The Capability-Institution Gap

• A field study of 5,172 customer support agents found 15% more issues resolved per hour with AI assistance, with the largest gains for less-experienced workers — a single deployment in one firm, not a cross-occupation effect.
• The ILO-NASK occupational exposure index estimates roughly one in four workers globally is in an occupation with some GenAI exposure — technical potential at task level, not a job-loss forecast; the authors judge transformation more likely than full replacement.
• A high-school mathematics experiment found unguarded AI assistance could improve practice results while weakening later unaided performance, with tutor safeguards mitigating the harm — education evidence that motivates a workplace test rather than proving employee deskilling.
• In a 2025 trial, 16 experienced developers took 19% longer on familiar repositories when AI was allowed — a narrow, dated setting whose 2026 follow-up could not reliably estimate the current effect.
• The OECD identifies skills shortages as an adoption barrier and highlights data interpretation, problem-solving and managerial capability — survey associations that do not establish a single causal effect.

Concepts, frameworks, models and decision principles developed by Ravinder — practitioner synthesis, distinct from the external evidence cited above.

• The Capability-Institution Gap model — the author's practitioner framework for locating where changed individual capability meets unchanged institutional design. It is a mismatch to investigate, not a measured distance or a claim that every organization is behind.
• An eight-area diagnostic (work, tasks, role, expertise, experience, management, performance, governance) that asks for observable evidence rather than a general rating of AI readiness.
• A six-step L&D operating sequence — work, performance, capability, experience, support, evidence — that moves the function from the training request to the performance requirement.
• A know / find / do / judge design model that treats knowledge, retrieval, performance and judgment as distinct requirements instead of one content requirement.
• A 90-day bounded pilot with three decision gates (diagnose, test, decide), a pilot scorecard measuring workflow performance, human judgment, business value, control reliability and development access, and a one-page meeting worksheet.

• What has AI made faster, possible or unnecessary in this workflow?
• What should AI perform, assist or leave to people — and who decided?
• Does the responsibility still fit the changed task mix?
• What must people understand to judge the result, and which developmental tasks may disappear?
• What must the manager change for faster work to become better work?
• What result matters beyond output volume, and who may act, approve, stop and escalate?

Composite scenarios for illustration — they describe hypothetical situations, not client results.

• A junior analyst produces a market assessment in two hours with AI, but cannot explain the demand assumptions or identify which figures are estimates — the document proves assisted output, not dependable judgment.
• Six employees each produce a weekly report; AI cuts preparation from hours to minutes, yet the manager still requests six documents in the old format and the team measures success by punctual submission.
• An organization reports 98% completion of a 45-minute conduct course with an AI tool available to explain policy — neither fact shows whether employees will act appropriately when a manager asks them to bypass a control under deadline pressure.
• An illustrative weekly calculation releases 12 hours of preparation across six employees, spends 3 on verification and rework and 2 on protected practice, leaving 7 hours of capacity — hypothetical figures for illustration, not a forecast or reported result.

• Begin the needs analysis one step earlier: before accepting a request for AI training, examine the outcome, current task mix, bottlenecks and decision rights — and observe actual work alongside the job description.
• Assess judgment in context: give employees an output containing a plausible mistake or conflicting assumption, and score the reasoning and corrective action rather than the fluency of the explanation.
• Managers decide whether faster work becomes better work — which reports still serve a decision, what standard output must meet, and where released capacity goes.
• Treat released time as capacity, not cash. The hours become savings only when actual spending falls; report capacity released separately from cash savings and include review, rework and development time in the calculation.
• Give every responsibility an owner and a usable control. A reviewer needs competence, evidence, authority, time and a route to intervene; if any is missing, a sign-off creates the appearance of control without its substance.

• Run a 90-day pilot across up to three AI-exposed roles: one consequential workflow per role, a baseline, agreed human-AI responsibilities and a small tested change.
• Use the diagnostic to locate the actual constraint, and assign an owner to each unresolved decision before commissioning development work.
• Take three decisions into the next leadership meeting: select the work with an accountable sponsor, agree the change and the human capability that must remain dependable, and require evidence with a 90-day decision date.
• Scale only when the workflow produces an agreed improvement, controls work in practice, and people demonstrate the capability required of them — and stop or narrow the test if material errors go undetected or authority is unclear.

Research shows gains in some settings and weaker results in others. The Capability-Institution Gap is the author's practitioner framework — a hypothesis to test in context — not a validated scale, diagnostic instrument or universal finding. The diagnostic, operating model and pilot gates are recommendations, not research instruments or universal thresholds.

• The cited evidence combines workplace studies, occupational analysis, education research and policy synthesis — methods that answer different questions and should not be combined into a single estimate of AI value.
• Some findings are dated snapshots and some rely on self-report; the school experiment does not establish workplace deskilling.
• The analyst, management-reporting and compliance scenarios are illustrative composites; their figures describe hypothetical situations, not client results.
• The long-term effects on employment, expertise and career progression remain uncertain; the pilot is designed to produce local evidence without relying on a precise forecast.