when ai learns to use a mouse

The tech commentariat spent two years convinced AI would quietly ingest enterprise software through immaculate, bespoke APIs.

Silicon Valley loves clean abstraction layers. The real world runs on duct tape.

Anthropic just drove a tank through that fantasy. With Claude's "computer use" capability, the model doesn't wait for engineering teams to build integration pipes. It looks at pixels on a monitor, grabs a virtual mouse, navigates a cursor, clicks buttons, and types into fields. Exactly like a caffeinated junior analyst staring down an 11 p.m. deadline.

Pixels Eat APIs for Breakfast

Enterprise software is not a sleek demo from a Series A deck. It is an unholy agglomeration of five browser tabs, an on-prem ERP built during the Howard administration, and an internal spreadsheet managed by someone who resigned in 2018.

Connecting those disparate systems via custom APIs requires months of dev time and hundreds of thousands of dollars in consulting fees. It is capital-inefficient friction.

Giving vision-enabled agents mouse controls shortcuts the entire integration tax overnight. The agent logs into an archaic billing portal, pulls an invoice, cross-references it across a legacy database, and drafts the approval email. Zero middleware required.

At Green Daisy, we spend our days building agentic products and testing where automation generates actual alpha versus novelty. What matters here is the violent collapse in time-to-value. Workflows that were economically non-viable to automate are suddenly fair game for lean organisations.

The Threat Vector Nobody Wants to Name

Do not mistake progress for perfection. Granting an algorithm the keys to your desktop is fraught with operational risk.

Models still hallucinate clicks. Front-end engineers push CSS updates that break visual selectors. Worse, giving an autonomous agent access to an authenticated browser session blows open a massive attack surface. Handing an AI your corporate banking credentials while you step out for a flat white is not innovation; it is negligence.

For bounded, repetitive tasks with strict verification loops, however, the paradigm has shifted. We have officially moved from chatbots that suggest prose to digital labour that executes keystrokes.

The software moat used to be an API ecosystem. Now, if a human eye can parse your interface, an agent can operate it.

Which legacy software provider on your balance sheet becomes irrelevant first?