Anthropic’s Computer Use Capability and What It Tells Us About Where AI Is Going
Our discussions with healthcare leaders at HLTH last week ranged from tactical to strategic, though a few themes stood out:
The challenges in evaluating accuracy, precision, and bias of AI software across use cases
The dichotomy between vendors trying to find their niches within a market and those trying to expand from point solution to platform
The question vendors now seem to be asking: Will advancing AI capabilities help or harm our business?
Timed perfectly to hint at some of the resolutions to these themes, Anthropic released its computer use capability, which allows a model agent to interact with the computer in the same way as humans.
Essentially, this feature allows the user to prompt the agent to perform a task on the computer, and it will navigate through tabs and programs, take screenshots of pages, interpret those screenshots, and type and click around to accomplish the task. While all of these individual capabilities exist in robotic process automation (RPA), the novelty of this feature is its flexibility: it doesn’t require a pre-defined process or coding to accomplish its task—it is capable of reasoning and common sense in deciding how to proceed. This is a reflection of a large part of many tasks that many humans today have to do manually.
It’s a limited demo, but the impact of an agent that can reason and interact across computer systems is vast, especially in healthcare, which is burdened by a range of administrative duties. An agent capable of reasoning and interacting flexibly across systems like EHRs, practice management systems, and claims management systems will be able to automate tasks from patient registration to billing and compliance reporting. The need might be most acute on the billing side, where there are already a number of revenue cycle automation vendors, and where administrative staff struggle to keep up with tasks like prior authorization and eligibility checks, denials management, and payment posting and reconciliation.
While there are a number of regulatory and compliance hurdles to overcome, what makes this piece of technology relevant across the spectrum is that it’s being released as an API for potential integration directly by either vendors or providers. Suddenly, it’s possible to implement without having to replace entire systems of record, but instead interact across multiple systems and bridge the gap between them.
Coming back to the themes we saw at HLTH, tools like computer use enable implementing specific workflows as point solutions, but with the shared architecture of a common platform. With this, evaluating AI software will start to look more like evaluating an employee’s performance. As for how advancing AI will affect a vendor’s business, it’s starting to look more and more like the use of frontier models will be key in many of the most advanced workflows.