AI Coworkers Need Human Judgment: Context, Heuristics, and Delegation

Same Person, Different Context, Different Needs

TLDR: The chat era of AI is ending. New AI coworkers like Anthropic's Cowork, Manus, and Devin require delegation, not conversation. But organizations treating them as "smarter spreadsheets" are missing the critical insight: these are systems of action, not systems of record, and they need human heuristics to succeed.

Here's the uncomfortable truth about AI agents in 2026: we've been thinking about them completely wrong. The chat paradigm is dying. The tools gaining traction today - Anthropic's Cowork, Manus (acquired by Meta late last year), Devin - they don't want you to chat with them. They want you to delegate. Define the goal, assign the task, come back when it's done.

The "AI coworker" framing is actually helpful because it exposes both the potential and the pitfalls. Like any new hire, the better your onboarding, the faster they can deliver value. Like any new hire, they will only deliver outstanding work if they can complement the existing team. And here's where most organizations are getting it catastrophically wrong.

They look at AI agents as smarter spreadsheets and fail to account for the people and process side. This is fundamentally misguided because unlike previous IT systems, agentic AI systems aren't systems of record - they're systems of action. Which means these AI coworkers will inevitably run into the same issues everyone else on your team faces: information gaps between systems, unclear guidelines, lack of corporate strategy, and the general ambiguity of any modern business environment.

There's a fascinating theoretical framework here worth understanding: Mischel and Shoda's CAPS model (Cognitive-Affective Processing System). Their key insight is counterintuitive: people aren't consistent in general - they behave consistently across similar contexts. These stable if-then patterns are called "personality signatures." Once you know someone's signatures, their behavior becomes much more predictable.

In business, different situations lead to different behaviors from the same person based on: what's at stake if something goes wrong, whether mistakes are reversible, how much tacit knowledge the task requires, and whether the output represents them personally.

For teams evaluating AI coworkers, this means asking two critical questions. First: will it empower your current top performers? Trust them to have developed the right heuristics for your business environment. Rather than forcing them to change their ways of working - or using AI to level up underperformers - make sure the AI complements your best people.

Second: does it actually learn your business, or just pretend to? Most agentic AI tools today are basically amnesiac. They aren't built to learn the heuristics that make people great. This lack of business context - technically, a lack of continuous learning - makes human supervision absolutely necessary even for the most advanced AI coworkers in 2026.

Map the if-then patterns that drive business outcomes for your top performers: If it's client-facing, then I verify. If it's routine, then I let it run. If there's ambiguity, then I take over. If it's time-sensitive, then I want speed over control. If your AI tool can accommodate these patterns, you've potentially found a winner. If it can't, you'll get adoption from some people and friction from others.

Key takeaways:

• AI coworkers are systems of action, not systems of record - they face the same organizational challenges as humans
• People behave consistently across similar contexts (if-then personality signatures)
• Evaluate AI tools by whether they empower top performers, not whether they can replace underperformers
• Most current AI agents lack continuous learning - humans must fill the context gap
• Watch for flexibility tells: can users see reasoning, dial autonomy up or down, intervene at any stage?

Tradeoffs:

• Full delegation to AI gains speed but sacrifices the contextual judgment humans provide
• AI tools that force single workflows gain simplicity but lose your best people first
• Less human supervision increases efficiency but risks missing context-dependent mistakes

Link: Same person, different context, different needs

This Week in AI News

OpenAI Bringing Ads to ChatGPT: Starting in the coming weeks, OpenAI will test "Sponsored Recommendations" at the bottom of responses for US-based free and Go tier users. Paid subscriptions remain ad-free. This follows the ChatGPT Go launch and signals OpenAI's push for revenue diversification.

Healthcare AI Race Heats Up: Days after OpenAI announced ChatGPT for Health at J.P. Morgan Healthcare Conference, Anthropic responded with Claude for Healthcare. US subscribers on Pro and Max plans can now connect health records via HealthEx and Function, with Apple Health and Android Health Connect integrations coming.

NVIDIA and Eli Lilly's $1 Billion Drug Discovery Bet: A first-of-its-kind co-innovation lab announced at J.P. Morgan will co-locate Lilly scientists with NVIDIA AI engineers in San Francisco, built on NVIDIA's BioNeMo platform and upcoming Vera Rubin architecture.

Malaysia and Indonesia Block Grok: Both nations temporarily restricted access to Grok after determining it lacked safeguards against generating non-consensual sexual deepfakes. Indonesia's Communication minister called such practices "a serious violation of human rights, dignity, and citizens' security in the digital space."

Boston Dynamics + Google DeepMind Partnership: A CES announcement that flew under the radar could be huge - Gemini is coming to humanoid robots. The partnership integrates DeepMind's Gemini Robotics foundation models into Boston Dynamics' commercial Atlas humanoid, destined for Hyundai auto factories with all 2026 deployments already committed.

This article was generated from the Metacircuits newsletter. The summaries above are based on the newsletter content. Always refer to the original sources for complete information.