AI Agents Can Now Use the Software You Already Run

On 24 June, Google quietly shipped one of the more consequential changes of the year for small business automation. It built computer use into Gemini 3.5 Flash, its fast, low-cost model. In plain terms, an AI agent can now see a screen and operate it the way a person does: read what is on the page, click a button, fill a field, move between tabs and work through a multi-step task across a browser, a phone or a desktop.

That sounds like a developer detail. It is not. It changes which jobs in a business can be handed to software at all, and the answer just went from a short list to almost anything you do with a mouse and a keyboard.

Why this is bigger than another model release

For years, automating a task meant the tools involved had to offer a clean way in: an API, an integration, an export button, some agreed plumbing between systems. The trouble is that most of the software a small business actually runs on offers none of that. The wholesaler's ordering portal, the old booking system, the council or government website you log into every quarter, the niche industry app your trade depends on: these were never built to talk to anything else. So the work stayed manual, and an owner or a staff member kept clicking through it by hand.

Computer use removes that requirement. An agent that can look at a screen and use it does not need a tidy integration, because it works the same surface a human does. The practical effect is that the no-API systems, the ones holding the most tedious, repetitive work in a business, are suddenly in reach of automation for the first time.

This is the direction the frontier labs have been pushing all year. Demis Hassabis, who leads Google DeepMind, has been consistent that agents which can take action, not just answer questions, are the next phase of what this technology is for. The 24 June release is that idea arriving in a model cheap and fast enough for everyday use.

What this unlocks for a small business

Forget the demos of an agent booking a holiday. The real prize for a small business is duller and far more valuable: the recurring screen-work that quietly eats hours every week, now done by something that does not get bored, distracted or sick of it. Here is what good looks like once this is set up properly:

• The supplier portal with no download is checked for this week's prices and stock, and the numbers land where you need them, without anyone keying them in.
• The booking or roster system that will never speak to your accounts gets reconciled automatically, instead of someone copying figures across by hand each Friday.
• The same form you fill in every quarter, on a website that has not changed since 2014, is completed and submitted while you do something that actually grows the business.
• Two systems that were never going to integrate are kept in step, because the agent moves the data between them the way a person would.
• The repetitive afternoon admin is done overnight, so the team starts the day on the work only people can do.

None of that is glamorous. All of it is the kind of friction owners usually just absorb, the tax of running on software that was never designed to cooperate. Lifting it off the team is where the hours, and the morale, come back.

The breakthrough is not an agent that can do anything. It is an agent that can finally touch the unglamorous, no-API systems where a small business loses its hours.NextAura

The capability is the easy part

Here is the honest caveat, and it is the same one Google flags. An agent that can use any screen can also be steered by whatever is on that screen, which is why the same week brought a fresh round of warnings about agents being targeted. We wrote about that risk separately: an agent will do what it is told, even by a stranger, if a hidden instruction is sitting on a page it was asked to read. Google ships guardrails for this, confirmation on sensitive actions and the ability to stop a task if something looks like an attack, but it is explicit that these are layers to combine, not a finished answer.

So the capability is now genuinely available, and genuinely cheap. The hard part, the part that decides whether this saves you hours or quietly causes a mess, is the judgement around it: choosing the jobs worth automating, building the agent to do them reliably on your actual systems, and keeping a human hand on anything that moves money or cannot be undone. That is craft, not a download.

If you have been waiting for a reason to take automation seriously, this is it: the wall that kept the messiest, most time-consuming jobs off-limits has come down. The opportunity is in the boring systems, and the businesses that point an agent at them first will quietly claw back a day a week while everyone else keeps clicking. This is exactly the work we do at NextAura. We figure out which screens in your business are worth handing to an agent, build it to run reliably and safely on your real tools, and keep a hand on the wheel so you get the hours back without the exposure. Get in touch, and we will take it from there while you get back to the business.