AI Can Now Operate the Software Your Business Runs On

On 24 June 2026, Google DeepMind introduced something that quietly moves the goalposts for what AI can do at work: a built-in computer use tool in its Gemini 3.5 Flash model. In plain terms, it lets an AI agent see a screen and operate software the way a person does, clicking buttons, typing into fields, scrolling and moving across a web browser, a desktop application, even a phone. For two years, AI has mostly handed you words. This is AI that goes and does the work inside the tools your business already runs on.

Google's announcement put it plainly: computer use is now a built-in tool supported in Gemini 3.5 Flash, delivering, in its words, the best performance yet for agentic computer use tasks. The detail that matters for a small business is not the benchmark. It is that the model doing this sits on the fast, low-cost Flash tier, the version built for high-volume everyday jobs rather than the expensive flagship. The kind of work that used to be too fiddly or too cheap to bother automating is exactly the work this is aimed at.

It is worth being precise about what this is and is not. Right now it is a capability that people build agents on, not a single button inside an app you already use. But the direction is unmistakable, and it is the direction every major lab is running in. The team behind it is led by Demis Hassabis at Google DeepMind, the same lab that has spent this month publishing openly about how to make AI agents safe to deploy. When the people building frontier models start shipping agents that operate ordinary software, the question for a business stops being whether this arrives and becomes how soon you put it to work, and how safely.

From AI that talks to AI that does

Here is the shift in one line. A chatbot drafts the email, summarises the document or answers the question, then hands it back for you to actually go and do something with. An agent that can use a computer closes that last gap. Instead of telling you what to do, it goes and does it across the screens where the work actually lives: the booking system, the supplier portal, the spreadsheet, the admin panel that never had a neat integration. It does not need a special connection built for it, because it works the software the same way a staff member would, by looking at it and acting.

Why this lands hard for a small business

Every small business runs on a quiet tax of screen work. Someone copies an order out of an email into the accounting system. Someone updates the same price across three listings. Someone reconciles a booking, chases a missing detail, re-keys an invoice, logs into a portal that has no app and no export. None of it grows the business. All of it eats the hours of the people who could be serving customers instead. This is precisely the work an agent that can operate software is built to absorb, and for a small team that could never justify a custom software project, that is a genuinely new option on the table.

The catch worth knowing about

There is a reason this is powerful and risky in the same breath. An agent with its hands on your real software can do real things, which means a mistake is a real mistake and a tricked agent is a real problem. Google has built in safeguards: it trained the model against adversarial attacks, it offers a setting that makes the agent ask for explicit confirmation before sensitive actions, and it can automatically stop a task if it detects a prompt injection, the trick where hidden instructions on a web page try to hijack the agent. DeepMind itself has been candid this month that running agents at scale is not yet a fully solved, safe thing. That honesty is the point, not a reason to wait. This is a capability that pays off handsomely when it is set up by someone who knows where the landmines are, and goes wrong quietly when it is not.

So the opportunity and the warning are the same sentence. The businesses that win with this will not be the ones who move fastest and least carefully. They will be the ones who get a capable agent onto the dull, repetitive work first, with the right limits around it, while everyone else is still copying and pasting by hand. Here is what good looks like once it is set up properly:

• The repetitive screen work that fills your team's day, copying, re-keying, updating the same thing in three places, gets done in the background instead of by a person.
• Software that never had an integration, the portal with no export, the legacy system, the supplier site, can finally be part of an automated flow, because the agent uses it like a person would.
• The work runs with the right guardrails: confirmation before anything sensitive, clear limits on what the agent can touch, and a stop the moment something looks wrong.
• Your people spend their hours on the things only people can do, the customer in front of them, the quote that wins the job, the call that saves the account.
• You get the upside of automation without a six-figure software build, because the agent adapts to the tools you already have rather than the other way around.

The advantage no longer goes to the business with the biggest software budget. It goes to the one that puts a capable, well-supervised agent on the dull work first.NextAura

None of this means you should hand an AI the keys to your business this afternoon. The capability is new, the safe way to use it is genuinely fiddly, and the gap between a slick demo and a dependable system that runs every day is where most of the real work lives. But the prize is real, and it is arriving faster than most owners expect: the small team that gets an agent doing its screen work first will quietly out-run the one still doing all of it by hand.

This is exactly what we do at NextAura. We build and run the AI agents that take the repetitive screen work off your team, set up with the guardrails that keep them safe, and we steer the whole thing so it genuinely helps rather than just impresses. If you would like the upside of this without learning where the landmines are, much like the practical agents we wrote about in AI agents that actually help, get in touch and we will take it from here while you get back to your customers.