Workers should control the means of agentic production, suggests WorkBeaver boss

INTERVIEW "I think everybody is adopting AI irresponsibly and I think it's going to have a net negative outcome on the socio-economic standing of the world," said Bars Juhasz. "So let's see if we can't pitch more of a win-win future."

Juhasz, CEO of no-code automation biz WorkBeaver, told The Register in an interview that he hopes to encourage the adoption of AI-based automation – agents – from the demand side rather than the supply side, from the perspective of the office worker rather than management.

Around August 2024, he said, he sold most of his shares in Undetectable.ai, a company he co-founded, and found himself in a position to work on whatever he wanted.

Given his background in machine learning, he said, he started thinking about AI agents – AI models given access to tools – and how they're likely to affect society. The result was WorkBeaver.

"We think companies are going to rush in, they're going to adopt AI too quickly, lay off people as a result of that, realize they messed up, and have to bring people back in," said Juhasz. "And wouldn't you know it? Here we are a year later, and it's played out almost to the T of how we hoped they wouldn't go, but here we are."

Juhasz expects that within five to ten years workers will have to demonstrate some level of AI proficiency appropriate to their roles, a view evident in Purdue University's new AI competency requirement as a condition of graduation. But, he said, "a lot of people are just not going to be able to do that. They're just not in that position." Maybe they're non-technical, he said, or they're just resisting the technology.

Taking it slowly may be anathema in an industry that celebrates the Zuckerbergian zeal to "move fast and break things." However, the consequence of urgent AI adoption imperatives is that someone has to deal with the mess.

"Now we've got these stories coming out of having to backtrack, of AI janitors, right?" Juhasz mused. "A whole new career dedicated just to cleaning up messes left by this rush."

WorkBeaver, said Juhasz, represents an attempt to help people become comfortable with AI, regardless of their level of technical competency.

"So instead of focusing on the business side of the equation, the supply side, what if we focused on the demand side?" he said. "What if we focused on the people who are set to be displaced, who otherwise realistically have very little fighting chance to be in the competitive workforce in the next five to ten years?"

WorkBeaver presently functions by using a menu-driven interface that asks users for prompt-based descriptions of tasks. The underlying AI agent then attempts to carry out the described task as if it were the human user, without requiring code or APIs.

The company website contains a variety of use cases designed to help people understand the sorts of tasks its tech can handle, because people don't necessarily see how AI technology can help them accomplish things. Examples include automated form filling, reminder setting, appointment setting, email sending, data gathering, and data entry.

"Some of the people using WorkBeaver to create agentic automations don't even use ChatGPT," said Juhasz. "That's the level of non-technical we're targeting."

Juhasz said WorkBeaver's approach differs from other companies selling agentic tools through its focus on staff and what workers find helpful, rather than top-down directives to use AI without much understanding of the work being done.

"For example, I spoke with reporters at Reuters and they have top-down mandates from managers who are not editors, who are not journalists, coming in saying: 'Okay, you guys have to use these tools X amount of times per week.' To me, they're KPIs [key performance indicators], right? And the tools they're asking them to use are literally replacing the critical thinking process of the job. It's like put in the story, put in the sources, and it'll spin differently. Obviously, as you can imagine, they're not too happy about it. They don't believe that's where automation should be taking place."

WorkBeaver launched its private alpha in January this year and reached open beta in September. The company now has close to 4,000 users, according to Juhasz, primarily among small-and medium-sized enterprises. Customers are a mix of individuals paying out of their own pockets and of teams buying multiple licenses. Juhasz is also working on a deal with a laptop maker to have WorkBeaver preinstalled on business laptops in South America.

Juhasz said the company is a commercial success but has fallen short as a social enterprise that helps people become comfortable with AI.

One reason, he explained, is that deploying agents requires user education.

"Education is required for prompting," he said. "We're about three to six months ahead of even OpenAI and Anthropic in this space. They're about to find out what we found out three months ago, which is the education requirement to make an agent do what you wanted it to do effectively by prompting is going to immediately cut out a lot of normal people."

The other reason, he said, is that "People just didn't know what to automate."

WorkBeaver’s next release will address that issue. Due this month, the next iteration of the service replaces the menu-driven prompting process with an agent monitoring screen, mouse, and keyboard interactions as a way to learn the tasks users undertake and then repeat them.

"What we've done is we've created a very unique user experience where anybody, non-technical user, is able to create automations, agentic automations, with no prompting, no drag and drop," explained Juhasz. "The way it works is you sit down with the agent and you collaboratively work through a task together. So basically anything you do on your computer, the agent should be able to take over."

The collaborative training process is the key – the agent can catch and correct potential points of failure as the worker is walking through the task. This avoids the frustrating experience of creating an agent-based automation that fails at runtime and requires a technical person to revisit the code or the configuration process.

Juhasz said WorkBeaver built its tech using Google Cloud Vertex AI. But the agent is model-agnostic and can, with a bit of training, work with any of the foundation models and can run in enterprise private clouds.

"We are very much on the privacy end of the spectrum," said Juhasz. "Literally, the only things we store on our servers about you is your email address and your account balance. That's it. I don't need liability, so everything is stored locally on the users' devices, fully encrypted. We learned from Microsoft's mistakes."

Juhasz added that WorkBeaver has a zero data retention agreement with Google Cloud. "Your data lives and dies in ephemeral memory for a fraction of a second at most," he said.

Allowing an AI agent to control your computer is a security risk, one that Juhasz readily acknowledges.

"I don't want an agent being able to access stuff it shouldn't have access to," he said.

WorkBeaver addresses this concern through application and folder allow-listing, alongside a few additional security measures.

The litmus test, said Juhasz, will be whether, when the new version is released in a week or two, non-technical people use the technology in a way that makes them happier at work. ®