I’ve been vaguely keeping track of the progress of “digital employees”—AI-powered agentic systems that simulate traditional employees. There are plenty of startups trying to build them.
- Lindy: “Meet your first AI employee”
- ai.work: “Autonomous AI Workers designed for internal operations teams - IT, HR, Procurement, Legal and beyond.”
- Relay.app: “Augment your team with AI teammates that work for you”
- Teneo.ai: “Join Teneo and our +17,000 AI Agents revolutionizing customer service to fully automate Tier 1 support.”
- 11x: “Digital workers, Human results.”
- newo.ai: “Create your own AI Receptionist in just 3 minutes and start earning up to $30,000 more per month per location by ensuring you never miss a customer, even after hours.”
- Devin: “Crush your backlog with your personal AI engineering team.”
- Beam: “Hire Self-Evolving AI Agents to Run Your Operations”
- Relevance AI: “Build teams of AI agents that deliver human-quality work”
I’m currently listening to the second season of the Shell Game podcast. In the new season Evan Ratliff is doing his best to build and run a startup almost entirely using these digital employees. So far he’s had mixed results, and he and his other human teammate have had to put in a ton of effort adding scaffolding on top of the AI agent tools to get them to behave somewhat productively.
What really stood out the most was the degree to which his digital employees appeared to be role playing as employees, making up backstories for themselves, fabricating details about the company, and imagining how they spent their weekends. Ratliff finds this charming. I find it off-putting.
I do find the idea plausible that AI hallucinations aren’t entirely bad. They may be the key to the creativity and imagination AI brings to the table. But a digital employee who fails to draw the line between imagination and reality seems less than useful.
I don’t know to what degree this conflicts with effective prompt engineering. Is it necessary to tell the AI that it’s a human HR professional with 12 years of experience in the industry in order to coax the best performance out of the model? I hope not.
I expect these tools will become more and more pervasive as they improve, though I’m less certain whether the hyper personification of some of these tools will persist. Will we prefer our AI agents have human-appearing names and avatars, or would we rather they present as what they are, AI-powered software bots?