Google Gemma 4 Just Launched — And It Might Solve Finance's Biggest AI Privacy Problem
There's a tension that most finance professionals who use AI haven't fully resolved yet.
On one hand, AI tools are genuinely useful for finance work — drafting, summarising, reviewing contracts, analysing data. On the other hand, finance sits at the centre of an organisation's most sensitive information: payroll data, board papers, funding agreements, audit findings, legal correspondence. Putting that information into a cloud-based AI service creates a data question that most organisations haven't answered clearly.
Google's release of Gemma 4 on 2 April 2026 is worth understanding in this context. It doesn't resolve the tension entirely — but it changes the practical options available to finance teams in a meaningful way.
What Gemma 4 Actually Is
Gemma 4 is Google DeepMind's latest open-weight AI model family. It was built on the same research foundation as Google's proprietary Gemini 3 models, and released under an Apache 2.0 licence — meaning it's free to use, modify, and deploy commercially without restriction.
What makes it different from Claude, ChatGPT, or Gemini as most people use them is this: Gemma 4 can run entirely on your own hardware, with no data leaving your device or network.
Released: 2 April 2026 | Licence: Apache 2.0 (free, commercial use permitted)
Four model sizes: E2B (smartphone/edge), E4B (laptop), 26B MoE (workstation), 31B Dense (workstation/server)
The 26B model runs on a standard laptop or desktop with 16–18GB RAM in 4-bit quantisation. The 31B model needs around 20GB RAM in 4-bit. Both run without internet access once downloaded.
Benchmark performance: the 31B model ranks #3 among all open-weight models globally on the Arena AI leaderboard as of launch.
The Privacy Problem with Cloud AI in Finance
When you paste a document into Claude.ai, ChatGPT, or Gemini, that text is processed on servers operated by the respective companies. Most enterprise-grade subscriptions include data processing agreements and contractual commitments that your data won't be used for model training — and those commitments matter and are worth taking seriously.
But for many NFPs, NDIS providers, and SMEs, the reality is that staff are using personal or standard accounts rather than enterprise tiers. And even with enterprise agreements in place, the question of whether sensitive financial data should leave the organisation's own systems at all is a legitimate governance question — one that boards and audit committees are increasingly asking.
The specific categories that give finance teams pause include payroll data (individual salaries, bank details, leave balances), board and audit papers, legal correspondence and settlement documents, funding agreements with confidential commercial terms, and any information covered by confidentiality clauses in contracts.
The practical response from most finance teams so far has been to de-identify or sanitise documents before using AI — removing names, replacing specific figures with representative ones, working with the structure rather than the raw data. That's a reasonable approach. But it adds friction and creates its own risk of not sanitising thoroughly enough.
What Running a Local Model Changes
A local model like Gemma 4 processes everything on your own hardware. Nothing leaves the device. There's no API call to an external server, no data processing agreement to navigate, no question about what happens to the text you've pasted in.
That changes the privacy calculus meaningfully for specific use cases — the ones where the sensitivity of the document has been the barrier to using AI assistance at all.
For a finance team, the most relevant applications are reviewing actual payroll variance data without sanitising it first, working with real board paper drafts including specific figures and context, reviewing contracts with identifiable commercial terms, and preparing sensitive correspondence where the specifics matter and stripping them out defeats the purpose.
The quality trade-off is real: Gemma 4 at 26B or 31B is capable but not at the level of frontier models like Claude Opus or GPT-5 for complex reasoning tasks. For summarising, drafting, reviewing, and structuring — the core finance use cases — it performs well. For nuanced analysis or tasks requiring deep contextual reasoning, the gap is more noticeable.
How to Actually Get Started
The practical barrier to running a local model has dropped significantly in 2026. The tool I'd point most finance professionals toward is LM Studio — a desktop application for Mac, Windows, and Linux that handles downloading, configuring, and running models through a simple chat interface. No command line required.
The steps are: download LM Studio, search for Gemma 4 in the model library, download the 26B or 31B model in 4-bit quantisation (around 16–20GB), and start a local server. The interface looks and works similarly to a standard AI chat tool — the difference is that the processing happens on your machine.
Hardware requirements for the 26B model at 4-bit: a Mac with Apple Silicon (M1 or later, 16GB+ unified memory) or a Windows/Linux machine with 16GB+ RAM and a modern GPU handles it well. The 31B model is more demanding — suited to M2/M3 Pro Macs or machines with a dedicated GPU. The smaller E4B model runs on most modern laptops and is a reasonable starting point for lighter tasks.
What This Doesn't Solve
Local models don't eliminate all AI privacy risk. They address the data-leaving-your-device problem, but they don't address the governance question of who in the organisation is using AI for what, under what policy, with what oversight.
If your NFP board has started asking about AI governance — as I wrote earlier this week — a local model is part of the answer, not the whole answer. The policy and accountability layer still needs to exist.
They also don't address the risk of AI-generated errors in sensitive documents. A local model can hallucinate or produce inaccurate output just as a cloud model can. Human review of AI-generated content remains non-negotiable in finance, regardless of where the processing happens.
If your organisation is thinking about how to use AI for finance work in a way that addresses data privacy and governance — not just efficiency — PFL is happy to work through that with you.
Talk to PFL →
Comments
Post a Comment