The Behaviors AI Cannot Do For You

Edmundo Barrientos · 2026-09-16 · 3 min read

TL;DR: Deloitte names three human behaviors that decide whether AI helps or hurts an organization: judgment, experimentation, and divergent thinking. You can't cultivate any of them through corporate mandates. To acquire them, each needs its own practice.

The call nobody double-checks

A claims adjuster works through a stack of files. The AI advises approval on a borderline case. She has forty more claims to process and another policy page to check. Most days the tool is right, so she signs off.

Deloitte research on why sensing behavior beats sensing usage found that only half of executives regularly verify AI output when making decisions. That gap is judgment: knowing when to accept, question, or reject AI guidance. At Roche, "a key part of judgment is knowing when to use AI and when not to. We don't want to rely on it for strategic thinking."

Judgment does not grow from a reminder to double-check. It grows through hands-on experience: catch a flawed call in a case you know, and explain the error. Build that into the adjuster's own workflow, scored against her own claim types. Skip it, and the next unusual claim gets approved by mistake.

The unrecognized effort of experimentation

Deloitte also found that 78% of employees are testing new techniques, but only 56% feel acknowledged for it. Canva gives 5,000 employees a week to explore AI tools. That shows engagement, not which practices held up.

Experimentation, run as practice, needs a shape: try something new, check it against a case with a known answer, then report what happened. This is the practice half of augmentation; most programs never get past trying. Our adjuster tries the AI's photo summary on three claims, checks it against her own read, and then trusts it on a new one. That report is what spreads through a team, not the memory of a fun week.

The reframe AI cannot generate

AI is built to match your case to a pattern it has seen, not to notice that this case matches none. That is divergent thinking, and Deloitte is blunt: AI is "not, by design, a source of" it. The room for that thinking is also shrinking. Among workers facing significant change from AI, Deloitte found 69% report a heavier workload and 43% name lack of time as the biggest barrier.

So this practice has to be scheduled: reason through a standing case before the AI's recommendation is opened. That way, the habit of noticing a new pattern does not fade from disuse. That habit is what catches the fraud scheme that looks like nothing the model was trained on.

One-size-fits-all training inevitably falls short

Judgment, experimentation, and divergent thinking fail differently, so one policy and one course cannot build all three. Augmentation wins only when people build the specific new judgment calls it demands. That means the practice has to adapt to what each person actually needs, not repeat the same scenario for every role. Skip it, and AI teaches people to trust whatever the screen says.

References

  • Deloitte. AI adoption to AI adaptation, July 2026.

About the publisher: this article is published by Grasply (grasply.ai), a capability builder. Grasply turns what a company already knows, its docs, calls and recordings, into skills people use at work, for employees, partners and customers. Practice turns knowing into doing, and Grasply shows proof that the work changed.