services the client did not buy, so generate each pack from the tier field, never from free text, and have a person read the first ten.
9. Receipt and bill capture
Capture tools read a receipt or supplier bill and extract supplier, date, amount and tax, then push a draft to the ledger. Two kinds exist: standalone tools such as Dext, and capture built into the accounting system. Xero's pricing page includes its own Smart Document Capture in all three plans, while Dext's pricing for businesses starts at $302.50 a year, billed annually, for 250 documents a month and five users (as listed in November 2026; check the current pages).
Why effort 4 and risk medium to high: the tool writes into the ledger, and it can read a total wrongly or pick the wrong tax rate. Run a trial: take 100 documents from one client, let the tool process them, and have someone compare each extracted field to the document. Record errors by type: wrong supplier name, wrong total, wrong date, wrong tax code. A single-digit error count in 100 documents might be acceptable with a review step, but decide the threshold before you look at the results, not after. A practice buying capture for its clients should price from Dext's separate practice pricing page, because business and practice plans are sold differently.
Watch for: duplicate documents (the same receipt sent by email and uploaded), and foreign-currency receipts where the capture reads the symbol but not the exchange rate.
10. Bank reconciliation suggestions
Last, because it has the highest risk and the most setup. Modern ledgers suggest matches between bank lines and invoices or bills, and some use AI to propose codings. Xero's JAX auto-reconciliation is listed on Growing and above, and Xero's pricing page also shows it as an add-on on Early; I would not quote an add-on price without checking the page. Intuit's help page on Intuit AI in QuickBooks Online says there is currently no way to turn AI features off individually; its community team adds that AI bank categorisation can be avoided by switching back to the previous banking page where that option still appears. In QuickBooks, then, the question is mostly how to review the suggestions, not whether to allow them.
A wrong match does not announce itself. Two invoices for the same amount from the same supplier, matched to the wrong payment, leave the bank balance right and both supplier balances wrong. The control is a review step: accept suggestions only after a person looks at each one, keep the first month's rate of rejected suggestions as a baseline, and have someone not involved in the work run a sample check monthly. The set-up and checks for one platform are in Xero JAX Bank Reconciliation: Set It Up and Check the Matches.
A worked hour-count: testing use cases 1, 3 and 7
An illustration, with made-up figures, of how to test before committing. The practice has five staff and 60 clients. For two weeks everyone logs minutes on three tasks: replying to client emails, chasing records, and writing month-end commentary for the 12 management-accounts clients. The log has four columns: task, minutes, who, and a "checked or reworked" minute count.
| Task (two weeks) | Before: minutes | After: drafting | After: checking | Net change |
|---|---|---|---|---|
| Client email replies (90) | 810 | 250 | 330 | -230 |
| Record chasers (34) | 150 | 40 | 85 | -25 |
| Month-end commentary (12) | 720 | 210 | 600 | +90 |
Read it honestly. Emails fell 230 minutes over the fortnight (810 before, 580 after, drafting plus checking); chasers barely moved, a saving of 25 minutes across 34 messages; commentary got slower by 90 minutes because the checking of the tool's figures took longer than writing the notes by hand. The decision: keep the first, make the second a template without extra software, and drop the third until the prompt returns questions rather than explanations. None of this is a forecast; your numbers will differ, and the whole point of the log is that you stop relying on anyone else's.
Two jobs deliberately left off the list
Tax advice and filing decisions are not on the ranking because no set-up level makes them safe to hand over. An assistant can summarise a rule you have already looked up; it cannot be trusted to know which rule applies to this client this year, and the consequences land on the practice. Treat any tax position that comes out of a chat window as a lead to verify against the primary source, not an answer.
Automated client-facing chatbots are also left off. They need a knowledge base, an escalation path to a person and clear wording about what the bot is not, and a small practice rarely has the spare hours to maintain those. If clients ask the same five questions repeatedly, answer them in a short page or a saved reply first (rank 1), and revisit a chatbot after a year of use shows which questions never go away. When that day comes, start with a bot that answers only from your own published pages and hands anything about a client's own figures to a person.
Rules that apply to all ten
- One owner per use case. A named person reviews output and decides if it stays.
- Data in, data out. Write down what client data each tool receives and where it is stored, before the first test.
- Review before send or post. Nothing goes to a client or into a ledger on the tool's say-so.
- Measure two weeks before and after. Include the checking time.
- Stop rule. If the error rate in your sample check rises for two consecutive months, switch the use case off and find out why.
The cost side of all this is mapped in How Much Does AI Cost a Small Accounting Firm?
Choosing your first two
Most practices should pick one from ranks 1 to 4 and one from ranks 5 to 7, run both for a month, and not touch ranks 8 to 10 until the first two have survived a month-end and a filing deadline. Rank 1 teaches the tool's habits. Rank 3 or 7 returns time in the busiest weeks. Ranks 9 and 10 can wait for the quiet season, when a trial can run without a deadline pressing on the review.
If a use case fails the hour-count, that is a result, not a setback; you have learned something about your routines for the price of a log sheet. If you would rather run the first test alongside someone, that is what my AI implementation consultation covers; you can get in touch here.
Questions before you pick the first use case
Which AI use case should an accounting practice start with?
Start with drafting and rewriting client emails and letters. It needs no software change, no client data beyond what is already in the email, and every output passes through a person before sending. It is the lowest-effort, lowest-risk place to learn how the tool behaves before moving to anything that touches ledgers.
Can I paste client figures into ChatGPT or Claude?
Only under a plan and settings your engagement terms and data-protection duties allow. Business plans from the main vendors do not train on business content by default, but check the plan, the retention terms and what your clients agreed to. If unsure, anonymise figures and names first, or use the AI built into your practice software.
How do I measure whether a use case is worth keeping?
Log minutes per task for two weeks before, then two weeks after, including the checking time. Keep the use case only if total time including review falls and error rates stay flat. Judge it on your own figures, because vendor claims and other firms' numbers do not carry over.
Should AI ever post entries to the ledger without review?
Not in a small practice that is still learning the tool. Let AI suggest codings or matches, then have a person approve them, at least until you have a few months of reviewed results showing the error rate. Keep a sample check even after that.
Do I need to tell clients I use AI?
Check your engagement terms and professional body guidance, but telling clients is sensible practice. A short line in the engagement letter saying which tools are used on their data, and that a person reviews all output, answers most questions before they are asked.
Further reads
- AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the basics are in place before the first trial.
- Rolling Out AI in a Bookkeeping Practice Without Losing Control — Rolling AI out in a bookkeeping practice without losing control.
- Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — Tool choices sorted by task if you want names.
- 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways AI use can damage client confidence.
- Free AI Tools for Accounting: Which Jobs Free Plans Handle — Which of these jobs free plans can handle.
- Best AI Accounting Agents: QuickBooks, Xero JAX and Sage Copilot — What the built-in agents in the big ledgers do today.
- Accounting Practice Management Software UK: A One-Person Firm Checklist — Practice software checklist for one-person firms.
- How Accountants Use AI to Explain Tax to Clients in Plain English — Turning tax figures into plain English clients can follow.
- AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — A small CPA firm's first four agent jobs, scored and sequenced, with a review gate, pass rule and stop condition for each, plus a 90-day start.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Xero pricing page, Intuit help (Overview of Intuit AI in QuickBooks Online) and QuickBooks community replies, Dext business pricing page, Otter pricing page, TaxDome workflow page, Microsoft 365 Copilot business page, OpenAI and Anthropic data-use pages (checked November 2026).