We are a Claude consultancy in Sydney, so read our bias accordingly. But the question lands in every partner meeting, the vendors will not answer it straight, and the honest matrix is more useful to us than a sales pitch, because a firm that buys the wrong tool blames the whole category. Prices and features in this space move monthly; treat every claim here as checked at time of writing and verify before signing.
The criteria that actually matter for a practice
Client-data handling: training defaults on business plans, certifications, admin controls
Xero access: native, connector, or copy-paste
Agentic depth: completes multi-step file work, or answers questions about it
Office integration: how it behaves inside Excel, Outlook and the daily grind
Cost per user in AUD at practice scale, including the tools it lets you cancel
Claude
Anthropic's Claude does not train on business-customer data by default and carries SOC 2 Type II and ISO 27001 certifications. Its position in accounting sharpened in May 2026 when Xero shipped a native Claude integration to all 4.5 million subscribers, making it the only assistant of the three living natively inside the ledger most AU and NZ firms use. The bigger differentiator is Claude Cowork, the desktop agent: it reads scoped client folders, drafts chase emails, prepares BAS workpapers and year-end files for review, and runs scheduled tasks. That is firm-side work, not chat. The trade-off is that a general agent needs your templates and procedures encoded as skills, which is setup work: a fixed-fee firm deployment runs $3,500 plus subscriptions.
ChatGPT
OpenAI's ChatGPT is the assistant most staff already use, sometimes with the firm's knowledge. As a generalist reasoner and drafting tool it is excellent, and its business plans offer training-off defaults comparable to competitors. The practice problems are governance and integration: no native Xero presence, agent capabilities that remain browser-centric rather than file-system and practice-folder centric, and the reality that most accounting-firm ChatGPT usage happens on consumer accounts where the governance story is whatever the staff member clicked. A firm that formalises ChatGPT on business terms gets a strong chat layer; it still does not get firm-side batch work across a client base.
Microsoft Copilot
Copilot's case is distribution: it lives inside the Microsoft 365 stack the firm already pays for, and for Excel manipulation, Outlook triage and document drafting inside Word it is genuinely convenient. Its accounting depth is thin, it has no Xero relationship, and its agent story on arbitrary practice workflows trails the other two. For firms whose pain is Office-shaped, Copilot earns its seat; for firms whose pain is compliance-season-shaped, it mostly decorates the edges of the problem.
The verdicts, plainly
Choose Copilot if the firm lives in Microsoft 365 and wants in-app assistance with minimal change
Choose ChatGPT as a governed generalist chat layer, and formalise the accounts staff are already using
Choose Claude for agentic firm-side work: the desktop agent, the scoped folder model, and the native Xero integration are a combination the others do not currently match
Running two is legitimate: Copilot for Office convenience plus Claude for practice workflows is a common landing point
The questions to ask every vendor, including us
Whichever way the matrix tilts for your firm, five questions cut through the demos. Does the business plan train on our data, and can we see that in writing? What happens to client files the tool touches: where are they processed, and under what agreement? Show me the tool completing our actual worst job, on our template, not a curated demo file. What does review look like: does the output arrive with sources and flags, or as confident prose we have to audit blind? And what is the exit cost if we change our mind in a year: templates, skills and habits should transfer, subscriptions should just stop. A vendor who answers all five without flinching is safe to pilot. A vendor who answers with a roadmap slide is asking you to fund their product development with your tax season, and an Australian practice mid-lodgement-program has no spare season to donate.
The cost frame
Seat prices across all three land within a band that rounds to trivial against practice economics: each is less per month than a single hour of a senior's $180 charge-out time. The real cost question is different: which tool removes billable-hour leakage rather than adding a subscription line? A chat assistant saves minutes per question. An agent that halves preparation on a $900-of-staff-time compliance job saves real money per job, every job. Price the decision on hours returned, not on the subscription delta, and the comparison gets much less close.
If you want this matrix run against your own stack, client mix and staff habits, book a brainstorm call or start from the fixed-fee Claude Cowork setup. We will say plainly where Claude is not the answer: a firm that only needs Excel help should buy Copilot and keep the change.
