Google’s Own Data Says AI Won’t Replace Your Staff — Here’s What It IS Doing

The fear makes sense. You’ve been watching the headlines for three years now, “AI to replace 300 million jobs,” “White-collar work under threat,” “Your staff are next.” If you run a small business in Australia, you’ve probably had the thought at least once: should I be planning to run this place with fewer people?

New data suggests you can put that particular anxiety down.

Google and DeepMind analysed 15 million real interactions with Gemini, not surveys, not projections, actual usage. What they found doesn’t match the replacement narrative at all. Fewer than 10% of those interactions involved full task automation. The vast majority? People using AI to diagnose problems, troubleshoot processes, learn things, draft content, and summarise information. In other words: people doing their jobs with a better tool in their hands.

What 15 million interactions actually look like

The study is a useful corrective because it cuts through the speculation. Most AI discourse is built on forecasts. This was observation. Across 15 million uses of a capable, widely deployed AI system, the dominant pattern was augmentation, not replacement.

People asked Gemini to explain things they didn’t understand. They used it to work through a technical error. They had it summarise a long document before a meeting. They drafted a proposal and used AI to tighten it. These are tasks where a person is still in the loop, still making the call, still responsible for the outcome. The AI is doing legwork, not taking over.

The less-than-10% automation figure is worth sitting with. That’s the share of interactions where someone handed a task to AI and it ran to completion without human input. Even at that share, it’s mostly repetitive tasks, data entry, formatting, scheduling, where the human was never adding much judgement to begin with.

What this means if you run a small business in Australia

The “AI will replace workers” story was always more useful as a headline than as a planning assumption. What the actual usage data shows is that AI is a productivity tool, and productivity tools don’t fire people. They change what people spend their time on.

Your bookkeeper using AI to categorise transactions faster is still your bookkeeper, now freed up to catch the anomalies that need a human eye. Your customer service person using AI to draft responses is still the one who knows when a frustrated client needs a phone call instead of an email. Your marketing coordinator running AI-assisted content checks still needs to know whether the piece actually sounds like your brand.

Three examples that are already happening in small businesses across Australia:

• Administration staff are using AI to process invoices, draft correspondence, and prep meeting summaries. The work is faster. The role hasn’t gone anywhere.
• Trades and service businesses are using AI to write quotes, handle inquiry responses after hours, and summarise job notes. The person doing the quoting still has to know what the job costs.
• Professional services firms (accountants, consultants, legal) are using AI for first-draft research, document review, and client briefing prep. Judgement, relationships, and sign-off remain human.

In each case, AI is handling the parts of the job that were always a bit mechanical. The parts that require context, client knowledge, and professional judgement are still sitting with the person.

The practical takeaway

If you haven’t introduced AI tools to your team yet, the Google data gives you a clearer picture of what you’re actually introducing: a capable assistant, not a headcount reduction plan. The question worth asking isn’t “which roles can AI replace?” It’s “which parts of each role could AI take off someone’s plate so they can do the harder work better?”

Start there. The data says that’s where most people land anyway.

More hot takes: Hot Takes →

Sources

Google DeepMind, Characterising AI Model Usage Patterns from Real-World Interactions (2025). Analysis of approximately 15 million interactions with the Gemini API across research and commercial deployments.

📖 More deep dives: News Deep Dives

📖 More deep dives: News Deep Dives →