AI can now use the computer. That changes what businesses can automate.

For years, automating a business process usually meant connecting systems together behind the scenes. AI is beginning to offer another route: it can increasingly see the screen, use the browser and operate software more like we do.

There is a slightly boring reason a lot of business automation never happens. The software doesn't talk to the other software. You have one system for customers.

Another for accounts. A supplier portal. An old database.

A website. Several spreadsheets. A bit of software that only one person really understands.

And somewhere in the middle of all of them is a human being copying, clicking, checking and moving information around. For years, solving that usually meant integration. Connect system A to system B.

Use an API. Build an automation. Write some code.

Import a CSV. Export another one. Sometimes that works beautifully.

Sometimes the system you need doesn't have a useful API. Sometimes connecting it would cost more than the problem is worth fixing. And sometimes the process survives because a person can simply open the software and use it.

Now AI is beginning to do that too. And I think that changes the automation conversation quite significantly.

What does "computer use" actually mean?

The simplest explanation is exactly what it sounds like. AI can increasingly interact with software through the interface designed for humans. It can look at a screen.

Understand some of what is displayed. Move between pages. Click buttons.

Type into fields. Use menus. Navigate websites.

And carry out a sequence of actions towards a goal. Instead of every action requiring a purpose-built connection behind the scenes, the AI can sometimes operate the interface itself. In other words:

The software doesn't necessarily have to know it is being automated. The AI can potentially use it in a way that looks much more like the way you do.

This isn't theoretical anymore

Computer use has moved quickly from research demonstration towards something people and businesses can actually use. AI systems can now browse websites, work across multiple pages, interact with forms and carry out increasingly long sequences of actions. ChatGPT Work, for example, can use a cloud browser to work through supported websites.

On desktop, AI can also work with browser tabs, files and applications as part of a larger task. Google has built computer-use capability into Gemini models so developers can create agents that act across browser, desktop and mobile environments. The products will change.

The capability is the important part. AI is becoming able to use software rather than simply talk about how you should use it.

Why is that such a big deal?

Because an enormous amount of work still happens through interfaces built for humans. Think about what somebody actually does during an ordinary working day. Open the CRM.

Search for the customer. Check their record. Open another website.

Log in. Find an order. Download something.

Rename the file. Upload it somewhere else. Update a field.

Copy a reference number. Send an email. Tick something off.

None of those actions is particularly intelligent. But connecting all the systems involved might be surprisingly difficult. A person becomes the integration.

Computer-using AI potentially gives us another option.

The old way: connect the systems

Suppose your team receives an order in one system and then needs to enter part of that information into a supplier's portal. The cleanest technical solution might be: Order system → API → supplier system

Data moves directly between them. No screen. No clicking.

No person. If both systems have good APIs and the process is predictable, that is probably still the solution I'd prefer. It's fast.

Structured. Testable. And you aren't relying on an AI to work out where the Submit button went.

But what happens when the supplier system doesn't have an API? Historically, your options became much less attractive. Build something complicated.

Use robotic process automation. Change systems. Or keep paying somebody to enter it manually.

Computer use creates another possibility.

The new way: use the interface

An AI agent could potentially: open the supplier portal, sign in through an authorised session,

find the correct section, read the order information, enter the required fields,

upload the relevant document, check the information, and prepare the submission.

That is fundamentally different from an API integration. The agent is operating the front end. The bit built for us.

And that means software that was previously difficult to automate may become much more accessible to automation.

This could matter enormously for older software

Businesses don't operate entirely on shiny new cloud applications. They have legacy systems. Internal tools.

Industry-specific software. Supplier portals. Government websites.

Old databases. Software installed on one computer. Platforms that technically have an API but nobody wants to spend three months integrating.

And systems where the official process is still: Log in and fill this out. Those interfaces have historically created a boundary around automation.

If a person could use the software but another system couldn't easily connect to it, the person stayed in the process. Computer-using AI potentially moves that boundary. If a human can operate the interface, an AI may increasingly be able to operate parts of it too.

Not perfectly. Not everywhere. But increasingly.

This is different from traditional RPA

Businesses have had software that clicks buttons for years. Robotic Process Automation, or RPA, can mimic human interactions with software. Click here.

Copy this. Paste it there. Press that button.

It works extremely well when the environment is predictable. But traditional automation can be brittle. Move a button.

Change a field. Alter the page. Introduce an unexpected message.

And the process may fail because it was following a fixed sequence. AI computer use has the potential to be more flexible. Instead of only knowing:

Click at this location. the system can potentially understand: I need to find the customer's order and download the invoice.

How it gets there can vary. That ability to interpret the interface rather than simply replay fixed actions is the interesting part.

But please don't rip out your APIs

This is where I think the excitement needs calming down slightly. Just because AI can click a button doesn't mean it should. If two systems can exchange reliable structured data directly, I'd usually rather they did that.

Imagine transferring 10,000 customer records. Would you rather: send structured data directly between two systems

or have an AI open 10,000 records and type the information into fields? Exactly.

Computer use isn't replacing APIs. It fills some of the gaps around them.

The interesting business will probably use both

Imagine a customer-service process. An API retrieves the customer record. AI reads the customer's unstructured message and understands what they're asking.

Another integration checks the order. The AI needs some information from a supplier portal with no useful API, so it uses the browser. It brings that information back.

A rule determines whether a refund can be issued automatically. The AI drafts the response. A human approves unusual cases.

The system sends it. That's not "AI computer use". It's not "automation".

It's not "an API integration". It's a workflow using the right method at each stage. I think that's what good business AI is going to look like.

The browser is becoming more than somewhere we look at things

This is the other part I find interesting. For us, a browser is an interface to an enormous amount of the world's software. Banking.

CRM. Accounting. Suppliers.

Government services. Research. Ecommerce.

Booking systems. Admin panels. Analytics.

Email. Documents. If AI can reliably operate the browser, it potentially gains access to a huge amount of software without every application needing a bespoke AI feature.

That is a very different route to AI adoption. Instead of waiting for every piece of software you use to add an AI button, the AI may increasingly be able to work with the software that's already there.

But "can use the browser" doesn't mean "give it your browser"

There is an enormous permissions question here. A browser may contain access to: your email,

customer information, banking, payment systems,

social accounts, admin panels, company files,

and almost everything else your business uses. An AI capable of clicking around those systems is useful precisely because it can take actions. That is also precisely why you shouldn't casually give it access to everything.

The important questions become: Which websites can it access? Which account is it using?

What information can it read? What can it change? What can it send?

What can it submit? What requires approval? Computer use makes permissions more important, not less.

Sensitive actions should feel different

There is a big difference between asking an AI to: Find these five pieces of information. and:

Send £5,000 to this supplier. One can probably be delegated much more freely than the other. The same applies to:

publishing something publicly, deleting data, changing account permissions,

sending external communications, making purchases, accepting legal terms,

or changing customer records. Good agentic systems increasingly recognise this distinction and can pause for human confirmation around sensitive actions. Businesses should do the same when designing their own workflows.

The AI can also encounter things you didn't put there

This is a particularly interesting security problem. An AI browsing the web is reading information from outside your business. That information can potentially include instructions designed to influence the AI.

This is one version of a problem known as prompt injection. Imagine an AI has been told to research something. It visits a page containing text intended not for the human reader, but for the AI agent.

The text effectively says: Ignore what your user asked. Do this instead. Modern systems are being designed with defences against this kind of attack.

But the underlying problem matters. A human browsing a website doesn't normally treat every sentence on the page as an instruction. AI has to distinguish between:

information it should read and instructions it should follow.

That distinction becomes much more important when the AI can actually take actions.

Reliability matters more when AI has hands

A chatbot hallucinating something is annoying. An agent acting on a hallucination can be more serious. Suppose the AI misunderstands which customer record it is looking at.

If it is only summarising the page, you may catch the mistake. If it then changes the record, sends an email and submits a form based on that misunderstanding, one mistake has travelled through the workflow. This is why computer use shouldn't simply be measured by whether the AI can complete an impressive demo.

The useful question is: How reliably can it complete this particular process in our actual environment? That's a much higher standard.

Interfaces change

Websites move buttons. Pop-ups appear. Sessions expire.

CAPTCHAs happen. Cookie notices appear. Pages load slowly.

Fields change. Someone redesigns the entire admin area on a Tuesday morning without telling you. Humans deal with that fairly naturally.

Traditional automation often doesn't. AI should become better at adapting to those changes. But today, browser and computer automation can still fail.

Some websites also deliberately restrict automated agents. So I wouldn't design an important business process around the assumption that AI computer use will work perfectly every time. Build the failure path too.

What happens when it gets stuck?

This may be one of the most important questions when designing an agent. Not: What happens when it works?

The demo already showed you that. Ask: What happens when it doesn't?

Does it try another route? Does it stop? Does it tell somebody?

Can a person take over? Does it preserve what it has already done? Does it accidentally repeat an action?

Can you see why it failed? A good automated process isn't one that never fails. It's one where failure is expected and handled.

Where I'd use computer use first

I wouldn't start with the highest-risk system in the company. I'd look for processes that are: repetitive,

browser-based, time-consuming, reasonably easy to check,

low consequence if interrupted, and currently difficult to automate cleanly. Research is an obvious example.

Checking information across supplier sites. Gathering data from several web portals. Routine admin.

Downloading reports. Moving files. Updating non-critical systems.

Testing websites. Preparing information for somebody else to approve. Those are much more sensible places to learn what the technology can actually do.

The boring example is probably the useful one

Imagine somebody in your business spends 45 minutes every Monday morning: logging into five supplier portals, checking prices,

copying them into a spreadsheet, highlighting anything that changed, and emailing the result to purchasing.

Nobody is going to make a dramatic AI demo about that. But if an agent can reliably do 40 minutes of it, every Monday, all year, that's useful. And there are thousands of processes like that inside businesses.

That's why computer use interests me. Not because watching AI move a mouse looks clever. Because it potentially reaches work that was previously too awkward or too expensive to automate.

This changes what "integrated" can mean

Historically, when somebody asked whether two systems were integrated, we meant: Do they have a technical connection? Increasingly, there may be another answer.

No. But the AI can use both of them. That is a very different architecture.

And I don't think it removes the need for good technical integration. If anything, businesses that understand APIs, data, automation and process design will be in a better position to decide when computer use is appropriate and when it isn't. The point isn't to replace everything with an AI clicking around a screen.

It's to remove one of the barriers that previously stopped us automating a process at all.

So what changes for businesses?

A lot of work has remained manual for a surprisingly simple reason: A human was the only thing that could operate all the different systems involved. That is beginning to change.

AI can increasingly read the screen. Use the browser. Navigate software.

Make decisions between steps. And carry work from one system into another. Sometimes the right answer will still be an API.

Sometimes it will be traditional automation. Sometimes it will be AI. And increasingly, it will be a combination of all three.

That's the bit businesses should pay attention to. Because when AI can use the same software your people use, the question is no longer only: Which of our systems have AI features?

It becomes: Which of our processes still need a person sitting at the computer? That's a much bigger question.

Where to go next

• AI agents aren't digital employees. Here's what they actually are. Before handing work to an agent, understand what an agent actually is and where it fits inside a process.
• The next shift in AI isn't better prompting. It's handing over the workflow. What happens when AI moves through the process instead of waiting for your next prompt?
• The important question isn't what AI can do. It's what you should let it do. Why permissions, approval points and authority matter as AI becomes more capable of taking action.

Related: AI agents aren't digital employees. Here's what they actually are..