Gemini, Grok, DeepSeek and the rest: do accounting firms need a fourth AI?

Let's save you the read: no, your firm does not need a fourth AI. If you're already running Claude, ChatGPT or Copilot properly (deployed on a business plan, with a policy, with people actually using it), adding another chatbot to the pile solves nothing. We've written before about what happens when every partner picks their own tool, and it isn't pretty.

But there's one big asterisk, and it matters for a lot of firms: if your firm lives in Google Workspace, Gemini isn't a fourth AI. It's your first. The same logic that makes Copilot the default answer for Microsoft 365 firms (it's already in the suite, already under your admin controls, already covered by your existing agreement) applies to Gemini for Google firms. The noise is all about the big three. The Workspace firms have been quietly left out of that conversation, and they shouldn't be.

So here's the honest sweep of everything outside Claude, ChatGPT and Copilot: what Gemini actually offers a firm, the one Google tool almost nobody is using that they should be, and a plain-English take on the challengers.

Gemini: the Copilot argument, for Google firms

The case for Copilot in a Microsoft firm was never "it's the best model." It was "it's already there." We made that argument in full in our Copilot piece, and Gemini earns the identical argument on the Google side, arguably a stronger one, because of how Google prices it.

Since early 2025, Google has bundled Gemini directly into Workspace Business and Enterprise plans rather than selling it as a separate add-on. In the US, Business Standard runs at roughly US$14 per user per month (around A$19.80 in Australia at full price, though Google runs aggressive intro discounts). There's no separate US$30-a-seat AI licence decision to agonise over, which is exactly the decision that stalls Copilot rollouts in Microsoft firms. If you're on Business Standard or above, you likely already have Gemini and simply haven't turned your attention to it.

What you get in practice:

• Gmail: drafting, summarising long client threads, and "help me write" that understands the thread context, genuinely useful for the endless client correspondence that eats a manager's week.
• Sheets: formula generation, data cleanup and "explain this data" prompts. Not a replacement for your working paper discipline, but a real accelerant for the ad-hoc analysis firms do constantly.
• Docs and Meet: drafting engagement letters and file notes from a prompt, and meeting summaries with action items in Meet, the same territory Copilot covers in Teams.
• Admin controls: Gemini sits under your Workspace admin console, so access, data regions and audit sit where your IT governance already lives.

On the question that matters most for a firm (client confidentiality), Google's position at business tier is clear: Workspace customer data is not used to train Gemini models and prompts aren't reviewed by humans, under the same enterprise data protections as Gmail and Drive. That's the same class of commitment Microsoft makes for Copilot and Anthropic and OpenAI make on their business plans. As always: that protection lives at the business tier. A staff member using a personal Gemini account is outside all of it.

Now the honest limits. Gemini's raw reasoning on complex technical work (multi-step tax analysis, standards interpretation, long-document review) has generally trailed the best from Anthropic and OpenAI, though the gap moves with every model release. And like Copilot, the in-suite features are grounded in your data, which means the output quality depends heavily on the state of your Drive. A firm with a chaotic shared drive gets chaotic answers.

NotebookLM: the sleeper tool nobody's firm is using

Here's the part of the Google stack that deserves far more attention than it gets. NotebookLM (which Google has recently been rebranding as Gemini Notebook, though nobody calls it that yet) does one thing, and does it in exactly the shape accounting work needs: you upload your sources, and it answers questions grounded only in those sources, with citations back to the passage.

That's a different proposition from a general chatbot. A chatbot answers from everything it was trained on, which is where hallucinated section references come from. NotebookLM answers from the fifty documents you gave it, and shows you where in those documents the answer came from.

The firm use-cases write themselves:

• Load the relevant accounting standards, the ATO or HMRC or IRS guidance, and your firm's technical memos into one notebook, then interrogate it. "What does the standard actually say about lease modifications?" gets an answer with a pinpoint citation you can verify in seconds.
• Load a client's engagement file (prior-year financials, the trust deed, the loan agreements, the correspondence) and brief yourself before a meeting by asking questions instead of re-reading everything.
• Onboarding: a notebook of your firm's procedures manual and precedent documents becomes a self-serve answer machine for graduates.

Since 2025, NotebookLM has been a Workspace core service for Business and Enterprise customers, which means it carries the same enterprise-grade data protections as the rest of the suite: your sources and queries aren't used to train Google's models. For a Google firm, that makes it about as low-risk an AI deployment as exists.

The caveat, and it's non-negotiable: source-grounded does not mean infallible. NotebookLM can still misread, over-summarise or miss the exception in paragraph 47. The citations make verification fast. They don't make it optional. Treat it like a very quick junior who read the file: enormously useful, never signed off unreviewed.

The challengers, quickly and honestly

Grok. xAI's model is capable, and xAI now sells business and enterprise tiers with the standard promises: business data not used for training, deletion windows, enterprise controls. But the governance story is thin next to the established vendors: consumer Grok conversations are used for model training by default, the compliance and certification track record is short, and European regulators have been actively probing xAI's use of X data for training. Nothing about a firm's workload requires taking that on. If Grok's enterprise offering matures, revisit; today, there's no professional case that Claude, ChatGPT or Copilot doesn't already cover with less risk.

DeepSeek. The model is impressive and the price is irresistible, which is exactly why this needs saying plainly: DeepSeek's consumer service stores user data (prompts included) on servers in China, where the law can compel disclosure to the state. It has been banned from government devices in the United States, Australia, and a growing list of other jurisdictions for precisely that reason. If it's not fit for a public servant's phone, it is not fit for your clients' tax files. The one legitimate caveat: DeepSeek's open-weight models can be run through Western hosting providers or on your own infrastructure, where the China data-residency issue doesn't apply, but that's an IT project for firms with real technical capability, not something to reach for because the app is free. For everyone else: admire it from a distance.

Everything else (Mistral, Perplexity, the vertical "AI for accountants" wrappers) mostly repackages the same handful of frontier models with a workflow on top. Some of those workflows are good. But evaluate them as software purchases, not as a fourth AI.

So what's the stack?

The decision tree is shorter than the vendor marketing suggests. If your firm runs on Microsoft 365, your suite AI is Copilot and your reasoning question is whether to add Claude or ChatGPT on top. That's the three-way comparison we've written up in full. If your firm runs on Google Workspace, swap Copilot for Gemini in that logic, switch on NotebookLM tomorrow, and ask the same second question: does our technical work justify adding one of the frontier chat tools alongside? For many Google firms the honest answer is "Gemini plus NotebookLM covers 80% of it. Add one more only when you hit the ceiling." One suite AI, at most one frontier assistant, a real policy, and actual training. That's the stack. Our broader guidance on rolling it out lives in the AI for accounting firms hub.

Frequently asked questions

Is Gemini good for accountants?

Yes, with the same framing as Copilot. Its strength is that it lives inside Gmail, Docs, Sheets and Meet under your existing Workspace admin controls, and it's bundled into Business plans rather than sold as an expensive add-on. It's the right first AI for a Google Workspace firm. For heavy technical reasoning, the frontier tools from Anthropic and OpenAI have generally had the edge, so some firms pair Gemini with one of them.

Should a Google Workspace firm choose Gemini or Copilot?

Gemini, almost always. Copilot's entire value proposition is deep integration with Microsoft 365. Bolting it onto a Google firm means buying Microsoft licences you don't otherwise need and losing the in-suite grounding that makes it worthwhile. The suite you already run picks your suite AI for you; spend your evaluation energy on whether to add a frontier assistant alongside it.

Is DeepSeek safe for client work?

No. DeepSeek's consumer app and web service store data, including your prompts, on servers in China, and it has been banned on government devices in the US, Australia and elsewhere on data-security grounds. Client information should never go into it. The narrow exception is the open-weight models self-hosted or run through a Western provider, a defensible IT project for a technically capable firm, but not the app on someone's phone.

What is NotebookLM good for in an accounting firm?

Source-grounded research: upload accounting standards, tax office guidance, firm manuals or a client's engagement file, and ask questions that get answered only from those documents, with citations. It's included with Workspace Business plans as a core service with enterprise data protections. Answers still need verification against the cited passage, but that takes seconds, not hours.

Does my firm need more than one AI?

Most firms need exactly one done properly, and at most two: the AI native to your suite (Copilot for Microsoft, Gemini for Google) for everyday email, documents and meetings, plus optionally one frontier assistant for heavy technical work. Beyond that you're buying overlap, fragmenting your data governance, and creating the tool-silo problem that quietly kills firm-wide adoption. Consolidate before you add.