What is the best AI for accountants?

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, XBert or Ignition.

Can bookkeepers use AI, and where does it help most?

Yes, and mostly inside the ledger rather than in a chat window: auto bank reconciliation and coding suggestions, document capture that reads and codes supplier bills, and data-quality review that flags miscoded or duplicate items before month-end. The job shifts from keying to reviewing exceptions, so track the exception rate rather than the hours saved.

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; the adoption playbook above links a free builder if you need one.

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.

What are an accountant's professional obligations when using AI?

The ones you already have. Confidentiality and competence-and-due-care apply under APES 110 and the TPB Code in Australia, the ICAEW, ACCA and AAT codes in the UK, and the AICPA Code in the US, where Section 7216 also restricts how preparers disclose tax return information. None of them bans AI; all of them hold you responsible for what goes into it and what comes out.

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, and XBert's gateway into practice data — 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.

Technology Evergreen guide

AI for accountants: what actually works

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

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.

01 · Which stack is yours

Four questions, one recommendation

Nothing is sent anywhere

Where does your team already work all day?

Which ledger runs most of your client base?

Who would own this inside the firm?

What is the first job you want it to do?

0 of 4 answered

Your recommendation appears here

Answer all four and we will name the stack to start with, the first move to make this week, and the one worth running alongside it.

02 · The stacks, honestly compared

Where each one sits

Five ways firms are actually deploying AI

Reading the map

Start at the bottom of this list: features inside tools you already pay for, inheriting permissions you already trust. Each one further down the page buys more reach across systems and more capability — and costs you more policy, pilots and sign-off. Nothing here is exclusive; most firms end up with the in-suite features plus one thing further out. Deciding between the big two? Our Copilot vs Claude comparison goes deeper.

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

Every firm, incrementally

Your practice tools are adding AI where you already work — Karbon's triage and drafting, Ignition's MCP proposal flows, Vinyl turning client meetings into notes, actions and drafted proposals, XBert auditing client data quality and now serving it to Claude or ChatGPT read-only through its MCP gateway, and the features landing across FYI, AccountKit, Kloud Connect and Dext. Adopt these first: zero rollout cost, immediate wins, no new vendor risk.

Intuit Assist

Firms whose book runs on QuickBooks

The same bet as JAX on the other ledger: categorisation, reconciliation prompts and drafted client comms, shipped into QuickBooks Online and QuickBooks Online Accountant. Nothing to roll out and nothing new to approve — but the same lock-in question applies, and the useful features keep arriving on the higher tiers.

Xero — JAX

Firms all-in on the Xero ledger

Auto bank rec and Smart Document Capture are live and genuinely useful; XeroForce, its no-code agent builder, is still 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.

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.”

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.

03 · The adoption playbook

Five steps, in order

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.
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.

Generate one free with the AI policy builder → Built by Vinyl & Strategic Group. Free, answers are not stored, and it returns a board-ready PDF.
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.
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.
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.

04 · AI for bookkeepers

Where it helps a bookkeeper most

Most of a bookkeeper's AI value sits inside the ledger, not in a chat window. Four jobs, in the order most practices adopt them:

Bank reconciliation and coding

Xero's JAX and the ledger's own rules now suggest and auto-reconcile routine transactions. The job moves from keying to reviewing exceptions, so measure the exception rate, not the hours saved.

Document capture

Capture tools such as Dext read, code and publish supplier bills and receipts. Set the supplier rules once and they compound every month after.

Dext in the directory →

Data-quality review

Review tools such as XBert flag miscoded, duplicate and unreconciled items before month-end, so the file an accountant inherits is already clean.

XBert in the directory →

The client question list

A platform assistant turns the list of unknown transactions into the 'what is this?' email to the client. It is the most disliked job in bookkeeping, and it now takes minutes.

05 · Risk and professional standards

The rules you already follow still apply

No regulator has banned AI in practice, and none has given it a free pass. The obligations you already carry apply to anything you paste into an assistant.

Australia: The TPB Code of Professional Conduct requires registered tax and BAS agents to keep client information confidential, and APES 110 applies the same confidentiality and competence-and-due-care principles to members of CA ANZ, CPA Australia and the IPA. If the Privacy Act applies to your firm, sending personal information to an AI provider that processes it offshore is a cross-border disclosure under APP 8, so know where your vendor processes data.
United Kingdom: The ICAEW, ACCA and AAT codes carry the same fundamental principles: confidentiality and professional competence. UK GDPR governs any personal data you send to a processor, so treat an AI vendor like one, with a contract, a lawful basis and a record of what goes in.
United States: The AICPA Code's confidential client information rule applies, and Section 7216 of the Internal Revenue Code restricts how tax return preparers use and disclose return information, so check whether sending it to a third-party AI tool needs the client's written consent. Circular 230 still governs practice before the IRS.

The control is the same in every market: business-tier tools that don't train on your data, a written policy, and a person signing off anything that leaves the firm. This is a summary, not legal advice; the rules section of our coverage goes deeper.

Start from ours: the AI policy template for accounting firms and the client AI data checklist, both free to read and edit.

06 · Every piece of our AI coverage

Grouped by what you need it for

33 articles, updated as we publish

The rules

What your regulator and your code actually require

8 pieces

How to safely connect ATO data to the AI tools your firm already uses

The Firm · 15 Oct 2026 · 5 min

BAS agents are asking a question about agentic AI that the regulators have not answered

The Firm · 15 Oct 2026 · 6 min

The AI clause every Australian engagement letter needs

The Firm · 7 Aug 2026 · 5 min

AI is now a Code obligation for every tax practitioner

The Firm · 31 Jul 2026 · 5 min

AI for Canadian accounting firms: CPA Canada's position, PIPEDA, and what's actually allowed

Trent McLaren · 22 Jul 2026 · 9 min

Client consent for AI: what to tell clients, when to ask, and the letter to send

Trent McLaren · 22 Jul 2026 · 9 min

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

Trent McLaren · 22 Jul 2026 · 9 min

AI and the IRS rules: what US accounting firms can and can't do

Trent McLaren · 22 Jul 2026 · 9 min

What firms built

Practitioners on their own builds, with the time and cost

3 pieces

Why I built our AI systems in-house instead of buying them

Ben Walker · 7 Aug 2026 · 6 min

How I built a tax intake tool in forty minutes with eleven prompts

Brian Clare · 31 Jul 2026 · 5 min

How to scope an AI project so it actually ships

Nick Ferguson · 31 Jul 2026 · 5 min