Your first AI hire: how to design the role before you fill it

“Our first AI hire” means two different things depending on who says it. One owner means a person: an operations or admin assistant whose day runs through Claude, ChatGPT or Microsoft Copilot. Another means software: an AI assistant set up to do part of a job that a person used to do. Both start in the same place. Before you hire anyone or switch anything on, you have to design the role.

A person whose job is AI-assisted

This is a human hire, usually an operations coordinator, practice assistant or junior bookkeeper, recruited on the expectation that a large share of their output will be drafted by AI and finished by them. The person is the accountable worker. The AI is their tool, the way a spreadsheet is.

An AI assistant deployed into a role

This is a configured tool: a Claude project with your SOPs loaded, a ChatGPT workspace with set instructions, Microsoft Copilot working across your Outlook and Teams. It does a defined slice of work, and a named person checks it. Nobody is hired, but someone’s job changes, because they become its reviewer.

Most firms end up with both. The person version suits work that needs judgement, client contact or chasing. The tool version suits work that is high-volume, repeatable and easy to check. The usual first mistake is hiring a person for what a well-briefed tool could do, or pointing a tool at work that needs someone to pick up the phone.

A traditional job description lists duties. An AI-era role has to be designed task by task, because each task now has two possible doers and always one checker. For every task, answer four questions: what goes in, what comes out, who produces it and who checks it before it leaves the building.

The checking column is the one firms skip, and it is where the time saving either holds or leaks away. If a draft takes the AI ten seconds and the reviewer fifteen minutes to untangle, the task has moved, not shrunk. Design the output format so it is fast to check: a summary with the source attached, a list with each item linked to the document it came from, a draft email with the facts it relied on listed underneath.

The firms with clean SOPs and documented processes are the ones where agents will be able to do real work, and your team will start to become reviewers rather than doers.

Your existing people’s roles move too. Whoever reviews the AI’s work is doing a new job, and it needs time in their week, or it gets done badly at 6pm.

Start with tasks that are high in volume, low in judgement, easy to check and cheap to get wrong once. Good first candidates:

• Chasing documents and drafting reminder emails from a list of outstanding items.
• Turning a recorded process walkthrough into a draft SOP.
• Cleaning and checking client data exported from your practice management system, such as contacts missing an email or mobile number.
• First-draft replies to routine client questions, held for a person to send.
• Summarising a long email thread before a manager picks it up.

Some work never goes to the tool on its own, whoever you hire:

• Signing, lodging or submitting anything to a tax authority.
• The final tax position or advice a client will rely on.
• Accepting a client or an engagement, and any risk rating that sits behind that decision.
• Anything sent to a client without a named person reading it first.
• Putting client data into a tool your firm has not approved, or before the client has agreed to it.

On the last point, the Tax Practitioners Board’s guidance on AI, TPB(GS) 55/2026, issued 22 July 2026, says entering client information into AI tools can be a disclosure to a third party under Code item 6, “depending on how these tools are configured and used”, and recommends getting client permission first. Our guide to client consent for AI has the wording.

The firm owns everything that leaves it, and one named person owns each output before it does. The AI owns nothing.

For Australian tax and BAS agents this is written down. TPB(GS) 55/2026 says practitioners “are still ultimately responsible for the tax agent services they provide to their clients”, that output should be verified and reviewed “throughout each step of the workflow”, and that “Each of these steps should be documented.” Section 35 of the Code Determination already requires anyone providing services on your behalf to be appropriately supervised. US firms will find the same idea in the due diligence standard in Circular 230, and UK firms in Professional Conduct in Relation to Taxation. This is general information, not legal advice; check the guidance from your own regulator or professional body.

That’s what human-in-the-loop means: you let the agent go up to a certain step, then a human approves the final step before it moves on.

In practice, supervision means three habits. The reviewer can see what was AI-drafted. The reviewer has the source next to the draft. And the firm keeps a short log of what went wrong, so the instructions improve. Write these into your one-page AI policy so they survive the first busy month.

If you are hiring a person

Lead with outcomes, not tools. “Every client job has its outstanding documents chased within two working days” tells a candidate more than “proficient in ChatGPT”. Then name the daily tools, say plainly that much of their first draft will be AI-generated, and say who reviews it.

Hire for the checking instinct. The skills that matter are writing a clear process, spotting a figure that does not match its source, and knowing when to stop and ask. A useful interview task: give the candidate an AI-drafted client email with two planted errors and ask them to fix it and explain what they changed. Our piece on hiring your first employee covers the capacity and revenue side of the decision.

If you are deploying an AI assistant

Write a role card, and paste it in as the project or workspace instructions. It covers the purpose in one sentence, the inputs it may use, the data it must never be given, the exact output format, when it should stop and flag something rather than guess, and the name of the person who reviews its work. Test it on five past jobs where you already know the right answer before it touches a live one. If the first project feels too big, choosing the first problem worth handing to AI is the step before this one.

Hiring a first AI assistant was among the most-registered themes for The Firm’s AI in Practice Summit on 11 and 12 November 2026. Natalie Lennon of Two Sides Accounting presents “I hired my first AI assistant, here’s what I learned and mistakes you can avoid”. The session has not happened yet. Notes will be published at the summit notes page after the event, alongside more from our AI for accounting firms hub.

The assumption most likely to be wrong is that a ten-person firm has someone with time to review. If the owner is the only reviewer, the delegation map will bottleneck on them, and the honest first hire may be a person who can review, not a tool that needs reviewing. We have also assumed firms can configure their tools to keep client data out of model training; that depends on the plan and settings you buy, and changes over time. We have not seen insurer or regulator guidance that sets a specific ratio of reviewers to AI-drafted work, so the “one named checker per task” rule is ours, not a standard.

Should my first AI hire be a person or a tool?

Look at where the hours go. If they go on chasing, calling and judgement calls, hire a person who uses AI. If they go on repeatable drafting that is quick to check, configure a tool first and spend the saving on review time.

Can a junior staff member supervise an AI assistant?

They can do the first check on low-risk work, such as chasing emails or data clean-up. Anything a client relies on for tax or financial decisions needs someone with the knowledge to judge it, and for Australian tax agent services the registered practitioner stays responsible.

Do I need to tell clients that an AI assistant works on their file?

If client information goes into the tool, generally yes. The TPB’s guidance recommends client permission before client information is entered into AI tools where that is a disclosure, and an engagement letter clause is the usual way to get it.

What should go in an AI assistant’s instructions?

The six parts of the role card above, on one page. The test is whether a second reviewer reading it cold would understand the job, and it gets updated every time the same mistake is logged twice.

How do I know the delegation is working?

Track review time per item and the error rate for a month. If review takes longer than doing the task by hand, take the task back.