Technology · Evergreen guide

AI for accounting firms: what actually works

Every vendor says "agentic" now. This is The Firm's living guide to what AI is genuinely doing inside working practices — which tools earn their keep, what to adopt first, and where the traps are.

Updated 22 July 2026 · Built from our reporting and working firms' real builds — every claim links to the coverage behind it.

The state of play, in one paragraph

AI in accounting has split into two layers. In-suite AI — JAX inside Xero, Karbon's AI, Ignition's MCP-powered flows — automates tasks where your data already lives, and is the easy yes. Platform AI — Claude (and Claude Cowork), Microsoft Copilot — is the bigger prize and the bigger decision: an assistant that can work across everything, if you design the guardrails. The firms winning right now (the data is clear) aren't the ones with the most tools — they're the ones that gave AI specific jobs and rebuilt the process around the freed-up hours.

The stacks, honestly compared

Four ways firms are actually deploying AI — and who each one is for.

Claude (Anthropic)

Firms that want an AI colleague, not a feature

The strongest reasoning and the tool practitioners in our community keep building real workflows on — from 40-app firm platforms to one-hour Karbon integrations. Cowork brings it to non-technical staff. The catch: it lives outside your suite, so you design the guardrails.

Microsoft Copilot

Firms that live in Outlook, Excel and Teams

The default for a reason: it inherits your Microsoft 365 security posture and IT sign-off is easy. In our survey coverage, nine in ten firms quietly standardise here. Less capable at open-ended work than Claude; unbeatable at "inside the documents you already have open."

Xero JAX + XeroForce

Firms all-in on the Xero ledger

Auto bank rec and Smart Document Capture are live and genuinely useful; XeroForce (no-code agents) is early access. It's AI where your clients' data already sits — but it's Xero-shaped: expect it to deepen the moat, and watch pricing.

In-suite AI (Karbon AI, Ignition, Dext…)

Every firm, incrementally

Your practice tools are adding AI where you already work — Karbon's triage and drafting, Ignition's MCP proposal flows, and the features landing across FYI, AccountKit, Kloud Connect and Dext. Adopt these first: zero rollout cost, immediate wins, no new vendor risk.

Deciding between the big two? Our Copilot vs Claude comparison goes deeper — including why most firms quietly run both.

The adoption playbook

  1. 1

    Start with exports, not integrations.

    A spreadsheet export dropped into Claude or Copilot answers real questions on day one — contact-data gaps, time-by-task dashboards — with zero connection risk. Most firms should live here for a month before connecting anything.

  2. 2

    Write the client-data policy before the first connector.

    Decide what may leave your systems, which tools are approved, and who signs off. The firms that skip this step are the ones that end up in our stories about AI going wrong — and as we've said before, no AI is ever going to jail for you.

  3. 3

    Adopt in-suite AI immediately, platform AI deliberately.

    Features inside Xero, FYI or Ignition inherit permissions you already trust — turn them on. A platform assistant (Claude, Copilot) is a bigger decision: pilot with two people and five defined jobs, not the whole firm and a vibe.

  4. 4

    Give it jobs, not access.

    The wins in our coverage are specific: month-end close prep, client email triage into work items, meeting notes into actions, template drafting. Pick five jobs, measure the hours, expand from evidence.

  5. 5

    Have a strategy on paper.

    98% of accountants already use AI; only one in five has a formal policy. The gap between those numbers is where the inefficiency — and the risk — lives. A one-page policy beats an AI committee.

Every piece of our AI coverage

20 articles and guides from The Firm on AI in practice — newest first, updated as we publish.

AI for UK accounting practices: ICAEW and ACCA guidance, GDPR, and the practical path

UK accounting practices have more official AI guidance available than most realise: ICAEW has published a full generative AI guide, ACCA has a formal position paper, and the ICO has detailed guidance on AI and data protection. This article distils what those documents actually require, explains how UK GDPR applies when client data touches an AI tool, and lays out a practical adoption path timed against the MTD workload arriving in April 2026.

Trent McLaren · 22 July 2026 · 9 min read

Is AI safe for client data? The accountant's straight answer

The honest answer to whether accountants can use AI with client data: yes, if you use business-tier tools and a handful of non-negotiable rules. The real risk is your process, not the model. We break down what ChatGPT, Claude and Copilot actually do with your data as at July 2026, what your confidentiality obligations require in the US, Australia and the UK, and the five rules that keep you on the right side of both.

Trent McLaren · 22 July 2026 · 10 min read

Common questions

Frequently asked questions

What is the best AI tool for accounting firms?
There's no single answer — it depends where your firm already lives. Microsoft-centric firms get the fastest safe start with Copilot; firms wanting deeper reasoning and custom workflows get more from Claude (and Claude Cowork for non-technical staff); Xero practices should turn on JAX's auto bank reconciliation and document capture first; and everyone should enable the AI already inside their practice tools, whether that's Karbon, FYI, AccountKit or Ignition.
Is AI safe to use with client data?
It can be, if you treat it like any other data processor: use business/enterprise tiers (which exclude training on your data), write a policy covering what may be shared, prefer read-only connections, and keep a human approving anything that goes out. The risk is rarely the model — it is firms adopting tools with no policy at all.
Will AI replace accountants?
The evidence in our coverage points the other way: AI is removing the manual layer (reconciliation, data entry, document chasing) while the advisory, judgement and relationship work grows. The firms pulling ahead are not the ones with the most software — they are the ones that redesigned people and processes around the new capacity.
How should a small firm start with AI?
Three steps: export a spreadsheet and ask questions of it in Claude or Copilot (day one, zero risk); turn on the AI features in tools you already pay for; then pick five recurring jobs — month-end prep, email triage, meeting notes — and pilot a platform assistant on those with two people before rolling out wider.
What is MCP and why does it matter for accountants?
MCP (Model Context Protocol) is the open standard that lets AI assistants talk to your software — Xero, Karbon, Ignition — in plain English, with controlled permissions. It matters because it turns 'AI that answers questions' into 'AI that does the work in your actual systems', with read-only scopes and human sign-off as the safety rail.
How much does AI cost for an accounting firm?
Meaningfully less than the time it returns, if you adopt deliberately: in-suite AI is usually included in subscriptions you already pay; platform assistants run roughly US$20–40 per user per month on business tiers; the real cost is rollout time. Budget for a pilot, a policy and training rather than for software.