How to use ChatGPT for accountants (and where you really shouldn't)

Every accountant we talk to has tried ChatGPT. Most use it weekly. Fewer have thought carefully about which parts of their work it should touch and which parts it absolutely should not. This piece is that conversation.

ChatGPT (and the equivalent tools from Anthropic, Google, and Microsoft) is a genuinely useful workstation companion for accounting practices. It is also a tool that will cause real problems if used carelessly in a regulated environment where accuracy is billable, GDPR is a live concern, and clients are trusting you with sensitive data.

The goal here is not to tell you to stop using ChatGPT. It is to help you use it well: where it earns its keep, where it does not, what safety rules make the difference, and what comes next when your practice outgrows what a general-purpose chatbot can do.

Where ChatGPT genuinely helps in an accounting practice

These are the workflows where ChatGPT is a real productivity multiplier for accountants, provided the safety rules further down are respected.

**Explaining tax and accounting concepts to clients.** A client asks about the difference between salary and dividends, or about the timing of a tax deadline, or about how R&D relief works. You could write the explanation from scratch every time. Or you could ask ChatGPT to draft a plain-English explanation, edit it in ten seconds, and send it. Saves ten minutes per client message. Multiply by every routine explanation your team writes in a week.

**Drafting standard client communications.** Engagement letter cover notes. Onboarding welcome emails. Reminders for missing information. ChatGPT drafts them in your practice's tone (once you have given it a few examples) and you edit and send. Not a huge time saving per message but consistency and speed across a team of five to ten staff adds up.

**Turning bullet-point notes into finished narrative.** A partner takes rough notes during a client meeting. ChatGPT turns those bullets into a proper file note or a summary email to the client. Twenty minutes of writing becomes two minutes of editing.

**Explaining new legislation to your team.** New Making Tax Digital rules, changes to VAT treatment, updates to filing thresholds. ChatGPT can summarise the change and produce a one-page internal briefing that gets partners and juniors on the same page fast. You verify against the actual legislation (this is important, see below) but the drafting is quick.

**Drafting proposals for new clients.** Someone in your practice wins a new client and needs to send a proposal. ChatGPT drafts the scope, deliverables, and process section based on the practice's standard proposal template plus the specifics you paste in. The pricing and signature block stay human.

**Improving the readability of tricky reports.** You have written a technical report or a management-accounts narrative that is accurate but dense. ChatGPT rewrites it in plainer English while keeping the substance. Clients understand and value your work more when the writing is clear.

**Learning and revising for exams or CPD.** Junior team members can use ChatGPT to test themselves on ATT, ACA, ACCA, or CTA material, get explanations of concepts, and work through practice scenarios. It is a decent study companion (though obviously not a substitute for the official study materials).

Where ChatGPT will cause real problems in an accounting practice

These are the workflows to keep ChatGPT away from, either because the accuracy risk is too high or the compliance risk is unacceptable.

**Actual tax calculations for specific clients.** ChatGPT is bad at arithmetic. It is worse at legally-specific tax computation. It confidently produces numbers that look right and are wrong. If you use ChatGPT to compute someone's tax liability without verifying every step against the actual rules, you will eventually make a filing error that costs the client (and expose you to a professional-conduct issue). Use your tax software or your own working papers for calculations.

**Verifying legislation or citing case law.** ChatGPT will confidently cite a case that does not exist, a section number that is wrong, or a rule that has been superseded. It has been trained on general text, not on live authoritative sources. Never cite ChatGPT's output as the source for a technical position. Use HMRC guidance, the Institute's technical materials, and current tax software as your source of truth.

**Anything involving specific client financial data pasted into the public web interface.** This is the biggest one. Do not paste client bank statements, VAT returns, payroll data, or any personally identifiable information into ChatGPT unless you are using a version specifically approved by your firm's data policy and appropriate DPA is in place. The public ChatGPT product may use your input to improve their models unless you explicitly opt out at the account level. Even with opt-out, there is a reputational and regulatory risk in sending client data to a third-party AI without a proper agreement.

**Direct client communication without human review.** ChatGPT drafts. Humans send. Always. Never automate ChatGPT sending emails to clients from your accounts. The one time it hallucinates a wrong fact, misapplies your tone, or mentions the wrong client is the time you lose the relationship.

**Financial or investment advice framed by ChatGPT.** If you are FCA-regulated (or your equivalent local regulator) and part of your practice offers regulated advice, ChatGPT-generated advice content is a compliance risk. All advice content must be reviewed and signed off by an authorised individual before it goes to a client.

The four safety rules for using ChatGPT in a regulated practice

Practices that use ChatGPT responsibly follow the same four rules. If you adopt these as practice-wide policy, most of the risk goes away.

**Rule 1: never paste identifiable client data into a public chatbot.** Client names, account numbers, tax reference numbers, bank details, addresses, income figures tied to a name, all off limits. If you need ChatGPT to reason about a specific client scenario, anonymise the details first (change names, round numbers, remove identifiers). Or use a version of the tool with an enterprise data agreement in place (ChatGPT Team, ChatGPT Enterprise, Claude for Work, Microsoft Copilot within your tenant).

**Rule 2: never trust ChatGPT on numbers or legal citations without verifying.** Treat every number and every citation as a claim that needs checking. Use your tax software for computation, use the source legislation for citations, use your professional bodies' guidance for technical positions.

**Rule 3: humans review everything client-facing before it sends.** ChatGPT drafts, humans send. Zero exceptions. This one rule prevents the vast majority of AI incidents in professional-services practices.

**Rule 4: keep an audit trail of significant AI-drafted work.** If ChatGPT drafted a proposal, a technical note, or a client letter, log that. If a professional standards question ever comes up later ('who wrote this and what were they relying on?') you can answer it honestly.

The point where a practice outgrows ChatGPT

For individual productivity, ChatGPT and its equivalents are excellent. For actual practice-wide workflow automation, they hit hard limits. You will notice the limits when you start asking questions like:

'Can we get ChatGPT to automatically pull the receipts from each client's inbox and post them to Xero or QuickBooks?' No. That requires integrations ChatGPT does not have.

'Can we set up ChatGPT to chase clients for their missing tax season documents automatically?' No. That requires scheduled, multi-channel outreach ChatGPT is not designed to run.

'Can we get ChatGPT to generate monthly management accounts from each client's ledger data and email them out?' No. That requires reading from your accounting platform, running variance analysis, generating PDFs in your template, and delivering on a schedule.

These are exactly the workflows that custom AI agents built specifically for accounting practices handle. A custom AI agent authenticates into Xero or QuickBooks, reads the data, executes the workflow, and hands off to humans where needed. It runs unattended on a schedule. It integrates with the other tools the practice uses. And it is designed around your specific practice's process, not around what a general-purpose chatbot can be prompted to do.

The good news: your team's ChatGPT usage is not wasted work. The team members who have learned to work well with AI are the exact people who make custom AI agent rollouts successful. The habit of writing clear instructions, checking outputs critically, and iterating on prompts transfers directly to running a custom agent.

Practical starting points for a practice today

If you want to level up your practice's ChatGPT usage in the next two weeks, three practical steps.

**Step 1: get everyone on the same version.** ChatGPT Team or ChatGPT Enterprise (or Claude for Work) with a proper data-processing arrangement in place. Public accounts using client data are a slow-motion incident. Enterprise seats are inexpensive relative to the risk.

**Step 2: write a one-page internal policy.** The four rules above, plus practice-specific additions. Every team member reads it. Every new joiner reads it. This alone prevents most incidents.

**Step 3: create a shared prompt library.** Half a dozen prompts your team uses most: 'draft an engagement letter cover note for client type X', 'summarise these meeting notes into a client email', 'explain concept Y in plain English for a small business owner'. Consistency plus speed, immediately.

Do those three things and your practice gets most of the ChatGPT benefit with almost none of the risk.

When to look at custom AI agents instead

The signal that your practice is ready for a custom AI agent build is usually this: your team is doing the same thing to prompt ChatGPT ten or twenty times a week for the same use case. That repetition is a workflow. And workflows are what custom AI agents automate properly, with the integrations and the safety controls that a general chatbot cannot provide.

Klevere builds custom AI agents specifically for accountancy practices, integrating with Xero, QuickBooks, and the other tools your practice already uses. The most common builds we do for accounting practices are receipt and bill automation, client document chasing during tax season, and automated monthly client reports. Full details on each of those live in the industry page for accountants and in our Xero and QuickBooks workflow guides.

If your team has hit the ceiling on what ChatGPT can do and you want a written opinion on where custom AI would earn back the investment fastest for your practice, book a free AI audit. Thirty minutes on a call, one week for the written roadmap.