Training fails on the wrong axis. Builders are short of a way to feel where capability stops, not of tool literacy: the frontier is jagged and invisible from inside, so the curriculum has to teach where the tool quietly breaks rather than how to drive it.
Why this matters
Leaders name insufficient worker skills as the biggest barrier to integrating AI into existing workflows (n=3,235; a leader-perception ranking, and the respondents are the people who set training budgets). The response runs to education and past the structural change education depends on. In the same survey, 53% educate the broader workforce for AI fluency and 48% design upskilling programs, while role redesign, career-path redesign and reorganizing around new work patterns all sit at 33% and below.
Resistance ranks low as a constraint, and two independent sources say so. Worker sentiment as characterized by leaders runs 13% highly enthusiastic, 55% at least open, 21% “prefer not to use AI but will if required,” and 4% actively distrustful, so roughly two-thirds fall on the willing side. In a separate survey, employee resistance ranks near the bottom of named challenges (n=110, India-only, respondents are the change leaders being graded; weak evidence, cited as one dissenting datapoint, not a finding). And in the US tracking study, employee resistance and lack of trust fell to #8 from #2 two years earlier, so people are willing and the gap is ability and time.
Time is the part enablement programs quietly assume away. The mechanism, from a named startup CEO rather than a study: employees are expected to find experimentation time inside unchanged workloads. Nothing about a curriculum survives that. If the training is real, the hours are on a calendar and the work those hours displaced is named.
Meanwhile investment expectations are moving the wrong way. Nearly half report technical skill gaps, yet the expected effort required for employees to reach fluency dropped 8 points and confidence in training as the path to fluency dropped 14: leaders increasingly believe less training will be needed, not that they will spend less (the underlying question is about expected effort, so this reports effort expectations rather than budgets). Hiring will not close it either: recruiting advanced gen-AI skills is a top challenge at 49%.
The substantive reason generic fluency training fails agent builders is the jagged frontier. Capability is uneven: two tasks that look equally hard sit on opposite sides of an invisible line, AI helps most inside it and quietly hurts outside it, and novices, the population fluency training targets, gain the most exactly where it works and have the least basis for noticing when it doesn’t. A curriculum that teaches prompting produces builders who are confident in both regimes; one that teaches where the frontier bites produces builders who escalate.
For the same reason, the training gate belongs on capability rather than on attendance. Build permissions attach to demonstrated judgment on cases with known-wrong answers, the design training curriculum specifies, and certification tracks the criticality of what the builder is allowed to ship. Only 47% of respondents in an AI-engaged sample agree their employees generally understand the value of agentic AI, split 66/46/24 across adoption cohorts; since the cohorts are defined by adoption stage, that split is close to definitional and cannot be read as evidence that understanding causes adoption.
Where you stand
| Level | Looks like | Cheapest next move |
|---|---|---|
| Crawl | No training. People learn from each other and from whichever model they pay for personally. | Publish what is explicitly permitted, with three worked examples of good and bad use. A green zone teaches faster than a policy. |
| Walk | Generic AI-fluency sessions, optional, on top of unchanged workloads. Attendance is the metric. | Put the hours on the calendar and name the work they displace, then add one module on where the model fails silently. |
| Run | Role-differentiated curriculum tied to what a builder may ship; assessment on judgment rather than recall; office hours people actually use. | Add a known-wrong-answer exercise to the assessment. Anyone who accepts the plausible wrong output is not ready for tier 2. |
| Fly | Certification renewal on model or platform change; reviewer training distinct from builder training; incidents feed the curriculum within a quarter. | Trace one incident back to the missing lesson and add it. If nothing traces, your training is decorative or your incidents are unexamined. |
Concerns this dimension covers
- Oversight decay — untrained reviewers decay fastest, and reviewer training is the part everyone skips.
- Unreliable output acting on your systems — the failure a trained builder catches and an enthusiastic one ships.
- Quality debt and orphaned apps — what builders leave behind when nobody taught them what maintenance means.
- the jagged frontier — the capability shape every curriculum decision answers to, and what this dimension is grounded in.
Controls that answer them
- Human approval gates — a gate is only as good as the judgment behind it, and judgment is what training buys.
- Agent inventory and registry — registration is the first behaviour training has to produce.
Who does the work
- Training curriculum — the canonical program, keyed to criticality rather than to job title.
- The support model — enablement between formal training, and why response time sets the shadow rate.
- Citizen developer roles — who is being trained for what.
Sequencing and where this is checked
- Day 3 sequencing — which build permissions each training gate opens, and when to install it.
- Promotion gates — builder certification is an input to the gate, never a substitute for it.
- Evals — the organizational analogue of the same question: can anyone tell when output is wrong?
Open questions
- Has judgment-based assessment ever been validated against downstream incident rates? Not in anything we found. The argument for it is mechanistic, and the alternative, assessing recall, is known to measure nothing useful.
- Nobody has published a training-hours-to-competence figure for agent building other than a vendor’s course length.