Training junior accountants when AI does the grunt work
A step-by-step playbook for small firms training graduates in a practice where AI drafts much of the routine work. It sets out what juniors do by hand on purpose, how reviewing AI output becomes the training, weekly explain sessions, rotation, a sign-off ladder, how to measure capability, and a 12-month pathway to copy.
Part of our AI in accounting coverage. See the full AI for accountants guide →
Learning check · before you start
What do you already have in place?
A few quick questions, so you can skip the steps you have already done. Each one has a line explaining what it means. Not sure, or never heard of it? Choose “Not yet / not sure”: that is where most firms start, and the playbook walks you through it.
This playbook is for the owner or manager of a small firm who has hired, or is about to hire, a graduate into a practice where AI already drafts much of the routine work. By the end you will have a 12-month training pathway you can copy: what the junior does by hand on purpose, how reviewing AI output becomes the training, a weekly explain-your-working habit, a rotation, a sign-off ladder and a way to measure whether any of it is working. Setting it up takes an afternoon. Running it takes about an hour a week of senior time.
We covered the problem in when the junior stops asking the partner, how do they learn?: the slow loop of getting stuck and asking someone senior used to build judgement as a side effect, and AI has compressed it.
Step 1: Choose what juniors still do by hand, on purpose
Your task write a short list of tasks your junior does by hand, without AI, into their first-quarter plan, with a number against each and a named senior who checks them.
You said you already have this covered. Skim this step, or skip to the next.
Some work stays manual in the first months, not because AI cannot do it, but because doing it by hand is how a junior learns what correct looks like. Without that baseline they cannot check a machine’s version.
Pick a short list, and make it the work that teaches the most per hour. For most small firms that means:
- One bank reconciliation end to end, from statement to ledger, with no suggested matches switched on.
- One simple individual or sole trader return prepared from source documents, not pre-filled data.
- One GST, VAT or sales tax reconciliation traced line by line to the ledger.
- One set of year-end adjusting journals worked out on paper before they are posted.
Set a number: say two of each in month one, then one a quarter as a refresher. Written down, a busy month cannot quietly delete it.
Done when: the hand-work list is written into the junior’s first-quarter plan with a number against each item and a named senior who checks them.
Learning check
What would you do? Pick one; every answer is explained.
Step 2: Make reviewing the AI the training method
Your task start an error log for every AI-drafted task the junior reviews, and book the first monthly check where a senior plants an error for them to find.
You said you already have this covered. Skim this step, or skip to the next.
Once a junior has done a task by hand, switch the direction. The AI drafts, the junior reviews. Finding what is wrong in something that looks finished is harder than preparing, and it is the skill a future reviewer needs.
Make it structured. The junior checks every AI draft against the source, marks each correction and records it in a simple error log: what the AI got wrong, how they spotted it, and which rule or document proved it. Once a month, a senior plants one or two deliberate errors in a draft and sees whether the junior catches them.
The log is also the firm’s training plan. Ben Walker described doing this at scale with review points from year-end jobs.
If Division 7A is a big area for mistakes, let’s direct training there and get a better result.
For Australian tax work there is a regulatory reason to train this way too. The Tax Practitioners Board’s AI guidance, TPB(GS) 55/2026, says practitioners “should verify and review AI generated content for accuracy throughout each step of the workflow” and that “Each of these steps should be documented.” A junior who has been taught to review and log is already working to that standard.
Done when: the junior keeps an error log on every AI-drafted task, and the first planted-error check is in the calendar.
Learning check
What would you do? Pick one; every answer is explained.
Step 3: Run explain-your-working sessions
Your task book a recurring 30-minute weekly session where the junior talks a senior through one job, and give them a one-line template: the question, three options, my pick and why.
You said you already have this covered. Skim this step, or skip to the next.
The old loop had a conversation in it. Put it back on purpose. Once a week, for 30 minutes, the junior brings one job and talks a senior through it: what they did, why, what the AI suggested, and where they disagreed with it.
The rule that makes these sessions work is that the junior never brings a bare question.
Then they come to you not with the question, but with the question, three proposed answers, and the one they reckon you should go with.
AI makes the three answers cheap to produce. The learning is in choosing between them and defending the choice. The senior adds what the tool cannot know: the client’s history and what happened last time.
Done when: a recurring weekly slot exists, the junior has a one-line template (“the question, three options, my pick and why”), and the senior notes one thing learned after each session.
Learning check
What would you do? Pick one; every answer is explained.
Step 4: Rotate across the work
Your task plan the junior's year as four three-month rotations across different areas of work, each with a named senior lead.
You said you already have this covered. Skim this step, or skip to the next.
A junior who reviews one kind of AI output learns one kind of error. Rotate them, in three-month blocks, across bookkeeping and reconciliations, compliance and returns, advisory and meeting preparation, and client communication. Each rotation should include sitting in on client calls, because reading a client is not something AI drafts for you.
Keep the system of record the same through every rotation. Whether the firm runs FYI, AccountKit or Xero Practice Manager, the junior should see each job from intake to sign-off in one place, so they learn the shape of the whole job, not a slice of it.
Done when: the junior’s year is mapped into four rotations with a named lead for each.
Learning check
What would you do? Pick one; every answer is explained.
Step 5: Build a sign-off ladder
Your task write down a five-rung sign-off ladder showing what a junior may approve on their own at each stage, list the work types for each rung, and tell the junior which rung they are on.
You said you already have this covered. Skim this step, or skip to the next.
A sign-off ladder says what a junior may approve on their own at each stage, with evidence for each step.
- Prepare. Drafts work, by hand or with AI. Everything is reviewed.
- Prepare and self-review. Submits work with their own review notes and error log attached.
- Review AI drafts. First check on AI-drafted low-risk work, such as reconciliations and document requests, before a senior’s final look.
- Approve low-risk outputs. Can send routine, non-advice client communication and approve agreed low-risk work types on their own.
- First reviewer. Reviews another person’s work, with a senior signing off.
Two limits sit outside the ladder. Tax advice and lodgement stay with whoever is accountable for them; in Australia that is the registered practitioner, and section 35 of the Code Determination requires anyone providing services on their behalf to be appropriately supervised. And nobody moves up a rung on time served: they move when the evidence from Step 6 says so.
Done when: the ladder is written down, each rung has a list of work types attached, and the junior knows which rung they are on.
Learning check
What would you do? Pick one; every answer is explained.
Step 6: Measure capability, not hours
Your task start recording four measures of the junior's progress each month (review points, planted-error catch rate, an explain check and error log quality) and set a date to review them at the end of the first rotation.
You said you already have this covered. Skim this step, or skip to the next.
Measure what the pathway is meant to build. Four measures are enough:
- Review points per job, by category, trending down over the year.
- Planted-error catch rate from the monthly check in Step 2.
- Explain test: can the junior defend a position to a senior without notes?
- Error log quality: are they catching new kinds of AI mistakes, or the same ones?
These records double as evidence for professional body experience requirements. As at November 2026, CPA Australia asks CPA Program candidates for 36 months of relevant experience demonstrating at least 10 skills from its skills list across four categories, and CA ANZ’s Mentored Practical Experience requires three years, and not less than 500 days, of approved experience with six-monthly mentor meetings. Check the current rules with the junior’s own body; UK, US and Canadian bodies and state boards set their own.
Done when: the four measures are recorded monthly and reviewed at the end of each rotation.
A 12-month junior pathway for an AI-assisted firm. Adjust the work types to your client base. Move up a rung only on evidence.
| Months | Rotation | By hand, on purpose | Reviewing AI | Sign-off rung | Measured by |
|---|---|---|---|---|---|
| 1 to 3 | Bookkeeping and reconciliations | Two bank recs, two GST or sales tax recs | Reviews AI coding suggestions against source | 1, moving to 2 | Review points per job; first planted-error check |
| 4 to 6 | Compliance and returns | Two simple returns from source documents | Reviews AI-drafted workpapers and logs every correction | 2, moving to 3 | Planted-error catch rate; error log quality |
| 7 to 9 | Advisory and meeting preparation | One set of year-end adjustments on paper | Checks AI-drafted meeting packs and client summaries | 3, moving to 4 | Explain test on one advisory file |
| 10 to 12 | Client communication and first review | One refresher of each earlier hand task | First reviewer on a peer’s AI-assisted job | 4, moving to 5 | Review points they raise on others’ work |
Every month: weekly 30-minute explain-your-working session, one planted-error check, error log reviewed.
Learning check
What would you do? Pick one; every answer is explained.
Free to join
Join The Firm free to read the rest
Members get every playbook in full, plus free templates, downloads and tools across the site. Join more than 35,000 accounting professionals. All it takes is your email.
Where we could be wrong
The assumption most likely to be wrong is that an hour a week of senior time is available. In a firm where the partner is the only senior, the explain sessions are the first thing to slip, and the pathway turns back into “use the AI and hope”. If that is you, protect the explain session over everything else in this plan. We also cannot tell you how much hand-work is enough; the numbers above are a starting point, not a researched standard. And as tools improve, planted errors may become the main source of practice, which is a weaker teacher than real ones.
Questions like these are on the agenda for The Firm’s AI in Practice Summit on 11 and 12 November 2026, where “Teach your team how to use AI productively and securely” (Beau Gaudron of Growthwise, Gus McLennan of AI Dojo, James Rose of Content Snare and Travis Bacon of TaxDigital) was among the most-registered sessions. Notes will follow at the summit notes page after the event. For the rollout side, see why teams stop using their AI logins, and for the wider picture, our AI for accounting firms hub.
Frequently asked questions
Should graduates be allowed to use AI from day one?
Yes for drafting emails and finding their way around the firm’s tools, but not on the hand-work list. A graduate who has never done a reconciliation by hand has nothing to check the AI’s version against.
Is reviewing AI output really training, or just checking?
It becomes training when the junior has to prove each correction from a source and log it. Without the proof and the log, it is checking, and the junior learns to trust whatever looks finished.
How do I plant errors without wasting a senior’s time?
Keep a small bank of past jobs where you know the right answer, and change one figure or one treatment. Ten minutes to set up, reusable for every future junior.
What if a junior is faster than the pathway?
Let the evidence move them, not the calendar. If their catch rate and explain test are strong in month four, move them up a rung early and keep the hand-work refreshers.
Does this work for offshore or remote juniors?
Yes, with the explain session on video and the error log where everyone can see it. Remote juniors miss overheard conversations, so the weekly session matters more.
Your one-page summary
Your answers, choices and decisions, on one page
Sponsored
Junior AccountantsTrainingGraduatesSupervisionTeam DevelopmentAI in Practice Summit
Get the editable junior accountant pathway tracker Download the junior accountant pathway tracker
The Excel workbook (.xlsx), ready to fill in. Your profile unlocks every template on The Firm.
Or build it with AI
Open Claude or ChatGPT with the prompt filled in. It builds the workbook for your firm straight away, then asks a few questions to tailor it. No client data needed.
Works without client data. Don't paste client names, tax file numbers or ledgers into an AI tool your firm hasn't approved for them. AI data safety playbook.
