Stop experimenting with AI: How AI operational skills are defining the accounting profession
Firms that move fastest are not ones that dabble with clever prompts. They’re the ones with repeatable workflows, and outputs that turn AI into a controlled, client-facing advantage.
Too many practices are still treating AI as a toy with which to play around.
Perhaps one person uses ChatGPT for emails. Maybe somebody else is testing Microsoft Copilot in Excel. Another asks a practice tool to draft a sentence of commentary.
This might be useful, but it isn’t strategic. And it needs to be.
In this article, we take a look at moving from the experimentation mindset to one that’s firmly operational. Here’s what we discuss:
AI assistants vs AI agents: Know the difference
A critical distinction is understanding the difference between AI assistants and AI agents.
Assistants respond to a prompt. They might draft a variance note, summarise a client email, rewrite a paragraph or explain a movement.
They make a single task faster, but the human still decides the sequence, finds the source data, checks completeness, determines the next action and carries the work between systems.
Agents are a whole different level of sophistication.
An agent is designed around a goal and a repeatable sequence, and works largely autonomously.
In a month-end workflow, that could mean pulling bank feed transactions, matching predictable items, flagging unreconciled entries, comparing current month results with prior periods, drafting commentary, preparing client questions, creating supporting schedules—and then notifying a reviewer.
The accountant or bookkeeper is still accountable, but the role moves from manual execution, to workflow direction, review and judgement.
That’s why this isn’t a technology story alone. It’s a professional development story.
The people who progress in accounting today are those who can map a process, define the controls, specify the data, set review points and explain the outcome to a client.
It’s not just professional judgement. It’s structured professional judgement.
Start where AI is already reliable
Current high-value use cases for AI include month-end commentary, board-pack narratives, quarterly update preparation, cashflow discussion notes, AP and AR coding support, journal entry risk spotting, working paper preparation and movement analysis.
These are areas where AI can convert raw data into a first draft, highlight anomalies, group issues and create a clearer starting point for review.
But it does not mean trusting AI blindly with numbers.
It means using it to reduce the administrative drag around the professional work.
A draft variance note isn’t the final advice. A suggested coding category isn’t a signed-off posting. A client-facing summary isn’t a substitute for understanding the client.
But the first pass matters because it frees time for the work clients actually notice: clarity, reassurance, prioritisation and judgement.
The missing AI ingredient is structure
So, we can say in summary: Most firms are not short of curiosity about AI. They are short of structure.
And care must be taken. Without common prompts, shared templates, workflow maps and review rules, AI can increase inconsistency.
One person’s output is formal, another’s is chatty, another’s misses the commercial point. Review time rises because no one has agreed what good looks like.
A structured approach should start with a prompt framework.
Every prompt needs a clear goal, relevant context, the source material to be used and defined expectations for tone, length, format and audience.
For example: ask AI to produce a plain-English client note, using this month’s profit and loss movements and prior-period comparatives, no more than 200 words, with three client questions and a separate list of items requiring review.
That’s materially stronger than asking it to “write a summary”.
What’s more, practices should aim to create an AI file for each significant AI-supported workflow.
This should record the purpose of the AI use, data inputs, model or tool version, prompt structure, known risks, testing evidence, accuracy over time, reviewer comments and final sign-off.
This mirrors the compliance mindset accountants and bookkeepers already use. It also aligns with the direction of AI governance: demonstrate accountability, protect personal and confidential data, document the process and keep a human in control.
Next steps: Build the 5-step AI agent workflow
A practical agent workflow has five steps:
- Intake: gather the data or documents
- Check: validate completeness, scope and exceptions
- Do: perform the calculation, transformation, classification or drafting
- Document: create the audit trail and explain the reasoning
- Notify: send the output to a human reviewer
This pattern is simple enough for every practice to use and disciplined enough to prevent AI becoming chaotic.
It also points to the new roles emerging inside AI-enabled practices.
An AI librarian manages prompt templates and agent instructions. A workflow owner ensures the human and AI steps remain efficient. A data quality lead protects the inputs. A model risk approver tests agents and signs off risk. A client communication lead ensures outputs are clear, accurate and consistent with the firm’s tone.
Smaller firms may combine these roles, but they still need the responsibilities.
Final thoughts: Your 30-day plan
Choose one workflow, not ten. Month-end commentary or quarterly update preparation is ideal because it’s recurring, visible and rich in narrative value.
Map the steps as they happen today. Identify where AI can draft, compare, classify or flag. Write one approved prompt template. Run it on past data. Compare the output with the final work you actually delivered.
Record known failure points. Add a human review checkpoint before anything reaches a client.
Then measure three things: time saved, review time added and output quality.
This is the difference between playing with AI and becoming AI-enabled.
The firms that do this now will build repeatable capacity, safer delivery and stronger client conversations. The firms that wait will still be busy, but increasingly surrounded by competitors who can deliver faster, clearer and more commercially useful work.
Frequently asked questions
Agentic AI refers to systems designed to autonomously work towards a goal across multiple steps, rather than simply responding to one prompt. In bookkeeping, that could mean an AI-supported month-end workflow that pulls data, checks completeness, drafts commentary, flags anomalies and notifies a human reviewer.
Use approved tools, minimise the client data entered, avoid unnecessary personal or confidential information, document the purpose of AI use, retain prompts and outputs, and keep a human review step before client-facing work. An “AI File” creates an audit trail for this process.
Start with repeatable, reviewable workflows such as month-end variance notes, quarterly update preparation, working paper drafts, AP and AR coding checks, cash flow commentary and client question lists. These use cases benefit from drafting and diagnostics while leaving judgement with the practitioner.
AI will remove parts of the manual workload, but it does not replace professional judgement, client empathy, prioritisation, ethical responsibility or commercial interpretation. Bookkeepers who learn to orchestrate AI workflows will be better placed than those who only perform manual tasks.
Pick one recurring workflow, create one standard prompt, test it on past work, document the risks, decide who reviews the output and track whether the process saves time without reducing quality. Avoid rolling AI across the firm before the first workflow is controlled.
Download the AI action workbook
Start with one simple task. Follow the steps. See progress in 30 days.
