EDITOR'S NOTE
Last week I wrote about what happens when organizations measure AI's impact without first defining what success looks like. This week I want to go one step upstream: why do so many deployments end up causing harm before anyone starts counting?
The frame I keep coming back to is borrowed from medical ethics: non-maleficence, do no harm. Most AI deployment conversations focus on productivity and job displacement, but they miss the quieter question of whether employees still feel they're advancing their craft, feel valued, and feel that their work is meaningfully theirs. Research in the Journal of Business Ethics describes what happens when organizations skip this question: workers default into monitoring and correcting AI output rather than practicing their craft, a dynamic the researchers call "tending the machine." Productivity metrics can look fine the whole time. The harm is in what employees are no longer doing. The organizations pulling ahead aren't the ones that automated the most. They're the ones that understood their people's individual strengths and designed AI to amplify those, not absorb them.
THIS WEEK'S READING
NN/G's 2026 state of the industry report makes a point easy to miss in the AI adoption rush: everyone using the same tools doesn't mean everyone gets the same results. The competitive edge isn't in how much you delegate to AI, it's in knowing which parts of your work to protect.
Why It Matters: Leaning on AI to shore up your weaknesses while doubling down on your craft is a strategy. Blindly following what everyone else is doing isn't. The interface is shifting from screen to agent, but the judgment about what to hand off and what to hold still belongs to the person doing the work.
That framing raises an immediate question: when you do delegate, what actually determines whether trust in AI forms correctly?
An N=900 experiment and N=57 eye-tracking lab study from this year's CHI surfaces a counterintuitive finding: making an AI more human-like doesn't automatically build trust. In some cases it actively undermines it.
Why It Matters: Cognitive engagement has to come first. If a user understands what the AI is doing and why, social cues then reinforce that trust — skip the explanation and lead with the personality, and you get the uncanny valley problem instead. Explanation also needs to be tailored to the individual context, not delivered as a one-size-fits-all disclaimer. The mechanism matters more than the feature.
Knowing how trust forms is the starting point. The harder question is how to build it incrementally as AI capability proves itself over time.
AI isn't the right tool for every job, and knowing when to apply it is as important as knowing how. This piece introduces "progressive delegation" — start narrow, expand autonomy as the agent earns it through demonstrated reliability.
Why It Matters: The goal isn't maximum automation, it's appropriate reliance. AI lacks the contextual awareness humans bring to any interaction: the unspoken cues, the situational judgment, the things that don't make it into a prompt. That gap means you have to evaluate how the model actually behaves in your specific context and let that guide how much you hand off.
At the organizational level, that principle scales into something larger: rethinking not just individual workflows, but how the whole operation is structured.
The question most organizations are asking — what can an agent do? — is the wrong starting point. McKinsey's framing flips it: start with the workflow, not the tool.
Why It Matters: Agentic AI is now spanning systems that used to operate in silos, and the real opportunity is in rethinking how work is organized across the whole operation. 71% of consumers expect personalized interactions and 76% get frustrated when they don't happen — AI is the first technology that makes individual-level tailoring possible at scale. The organizations capturing that aren't deploying the most agents. They're thinking biggest about where AI actually moves the needle.
The strategic picture is clear. The practical question is where to start.
Most people treat delegation to AI as a binary: either you're doing the task or the AI is. This framework breaks that into six distinct levels, from "look into it and report back" to "take action, no further contact required."
Why It Matters: Skipping levels is how things go wrong. The Amazon tokenmaxxing story from our June 2 issue is a good example — the failure wasn't in the measurement, it was in jumping to a high level of delegation before the trust infrastructure existed to support it. What makes this piece stand out is that you can pick it up and apply it to your own workflows today, no setup required. Start at Level 1 with something low-stakes and build from there.
The pattern across this week's reading is consistent: the organizations getting AI right aren't starting from the tools, they're starting from the people. What does this person need to feel like they're doing their best work? What does this workflow need to produce genuine organizational value? What level of trust has actually been earned? Answering those questions before you deploy is the design decision most organizations skip, and the one that determines everything that follows.
If you're building out an AI deployment right now, forward this to whoever is designing the handoff moment. That's where it's won or lost.
Have a take on any of this? Reply directly, I read every response.
Until next Tuesday,
Charles Costa, MLIS