Trending Articles
Trending Topics
Trending
Tired of clicking through digital forms, typing, clicking buttons and other dull digital paperwork? Anthropic’s Claude can do it for you. Well, so long as you don’t mind a bunch of errors.
AI startup Anthropic — founded by former OpenAI staffers — has unveiled an upgraded version of its Claude models, Claude 3.5 Sonnet and Claude 3.5 Haiku. Alongside performance and other improvements, Claude 3.5 Sonnet has a neat trick: it can click buttons for you.
According to a post on the Anthropic website: “developers can direct Claude to use computers the way people do — by looking at a screen, moving a cursor, clicking buttons, and typing text. Claude 3.5 Sonnet is the first frontier AI model to offer computer use in public beta.”
But this isn’t the end of shuffling digital paperwork — or the end of your job, though some companies are already making use of the tool. The company warns this is very much an early version for testing and feedback from developers: “At this stage, it is still experimental — at times cumbersome and error-prone.”
Microsoft has its own version, unveiling the ability to make autonomous agents in Copilot Studio. Soon to be available in public preview, companies are already using these AI tools to automate client onboarding, compile cases for human review, process sales orders and other administrative tasks. Honeywell said its efforts with the tool is the equivalent of adding 187 full-time employees — this is why people are scared of job losses.
If the efforts by Microsoft and Anthropic pan out, this could mark the beginning of the age of AI agents. In other words, when AI can actually do tasks rather than just answer questions, suggest text or code, or generate images. As impressive as those accomplishments may be for some industries, AI agents take it all a step further.
But it’s early days. Anthropic says the aim is to teach Claude general computer skills, enabling it to use software and other tools that are designed for people. That means developers could use Claude to automate repetitive processes, build and test software, and even conduct research.
Developers can use Anthropic’s API to access Claude, asking it to use data from a computer to fill out a form: first, telling it to check a spreadsheet, then move a cursor to open a web browser, navigate to the right web page, use the data to fill out the form, and so on.
That’s only really useful for a frequently completed task that follows a standard set of instructions, but of course there’s plenty of that in the modern world of work. The problem is Claude is even less accurate than a distracted intern: Claude 3.5 Sonnet scored just under 15% in a screenshot test, topping out at 22% with other tasks. (This is according to OSWorld, which benchmarks the effectiveness of AI compared to humans.)
Anthropic naturally expects this to improve with time — and feedback from developers — but warns anyone using it to begin with “low-risk tasks”. And though Claude isn’t very accurate as yet, Anthropic is already warning about the risk of hackers or spammers using this technology. It says it’s working on techniques to prevent such deviousness.
The Microsoft tool, meanwhile, has ten prebuilt autonomous agents that can be tweaked with filters and settings. Of course, it’s always been possible to automate basic tasks, but the idea here is that these AI-powered agents will be more powerful thanks to generative AI’s capabilities and come with built-in guardrails to avoid the worst mishaps. But again, it’s early days and this remains a preview product.
AI agents are like driverless cars: we need to decide what degree of precision is acceptable. Another way of looking at it is this: we accept mistakes from people, how many will we accept from AI? And, who is responsible for those mistakes?
Anthropic warns to use Claude in this way only for low-risk tasks, but think of anything you’d like to automate: filing taxes, booking a trip, entering sales information into a form. Wouldn’t it be delightful to point Claude at a spreadsheet, hand it your HMRC login details, and have it do your taxes? Of course it would. And humans make mistakes on taxes all the time. But who suffers the consequences if the wrong figure is imputed? Yes, it’s still you.
Imagine Claude booking a flight to a work conference, only to rock up at customs to realise you needed a visa to enter the country. Or wondering why your sales data isn’t adding up and having to unpick an AI’s maths. Such systems will save us clicks, but not anxiety if we can’t trust them perfectly — and that seems a long way off indeed. We want to ditch the burden of paperwork, but also the burden of responsibility for it.
One day, that may well be possible. In these examples, Claude or another AI agent isn’t generating data, it’s merely finding the right bits and pieces, and understanding where it goes. And making AI work in our world, rather than us having to bend to fit its ways, is a better way to go about building such systems. But in the meantime, do your own taxes — 22% accuracy isn’t going to cut it with HMRC or your boss.
