How to Diagnose an AI Adoption Plateau

Confirm that the plateau is real

Specify the metric, eligible population, window, tool scope, and data source. Check for changes in logging, identity resolution, licenses, team size, seasonality, or the definition of “active.” Plot both counts and rates. A flat user count may hide growth if the eligible population shrank; a flat rate may hide different cohorts entering and leaving.

NIST’s AI RMF Core calls for context-specific goals, metrics, risk monitoring, human oversight, and ongoing management rather than static adoption assumptions (AI RMF Core). The CDC’s training-evaluation guidance recommends defining evaluation purpose and questions before selecting methods (evaluation planning). These sources support diagnosis discipline, not a universal benchmark for acceptable adoption.

Sources: [1] [2]

Segment by opportunity, not only department

Group people by relevant task opportunity, approved access, workflow frequency, and consequence. Someone whose work rarely benefits from the tool should not be treated like someone blocked from a suitable use case. Distinguish never activated, tried and stopped, occasional users, regular users, and people working outside captured systems.

Combine telemetry with interviews, workflow observation, support records, and artifact review. Ask what job people were trying to complete, what alternative they used, where review effort increased, and what prevented a second use.

Worked plateau diagnosis

Imagine a synthetic programme with an activation rate near 45% for three months. The first explanation offered is insufficient prompting skill. Segmentation shows a different picture: one group lacks access to the approved data connector, another uses AI only during quarterly planning, and a third stopped after managers required duplicate manual review.

The diagnosis board records each hypothesis, evidence, counterevidence, affected population, owner, and smallest test. The connector group gets a controlled access fix. The quarterly group is monitored on an appropriate window. The duplicate-review group pilots a clarified verification procedure. Prompt training is not the default intervention because the evidence does not isolate skill as the main constraint.

A failed analysis says, “Usage stalled, so employees are resistant.” It turns an aggregate pattern into a psychological claim. Repair it by checking opportunity, access, task frequency, policy, and downstream burden.

Choose interventions as tests

For each likely cause, define one bounded intervention and a decision rule. Access fixes should change eligible access. Workflow redesign should reduce a specified handoff or ambiguity. Practice should target an observed skill gap. Manager calibration should clarify acceptable use and review.

Monitor quality and risk alongside activity. More use with rising correction burden may not be desirable. Conversely, stable use with better task selection can be healthy.

Keep a hypothesis register

Write each explanation before examining the next intervention result. Include the expected signal, disconfirming signal, required data, affected group, and review date. For example, an access-barrier hypothesis predicts failed or abandoned authentication among otherwise relevant users; it is weakened if access succeeds and interviews reveal no useful task. The register prevents teams from relabeling every result as support for the favored story. AI may organize evidence against the hypotheses, but owners should adjudicate ambiguous observations and retain competing explanations.

Plateau review checklist

A plateau is a prompt to investigate the work system. It is not, by itself, evidence of resistance, low capability, or poor programme value.

• Metric definition, population, window, and instrumentation are stable.
• Counts and rates use the correct opportunity denominator.
• Relevant task frequency and approved access are visible.
• Qualitative evidence tests multiple explanations.
• Quality, review burden, and risk accompany usage.
• Each intervention targets a supported cause and has an owner.
• Results are bounded to the tested population and period.