AI for accountants: what to use in your practice, and what to recommend to clients

AI for accountants comes in two forms: the AI built into your practice and ledger software, and general assistants like ChatGPT, Claude and Microsoft Copilot, which can now connect to Xero and answer questions from live client data. Used on a business plan, with a person checking everything before it goes out, they take real time out of chasing, drafting and month-end prep. This guide covers what to use in your own practice, what to recommend to your clients, and where it stops.

It is written for partners, managers and bookkeepers at UK practices who have tried ChatGPT, want to use AI properly, and keep being asked by clients which tools they should be using.

Most practices already use it somewhere. In research for Xero by Cebr and Censuswide (515 UK accountants and bookkeepers, August 2025), 98% of practices said they used AI in some way for day-to-day tasks, and around 47% used it at least weekly. The useful jobs share one feature: AI produces a first draft or a first pass, and someone qualified checks it.

Jobs that work well today:

• Bank reconciliation prep. The AI in your ledger suggests categories and matches. A general assistant turns the leftover unexplained transactions into a clear, numbered query list for the client.
• Chasing missing records. Give it the list of what's outstanding per client and it drafts a polite, specific chaser for each, in your house style.
• Drafting client emails. Explaining why a payment on account has gone up, or what a Making Tax Digital deadline means for them, from your notes, in plain English.
• Month-end and management-report commentary. Once you have reviewed the figures, it drafts the narrative ("revenue up 12% on last quarter, driven by…") for you to edit rather than write from scratch.
• Summarising long threads, meeting notes and lengthy documents into the three points that matter.
• Asking Xero questions in plain English. With a connector switched on: "which invoices for this client are more than 60 days overdue?" or "what moved in the P&L compared with last quarter?"

Where it is weak: technical answers from a general chatbot. ICAEW's generative AI guidance points out that many tools can't draw on the latest legislative developments and that public models may simply be wrong. Use AI to draft, find and summarise; don't use it as the source of the tax answer you put your name to.

Think in two layers rather than a long list of apps.

Layer one: the AI inside your software. Your ledger and practice software are adding AI features directly. Xero's is called JAX, which Xero describes as automating work so the books stay current. Because it works inside the accounting data, it is the safest place to start, and it may already be in what you pay for.

Layer two: one general assistant for the team. This is for writing, summarising and questions that cross systems. Pick by where your team already works:

• You run on Microsoft 365: Microsoft Copilot is the natural fit. It already reaches Outlook, Teams and SharePoint, and your admin controls it centrally.
• Your team already uses ChatGPT: stay with it, but move everyone onto a Business plan so client content is covered by business terms and managed in one place.
• You want the widest choice of ready-made connections, including Xero: Claude on a Team plan is the simplest to connect.

Pick one and standardise, then give staff a short written policy on what can and can't go into it. Staff will use AI either way: in Microsoft UK research with Censuswide (October 2025), 71% of UK employees said they had used AI tools at work that their employer hadn't approved. In a practice, that usually means client data in a personal account.

On the right plan, with sensible rules, yes. On a personal account, treat the answer as no. ICAEW's guidance on generative AI and ethics is direct: respecting client confidentiality means not loading confidential information into public generative AI tools, because you have limited control over who sees it, how it is secured and how long it is kept.

The difference between plans is real and checkable:

• ChatGPT: OpenAI says Business and Enterprise data isn't used for training by default. On personal plans (Free, Plus, Pro) it can be, unless each person turns off "Improve the model for everyone".
• Claude: Anthropic says it won't use inputs or outputs from its commercial products (Team, Enterprise, API) to train its models by default. On personal plans (Free, Pro, Max) it depends on each person's Model Improvement setting.
• Microsoft Copilot used with a work account is covered by Microsoft's data protection addendum, and Microsoft says prompts, responses and the Microsoft 365 data it accesses aren't used to train foundation models.

Training is not the only question. Under UK GDPR your practice is still responsible for client personal data you put into any tool, so check the provider's data processing terms cover your use, keep a note of which tools hold what, and limit who in the firm can use connected tools on which clients. Anonymising before you paste helps, but a combination of details can still identify a client.

When you connect an assistant to another service, such as Xero, that service's own terms apply to its data too. Xero says that through its Claude connector, Xero data is never used to train AI and access is session only.

Out of the box, an AI assistant only knows what you type or paste into it. Connecting it gives it a secure, signed-in link to another system, so it can look things up there itself. Each provider uses a different word: connectors in Claude, plugins (previously apps) in ChatGPT, connectors in Copilot.

Underneath, most of them now use the same open standard, the Model Context Protocol (MCP). Anthropic created it in 2024 and handed it to a Linux Foundation project in December 2025, co-founded with OpenAI and Block. The easiest way to think of it is USB-C for AI: one plug shape, so any assistant can connect to any system that offers it. You don't need to know how it works, only that a system with an MCP connection can be plugged into the assistant your practice uses.

Where Xero stands today:

• Claude: Xero has an official connector. Xero says it is free to add, needs an active Xero subscription and any Claude account, and is read-only: Claude "can't edit invoices, post transactions, or change anything in your books".
• ChatGPT: Xero announced its ChatGPT plugin as live in September 2026, for questions and insights on live Xero data.
• Microsoft Copilot: Xero is in Microsoft's catalogue of ready-made connectors. An admin switches it on, and each user needs a paid Microsoft 365 Copilot licence.

One detail that matters for practices: Xero says accountants and bookkeepers work in one client organisation at a time through these integrations. That suits client-by-client work, such as preparing for a review meeting, rather than questions across your whole client list.

Read-only first. Start every connection reading, not writing. The assistants let you control this per connection: Claude lets you set each action to allow, ask first or block, and ChatGPT asks before anything that changes data. Let it look things up and draft; keep posting, sending and changing records behind a person's approval.

We don't repeat the click-by-click setup here. The step-by-step for each assistant, including connecting Xero, is in our free guide below.

You are. ICAEW's guidance puts it plainly: generative AI tools "are not fully qualified accountants", and you are expected to understand a tool's risks and limitations and to apply professional scepticism to what it produces. Nothing about a connector changes that.

The working rule that holds up is simple: the AI prepares, a person approves. In practice:

• Review AI output the way you would review a junior's first draft, with the same sign-off.
• Keep anything that goes to a client, to HMRC or into the ledger behind a person's approval.
• Check figures in AI-drafted commentary back to the reviewed numbers, every time.
• Ask your professional indemnity insurer whether they expect anything specific about AI use.

There is one security risk worth understanding. An assistant that can read your inbox, see client data and send messages can be tricked by a cleverly worded email into sending data out. The developer Simon Willison calls this the "lethal trifecta". Only connect services you trust, and never give one assistant all three without a person approving what it sends.

Less than most practice software. Prices at the time of writing (checked 6 October 2026):

• AI in your ledger: check what your current plan includes before paying for anything else.
• Microsoft 365 Copilot Business (firms up to 300 users): £16.10 per user a month, paid yearly, ex VAT. Microsoft is running a first-year offer at £13.80 until 31 December 2026. The free Copilot Chat can't use connectors to systems like Xero.
• Claude Team: $20 per person a month paid yearly, or $25 monthly for a standard seat, priced in US dollars.
• ChatGPT Business: also priced per person per month; check OpenAI's pricing page for the current UK price.
• Xero's Claude connector: free to add.

The bigger cost is time: setting it up properly, tidying who can see what, and reviewing output. Most practices are better off licensing the people who will use it daily than buying a seat for everyone on day one.

Clients are using AI whether you advise on it or not, and not always well. In a Censuswide survey for Dext of 500 UK accountants and bookkeepers (December 2025), half said businesses had suffered financial losses from incorrect AI advice. That makes "what should we be using?" an advisory conversation you are well placed to lead, and it usually ends in better records, which means better numbers reaching you.

A sensible recommendation depends on the size of the business:

• Sole traders and micro businesses: the AI already in their accounting software (receipt capture, categorisation), plus one assistant on a paid plan if they want help with emails and admin. The one rule: no bank statements or customer details in a personal ChatGPT account with model training left on.
• Small teams on Microsoft 365 or Google Workspace: one assistant on a business plan, connected read-only to Xero and their files. Before connecting anything, tidy who can see what in SharePoint or Google Drive, because the assistant sees whatever each person can see, salary spreadsheets included. Add a one-page staff AI policy.
• Growing firms with several systems (Xero plus a CRM, a job or project system and a few critical spreadsheets): the same starting point, with a clear warning that ready-made connectors will answer questions about one system at a time. The questions owners most want answered, about margin, cash and capacity, usually need several systems joined up.

Ready-made connectors give an assistant access to data. They don't give it an understanding of the business. The gaps show up quickly in owner-managed firms:

• It can't price work the way they do. It can read a price list, but it won't know their discounts, margins, minimum charges or what's included unless those rules are built in somewhere.
• The data is split across systems. Each connector is a separate pipe. Nothing joins Xero's view of a customer to the CRM's or the job system's, so "which customers are we making the least money on?" usually can't be answered, and two people asking the same question can get two different answers.
• Some systems have no connector at all. When we checked on 3 October 2026, Sage Accounting, Sage 50, FreeAgent, Simpro, Tradify, Commusoft and Procore had no ready-made connector in Claude, ChatGPT or Copilot's catalogues, and nor does the pricing spreadsheet on the shared drive.
• Heavy use has limits. Xero now tiers and charges for heavier use of its API, and an assistant re-fetching everything for every question uses that up.

This is the work we do. We build custom connections (MCP servers) and tools around how a business actually works: a quoting tool that prices from the client's own rates and rules, or one central store that joins Xero with their CRM, job system and spreadsheets so every assistant gets the same answer. They work with Claude, ChatGPT or Copilot, start read-only, and keep a person approving anything that changes data. For one UK specialist supplier, pricing a quote went from about 1 to 2 hours to about 5 minutes.

If some of your clients sound like this, here is how working with us as an accountant or bookkeeper works. You stay their adviser; we build the system.

• Choose one assistant and put everyone who will use it on a business plan.
• Write a one-page policy: which tool, what client data can go in, and who approves output.
• Pick one job to start with, such as chasing missing records, where you can say exactly what goes in and what should come out.
• Connect Xero read-only and try it on your own practice's books before any client's.

Get the free step-by-step guide to connecting ChatGPT, Claude or Copilot to Xero and your other systems → It covers which plan you need, what happens to your data, and what to sort out before you connect anything. Free: enter your name and email to open it.