Zillow Offers Didn’t Fail Because Its AI Was Bad. It Failed Because the AI Was the Only One in the Room.
Matt Sheehan | SpatialNext
Zillow built one of the most sophisticated sensing layers in real estate. Then they made the Zestimate the offer price and removed human judgment from the purchase decision entirely. Nine months later: $500 million in losses, a quarter of the workforce gone. The AI didn’t fail. The decision architecture did.
From Human-in-the-Loop to Human-with-Agency: Why AI Oversight Fails When Humans Are Present but Powerless
Institute for Systems Integrity
The real question for any organisation deploying AI isn’t whether there is a human in the loop. It’s whether that human can actually intervene when the system is wrong. This paper introduces the Human Agency Framework — the distinction between symbolic human presence and protected judgement under pressure. Required reading for anyone who has pointed to a job title as evidence of oversight.
AI Governance in 2026: Is Your Organisation Ready?
David Talby | Dataversity
Governance that lives only in policy decks doesn’t govern anything. As agentic AI moves from analysis into execution, liability increasingly centres on actions, not outputs. This piece argues that oversight has to move closer to runtime, built into the system before it acts. The guardrails versus spine distinction in different language.
Automation Bias Persists with New AI Tools
International AI Safety Report 2026
The authoritative source on what happens when humans trust the model more than their own judgement. Automation bias isn’t irrational — it’s a predictable response to systems that present confident outputs with no visible uncertainty. In a randomised experiment with 2,784 participants, people were less likely to correct erroneous AI suggestions when doing so required extra effort. That finding belongs in every conversation about human oversight design.
The pattern across these four pieces is the same. The sensing layer is built. The reasoning layer is arriving. The human decision layer — the one that determines whether any of it produces a return — has no clear owner.
If that’s true in your organisation, that’s usually where my conversations start: mattsheehan@spatialnext.io