High-agency cognition: why few professionals leverage AI well

Two senior professionals at the same company, same area, same access to Claude Enterprise. In 6 months, one became 3× more productive; the other still delivers the same. Why? Not AI — what cognitive literature calls high-agency cognition.

This post connects research in metacognition, epistemic agency, and Karpathy’s “high-agency” framing to the real difference between professionals leveraging AI well or poorly in 2026.

The scientific question

“High agency” isn’t traditional academic vocabulary, but it maps to established constructs: locus of control (Rotter), self-efficacy (Bandura), metacognitive control (Flavell, recent Nelson & Narens), epistemic agency (Pickering, Knorr Cetina).

A high-agency person, in the literature, shows three empirically robust traits:

1. Attributes outcomes to their own action (internal locus of control).
2. Believes in their capacity to change outcomes (high self-efficacy).
3. Monitors and adjusts cognitive process while operating (high metacognitive control).

These traits are partially dispositional (stable inter-individual variation) and partially situational (change with context, training, modeling).

Why generative AI amplifies differences

AI is, for whoever uses it, a cognition amplifier. A high-agency person:

• Reformulates the task before prompts (“what am I really trying to solve?”).
• Iterates when first response is bad (“what’s missing in the prompt to make the model do X?”).
• Combines multiple tools (“Claude for text + Gemini for image + spreadsheet for calculation”).
• Calibrates confidence (“this response deserves independent verification, this one doesn’t”).

A low-agency person:

• Treats AI as oracle (“AI said X, so X”).
• Accepts first response (“didn’t work well, AI is bad”).
• Uses one tool for everything (“only ChatGPT”).
• Doesn’t calibrate (“believes everything equally”).

Result: the gap between the two grows, not shrinks, with time. AI is technology where Matthew 13:12 literally applies.

Empirical research available in 2026

Early research (2024-2025) already shows strong correlation between:

• Self-efficacy → Copilot productivity in Microsoft study (Microsoft Research, 2024).
• Internal locus of control → AI adoption in medical studies (Mayo Clinic studies, 2025).
• Metacognitive accuracy → prompt quality in cognitive studies (various Anthropic + academic publications).

Large-scale replication and longitudinal measurement still missing. But the signal is consistent.

What this means for adoption programs

The “classroom training for 200 people” model assumes AI is a technical skill. It’s not. It’s a meta-skill amplifier — those with high agency double; those without, hold steady.

Practical implications:

1 · Select champions by high-agency signal. Not by role, not by technical area. People with a track record of “solving problems without a manual” pull adoption 10× better than people waiting for formal process.

2 · Training delivers tooling, not agency. Expecting “everyone will become power user after 1 day of training” is fantasy. 20-30% will, 70% will continue as they were.

3 · Individual coaching beats mass training. 1-1 of 60 minutes with the professional active on their own task generates more agency than 8h class of 30. Similar total cost, different ROI.

4 · Organizational culture pulls or blocks. Company that punishes errors punishes iteration. Iteration is half of high agency. Cultural defect knocks down any technical program.

Link to the AI Agency Ladder

The AI Agency Ladder describes organizational levels. What this note adds: the individual base.

L1→L2 is unlocked by technical training + tool access. L2→L3 begins depending on individual high-agency — those without get stuck in perpetual L2.

L3+ requires high-agency in a critical mass of the team. That’s why L4 and L5 are rare — platform alone isn’t enough; needs people operating with agency.

The stewardship question

Before buying corporate AI platform for 200 seats, ask: does my team have 30+ people with high agency demonstrated in other dimensions? If not, train high-agency first (in any skill — own projects, decision autonomy, continuous learning). Then buy platform.

AI program without agency investment is fancy shelfware.

Where to go deeper

Karpathy’s “Highest-Leverage Tasks” notes; Bandura, “Self-Efficacy: The Exercise of Control”; Flavell, “Metacognition and Cognitive Monitoring”; Anthropic blog series on prompting (2024-2026). And the AI Agency Skills cluster accumulates applied material.