Why Your AI Employee Can't Do Their Job

There’s a lot of hype about AI employees right now. AI SDRs, AI engineers, AI marketers – the promise is that these digital workers will replicate what humans do, but faster, cheaper and better.

Most of them won’t work. Not because the AI isn’t smart enough (these models can solve math olympiad problems), but because we’ve handicapped them in the most fundamental way: they can’t have conversations.

The Silent Employee Problem

Imagine onboarding a new SDR with one rule: they can never ask questions. They can only read documentation and do their job. No clarifying questions. No checking in when things are ambiguous. Just silent execution based on what’s written down.

How long would they last? A day? Maybe a week if you have exceptional documentation?

Yet this is exactly how most “AI employees” work today. They can access your tools, read your data, generate emails, even make decisions – but they can’t tap someone on the shoulder and ask “Hey, is this what you meant?” or “Should I prioritize X or Y?”

The Real Work isn’t in the Script

The dirty secret about white-collar work is that the happy path – the standard, predictable stuff – is maybe 20% of the job. The other 80% is edge cases, ambiguity, and situations that require judgment.

Take an SDR. Sure, they send follow-up emails. An AI can write those emails beautifully. But that’s not the job. The job is knowing when to deviate from the script, how to handle that weird objection from the enterprise prospect, when to loop in the AE early, or how to navigate internal politics to get a deal unstuck.

These aren’t things you can document exhaustively. They require conversation. Questions. Back-and-forth with other humans who have context, experience, and judgment.

The Hierarchy of Skills

When you interview someone for a role, what’s the first thing you evaluate? For most knowledge work, it’s communication skills. Can they explain their thinking? Can they ask good questions? Can they collaborate?

Technical skills matter, but they’re table stakes. For a lot of roles, what separates good employees from great ones is their ability to navigate ambiguity through conversation.

Yet when we build AI agents, we focus on the technical execution – can it write code, can it analyze data, can it generate content – while completely ignoring the communication layer. It’s like hiring a brilliant marketer who refuses to attend meetings or respond to Slack. How useful would they actually be?

The Chief of Staff Test

Here’s another thought experiment: You ask your Chief of Staff to get a status update on your company’s biggest initiative. Are you on track? If not, why? What needs to change?

Now, imagine they can’t talk to anyone. They can only look at Jira tickets, read Slack messages, and review documents. No questions. No conversations. Just passive observation.

What kind of update would you get? Probably something surface-level about ticket counts and due dates. Nothing about the real blockers, the unspoken concerns, or the context that lives in people’s heads.

This is exactly what most companies are building with their “AI agents.” They give them read access to everything but no ability to engage. Then they wonder why the insights are mediocre.

Communication as Infrastructure

At Windmill, we’ve been wrestling with this problem. How do you give agents a voice? Not just the ability to respond when asked, but to proactively reach out, ask the right questions, and update their understanding based on the answers?

It’s not a model problem – ChatGPT, Claude, Grok and even small on device models are already excellent conversationalists. It’s an infrastructure problem:

• Who should the agent talk to? (You can’t spam everyone)
• When should they reach out? (Timing matters)
• What questions should they ask? (Relevance is key)
• How do they incorporate responses? (Learning from feedback)

This is an extremely difficult challenge. In order to do a great job you need to understand the org chart, understand how people work together, what are the existing communication patterns and even human psychology.

We haven’t fully figured this out yet, but we’ve made progress in two ways:

First, we let managers configure when Windy reaches out to people through the concept of routines. It’s not fully autonomous yet, but it’s a step toward agents that know when to engage.

Second, we’ve built conversation flows that can ask good follow-up questions. If someone mentions a blocker, Windy knows to dig deeper. If there’s ambiguity, it can clarify.

The Future: Agents That Think Out Loud

Here’s where this needs to go:

An agent working on a project update would first scan all available data – tickets, code, documents. Then it would identify gaps and uncertainties. It would reach out to specific people with targeted questions—maybe even suggest a quick sync if the situation is complex.

It wouldn’t just execute silently in the background, but it also wouldn’t make people do extra work. It would function the way good employees do – gathering context, asking questions, building understanding, then executing with confidence.

Why This Matters

The companies that figure this out won’t just have better AI tools. They’ll have actual digital employees – agents that can navigate ambiguity, handle edge cases, and improve over time through interaction.

The technical capabilities are already here. Models can write, code, analyze, and reason at superhuman levels. But without the ability to communicate – to ask questions, clarify, and learn from colleagues – they’re just very sophisticated automation.

Real intelligence, artificial or otherwise, isn’t just about processing information. It’s about knowing what you don’t know and having the ability to find out.

That’s the gap between AI tools and AI employees. And it’s why most of today’s “digital workers” are really just dressed-up automation. They can handle the happy path, but they fall apart the moment things get complicated.

The solution isn’t better models. It’s giving them a voice.