AI policy template for accounting firms

This policy sets out how [Firm name] and everyone who works for it may use artificial intelligence (AI) tools in client work and in running the firm. It exists to protect our clients' confidential information, keep our work accurate, and let us use tools that save time without taking on risks we cannot see.

• All partners, employees and contractors, and any outsourced or offshore team members who do work for the firm.
• Every AI tool used for firm work: chat assistants, AI features inside our practice software (ledger, practice management, document capture, meeting notes and proposals), and browser extensions or plug-ins with AI features.
• Firm devices, and personal devices when they are used for firm work.

• A named person is accountable for every piece of work, whatever tool helped produce it.
• Client data only goes into approved tools, on the approved plan, within the limits in section 6.
• Nothing an AI tool produces reaches a client, a regulator or a tax authority until a qualified person has reviewed it.
• We tell clients how we use AI in their work, and we ask first wherever the law or our professional code requires it.
• When in doubt, leave it out and ask the policy owner.

Only tools in this register may be used with client data. A tool that is not listed may only be used with information that is already public and says nothing about a client or our team. The policy owner keeps the register and reviews it every quarter.

Before a tool is added, the policy owner confirms and records:

• The plan does not use our inputs or outputs to train the vendor's models, either by contract or by a setting we have switched off.
• Where the vendor stores and processes our data, and how long it keeps prompts, files and outputs.
• Who at the vendor can access our data, and when.
• The tool supports multi-factor authentication, and single sign-on if we use it.
• We can export and delete our data when we stop using the tool.
• Any connection to our ledger, practice management system, email or document store is read-only unless write access is specifically approved (section 8).

We sort information into three levels. The client AI data checklist sets each level out in detail for the team; the rules are:

• Review every output before it is used. Check figures against the source, check every reference to legislation, rulings or standards against the primary source, and check it answers the client's actual question.
• Do not rely on an AI tool for a tax position, a technical accounting judgement or the conclusion of advice. It can draft; a qualified person decides.
• Keep the working. Save the final output, and what you relied on to check it, to the client file like any other working paper.
• Use the firm's account. Client data never goes into a personal account, even on an approved tool.
• Do not upload a document received under a confidentiality agreement, or one that belongs to a third party, until you have checked the agreement allows it.
• Do not use AI to produce anything that could be mistaken for an original record, such as a statement, receipt, invoice or letter from someone else.
• Recording or transcribing a meeting needs the agreement of everyone in it, asked for at the start. In some places consent is a legal requirement.

Tools that connect to our systems, or take actions on their own, carry more risk than a chat window: they can reach more data, and they can change it.

• Connections are read-only by default. Write access (posting journals, sending email, changing records) needs the policy owner's approval for a named task.
• Each connection uses the narrowest access the vendor offers, is recorded in the register, and is reviewed every quarter.
• Any automated action that reaches a client, a bank or a tax authority is approved by a person before it goes.

Our engagement letters include this clause, or wording to the same effect:

We use software tools, including artificial intelligence tools, to help us deliver our services efficiently. We use them under our AI policy, only on business plans that do not use your information to train their models, and a qualified member of our team reviews all work before it is provided to you. You can ask us at any time which tools we use in your work, or ask us not to use them.

If a client asks us not to use AI in their work, we record it on the client file and in our practice management system, and we follow it.

US tax work: check whether Section 7216 of the Internal Revenue Code requires the client's written consent before tax return information goes to a third-party tool, and obtain it in the required form first.

• Multi-factor authentication is on for every AI tool account.
• Accounts are created and removed by [role]. A leaver loses access on their last day.
• Shared chat links and public workspaces are off, or reviewed monthly by the tool owner.
• AI browser extensions and plug-ins are only installed if they are in the register.

If client data goes into a tool that is not approved, or an output with an error reaches a client, tell the policy owner the same day. Reporting early is always the right call, and nobody is disciplined for promptly reporting their own mistake.

The policy owner records what happened, asks the vendor to delete the data where that is possible, decides (with advice where needed) whether it is a notifiable breach and whether the client must be told, and changes the process so it does not happen again.

Breach obligations depend on where the firm operates. In Australia they include the Notifiable Data Breaches scheme where the Privacy Act applies to the firm; in the United Kingdom, UK GDPR's requirement to report certain breaches to the ICO within 72 hours; in the United States, state breach-notification laws.

• Everyone completes a short induction on this policy before they get access to an approved tool, and a refresher once a year.
• Whoever proposes a new tool walks the team through its settings and limits before it goes live.

The policy owner reviews the register every quarter and this policy at least once a year, and sooner when we adopt a new kind of tool, a regulator or professional body issues guidance, or something goes wrong.

I have read this policy, I understand it, and I will follow it.

This template is a starting point, not legal advice. Your obligations depend on where you practise, your professional body and your clients; take advice on anything you are unsure of.