IKEA's Raj Navakoti says agents should ask

AI agents in big companies keep failing in a very specific way. Not because the models are dumb, but because they’re missing the local stuff — permissions, system access, internal jargon, weird process rules, and all the tribal knowledge nobody wrote down. Raj Navakoti from IKEA spent a May 5 workshop arguing that the fix is not better guessing. It’s getting agents to ask. Explicitly. In structure. Before they act. ### What is he actually proposing? Basically, he wants agents to stop pretending they have everything they need. Instead of silently improvising, an agent should emit a clear list of needs: I need calendar access, I need a better confidence score on who spoke in this meeting, I need permission to update this enterprise record. That turns the agent from a black box into something more like a junior coworker who knows when to raise a hand. (youtube.com) ### Why is that a big deal? Because most enterprise AI design still assumes humans should pre-package the right context for the model. That sounds neat, but it breaks in real organizations. The knowledge is scattered across wikis, chat logs, legacy tools, and people’s heads. Navakoti’s broader work calls this an enterprise knowledge problem, not a raw model capability problem. In other words — smart model, zero institutional memory. (arxiv.org) ### What does “ask” look like in practice? Not a vague chatbot reply. A structured request. That’s the important part. If the agent can say what is missing, the system around it can decide what to do next — fetch more context, ask a human, deny access, or route the task somewhere safer. The examples from the talk matter because they’re operationally real: calendar access, diarization confidence(arxiv.org)ose are exactly the places where silent guessing creates bad automation. (youtube.com) ### Why not just give the agent everything? Because that creates a different mess. If you dump every document, tool, and permission into the prompt path, you get bloated systems that are hard to govern and easy to misuse. Navakoti’s March 2026 paper makes the same point in more formal language — top-down knowledge engineering tends to produce giant, untested knowledge bases, while (youtube.com)actually needed. (arxiv.org) ### So the agent learns by failing? More like the organization learns from the failure. His “Demand-Driven Context” idea flips the process around: give the agent a real problem, watch where it gets stuck, then curate the minimum missing knowledge needed to unblock it. That is a very different philosophy from trying to model the whole company in advance. It’s closer to test-driven development, but for enterprise context. (arxiv.org) ### Why does trust keep coming up here? Because trust in enterprise AI is less about whether the model sounds fluent and more about whether people can see its limits. An agent that says “I’m only 62% confident about speaker attribution” is much easier to supervise than one that confidently writes the wrong name into a CRM. Asking before acting also creates an audit trail — what the agent knew, (arxiv.org) That’s observability, not just UX polish. This part is partly inference from Navakoti’s examples, but it follows directly from the design he’s describing. (youtube.com) ### Why is IKEA the interesting setting? Because retail and logistics are full of messy, cross-system workflows. IKEA is not a toy environment where one model answers FAQs. It’s the kind of place where physical operations, legacy platforms, and modern software all collide. Navakoti’s role sits right in that overlap, which makes his framing useful beyond IKEA — especially for any co(youtube.com)sessionize.com) ### Bottom line The interesting idea here is not “agents need context.” Everybody knows that. The sharper claim is that agents should be designed to declare their missing context, tools, permissions, and confidence as part of the workflow itself. That makes them slower in the moment, maybe — but much more usable in the places where mistakes actually matter.