01 · The metric errorA full funnel can conceal an empty system

Imagine an organisation that has provisioned an AI writing assistant to every employee. The launch dashboard glows green: licences allocated, accounts activated, training completed. Yet approved documents still move through the old route, staff copy fragments between unconnected systems, reviewers discover more invented claims, and the people doing sensitive work quietly avoid the tool. The organisation has distributed software. It has not established adoption.

The familiar measures are not useless. A login shows that access credentials worked. A prompt shows that somebody tried an interaction. Repeated activity may indicate usefulness. None establishes that a recurring piece of work reached a better outcome after context, verification, exceptions, rework, workload and harm were considered. An activity measure becomes especially misleading when it rewards the behaviour it claims merely to observe. Tie prompt volume to performance ratings and prompts will rise. The evidence may become less truthful while unsafe use moves out of sight.

This article defines adoption as an emergent property of a work system. People adopt a changed way of accomplishing work inside a web of demand, tasks, roles, authority, information, capability, capacity, incentives, professional obligations, controls and consequences. Low use can therefore be a precise signal that the system is incomplete. High use can coexist with poor adoption when generated material creates errors, escalations or invisible checking labour. The governing question is not “Why will people not use it?” It is “What conditions make an appropriate use episode possible, worthwhile, legitimate, safe and learnable?”

Current evidence supports the distinction but does not license a universal adoption rate. UK government and ONS studies report uneven use across sectors and firm sizes, alongside barriers involving skills, cost, business need and integration. Their survey measures are population-bound and often self-reported. OECD employer and worker surveys add working conditions, training, trust and consultation, while the ILO finds heterogeneous changes to tasks, autonomy and job quality. These sources tell leaders where to look. They do not turn a reported productivity impression into audited value or causal proof.

02 · The unitTrace the outcome-producing use episode

The smallest useful unit is not a user. It is a recurring episode of work in which a defined actor uses or declines AI, obtains the necessary context, exercises judgment or delegated authority, completes controls, manages exceptions, hands work onward and contributes to an intended outcome.

Each episode record identifies the eligible work population and exclusions; the person, team and agent roles; model, tool, data, policy and workflow versions; required inputs and available context; and the choice to use, edit, override, escalate, abstain or stop. It captures verification, delay, rework and recovery, as well as notice, contestability and remedy for an affected party. It then connects output and decision to outcome, time, workload, cost, harm and distribution. Finally, it names the owner who can act on the learning signal.

A login cannot satisfy that record. Neither can a draft abandoned outside the document workflow. A successful episode may contain no generated output at all: a risk reviewer can inspect the case, recognise that the model is outside its validated boundary and document an abstention. That is not failed adoption. It is a system producing the right controlled behaviour. Conversely, an agent who copies a confident but incorrect answer into a customer record has completed a prompt and perhaps a transaction, but not a successful episode.

03 · The conditionsSix coupled pairs make work possible

The Work-System Adoption Map is this article's synthesis of twelve conditions in six coupled pairs. It is a map, not a maturity scale. A population may have excellent capability and no authority, or sound integration and no valuable demand. Averaging the conditions would conceal the binding constraint.

Demand and outcome ask whether a real recurring trigger enters the workflow and whether the intended customer, worker, operational or financial result is explicit. Task and flow ask whether the AI contribution fits the whole task bundle and whether upstream inputs and downstream handoffs exist at the moment of work. A plausible demonstration fails both tests if it solves an occasional fragment outside the production sequence.

Role and authority identify the user, decision owner, verifier, exception owner and affected party, then test whether permission, delegation, override, escalation, abstention and stop rights can actually be exercised. Context and integration cover relevant data, policy, history and versions as well as the systems, channels and devices where work occurs. Nominal access on a desktop is not reachability for a night-shift worker using a managed handset.

Capability and capacity separate knowing how to frame, verify, challenge, explain and recover from having the time, staffing, support and cognitive room to do so. Legitimacy and learning connect incentives, professional standards, privacy, fairness, worker voice and social norms with feedback, incidents and outcomes reaching somebody empowered to change or stop the system. These conditions are population-specific. The binding condition for a remote contractor may be authentication; for a clinician, liability; for an agent, queue pressure; for a customer, recourse.

04 · The denominatorA population ledger prevents convenient arithmetic

The Adoption Population Ledger moves from Eligible to Reachable, Enabled, Attempted, Repeated, Embedded and Outcome-producing. Every stage has its own numerator, denominator and exclusions. People, tasks, episodes, decisions and outcomes remain separate units.

Eligible means the capability is relevant and permitted for defined work and people. Reachable means that eligible population can obtain it in the actual location, shift, channel and device. Enabled adds context, authority, capability, time and support. Attempted records a real episode, including documented abstention where that is the correct response. Repeated distinguishes work from a demonstration. Embedded means the changed practice survives normal workload, exceptions, handoffs, staff changes and competing priorities. Outcome-producing means verified benefit remains after error, rework, control load, harm and distribution are considered.

Cohorts cannot be silently collapsed. One person can perform many episodes, and one episode can involve several people. A global percentage can hide an inaccessible language group, an excluded contractor population or a high-consequence task that should rarely use automation. The ledger requires leaders to explain each exclusion and choose the next decision supported by that stage, rather than treating movement to the right as inherently good.

05 · The diagnosisFriction is evidence before it is an obstacle

The Adoption Friction Ledger records observed behaviour without assigning a motive. It names the population, task, context, frequency and consequence; maps the affected condition; lists evidence and rivals; classifies the friction as protective, harmful or unresolved; and assigns an experiment, owner, expiry and rollback.

Protective friction is the deliberate difficulty that safeguards consequence, review or recourse. A second approval before an irreversible payment may be appropriate. Harmful friction adds no corresponding protection: repeated authentication, unavailable context, duplicate entry or a queue target that makes verification impossible. Unresolved friction is a hypothesis under test. A worker's decision to avoid a summarisation aid may reflect poor accuracy, slower review, unclear liability, professional identity, an accessibility problem or simply a weak product. Calling all of these “resistance” substitutes judgement for observation.

The ledger must also expose the cost of its own intervention. Removing a review step may reduce waiting while increasing error exposure. Adding a mandatory explanation may aid contestability but overload staff or exclude users of assistive technology. The proposed change therefore carries an enabling signal, an adverse signal, the new workload or risk it can create, and a date when the organisation will retain, adapt, narrow, replace or stop it.

06 · Before rolloutAssemble a minimum viable work system

Before an adoption campaign, the Minimum Viable Work System Test asks whether there is enough system to adopt. It is a go-to-work test, not a declaration that the model itself is safe or the business case is proved.

A consequential recurring population must exist, and the non-AI path and option not to proceed must remain credible. The team observes the full task bundle, including hidden verification and coordination. Inputs, context, integrations and handoffs work. Roles and authority are complete. Users can recognise limits, verify, escalate and recover. Workload and support are funded rather than pushed into personal time. Professional standards, worker voice, privacy, accessibility and affected-party rights are respected. Incidents, exceptions, recovery and recourse have owners. Evidence can be gathered without surveillance theatre. A stop, redesign or retirement path exists.

If these elements are absent, communications cannot manufacture them. Champions may temporarily bridge gaps through personal effort, but that produces a biased demonstration: unusually motivated people absorb extra load that the wider population cannot sustain. Mandatory usage conceals the problem more aggressively. It turns an incomplete system into individual exposure and converts the measurement process into a performance signal.

07 · The changeRepair the binding condition with a bounded test

The Adoption Intervention Matrix converts diagnosis into the smallest change that can reveal whether a condition is binding. Training is one option among workflow, technology, context, authority, capacity, incentives and controls. It is never the automatic answer to low use.

Each intervention names the condition and friction, affected population and episode, owner, contributor and independent challenger. It states the proposed change, expected enabling signal, adverse signal, comparison and test duration. It also predicts workload, job-quality, accessibility, fairness and privacy consequences. At expiry the owner must retain, adapt, narrow, replace or stop. An indefinite pilot merely postpones the decision while normalising unsupported work.

Suppose an approved sales assistant is ignored while staff use a public model. The smallest revealing change may be governed access to a limited slice of approved customer context for one team, not a global training course. Expected signals include fewer manual transfers and more completed approved episodes. Adverse signals include sensitive data exposure, slower customer response or new retrieval errors. If context repair does not displace shadow use or improve outcomes, usefulness, latency or workflow fit remains a rival explanation.

08 · The panelFive views must disagree in public

The Adoption Evidence Panel holds five views apart: population, episode, work, outcome and system health. There is no weighted total. Its value lies in disagreement, because disagreement reveals which attractive story is incomplete.

The population view shows who and what is eligible, reached, enabled and excluded, with reasons. The episode view shows attempts, repeats, abstentions, edits, overrides, escalations, exceptions and recovery. The work view makes time, waiting, rework, cognitive load, queue transfer, hidden checking and coordination visible. The outcome view compares quality, speed, service, revenue, cost, safety, harm and distribution with a credible alternative. The system-health view watches support demand, incidents, drift, versions, worker trust, affected-party feedback and unresolved friction.

High activity with rising rework is not a puzzling contradiction; it is a clear panel. Low activity among an ineligible population is not a motivation crisis. A fall in use after tighter boundaries can be an improvement if the remaining episodes are better calibrated and affected parties receive stronger recourse. Self-reported time saved can be an exploratory signal, but it cannot carry the outcome view without workflow and comparison evidence.

The evidence boundary

Brynjolfsson, Li and Raymond studied 5,179 customer-support agents in one staggered deployment and found heterogeneous effects, including larger gains for less experienced workers. It is strong field evidence for population-specific outcomes and possible learning, not a productivity constant for every occupation. NIST's human-AI material reinforces the need for context, differentiated roles, challenge and override evidence. Parasuraman, Sheridan and Wickens show that allocating functions to automation changes human work rather than deleting it; Bainbridge's warning endures because people may inherit abnormal conditions after routine involvement has weakened.

Acceptance research supplies the strongest rival. Davis's perceived usefulness and ease of use, and the performance expectancy, effort, social influence, facilitating conditions and contextual moderators in UTAUT, describe meaningful individual differences. A system account can become patronising if it erases agency, preference, professional identity or experience. Orlikowski's practice lens and May and Finch's work of embedding also warn against treating technology as a fixed object with automatic effects.

The Adoption Revalidation Rule reopens the population ledger, map and representative episodes whenever the task, population, workflow, model, data, interface, vendor, autonomy, policy, regulation, incentive, staffing, skill mix, workload or affected-party consequence changes materially. Historical usage describes an earlier configuration. It does not prove that the current work system remains adopted.

09 · Ten controlsCases that separate activity from adoption

These compact cases apply one discipline: define the eligible population, trace the episode, inspect all twelve conditions, classify friction, test the minimum system, choose a bounded intervention and read the five evidence views. The named owner must be able to change the binding condition.

Across the cases, repair is not synonymous with scale. The writing assistant needs workflow fit; customer support needs redesigned controls; clinical abstention needs authority and workload repair; shadow use exposes unmet demand and privacy risk; claims work needs capacity and incentive change. Selective risk use is retained, junior authority is gated, scheduling is paused for rights, the small business chooses a lighter system, and the mandated target stops. The same usage statistic would have prescribed the wrong action in nearly every case.

10 · Interactive instrumentWork-System Adoption Diagnostician

Use the instrument to structure a case, not to score a person. Select the evidence state for each field or load a named scenario. The result chooses one of exactly five routes and exposes the population ledger, episode, condition map, friction, minimum-system gaps, intervention, obligations, signals and revalidation trigger.

The instrument intentionally refuses three outputs. It does not diagnose “resistance” from low use, recommend training because a person hesitated, or calculate a universal adoption score. It directs attention to the condition an owner can change and preserves both the non-AI path and the possibility that no repair can make the intended outcome worth pursuing.

11 · The operating conclusionAdoption belongs to the work, not the personality

A sound adoption practice makes the work system answerable for the behaviour it enables. It refuses to make a worker personally accountable for missing context, impossible queues, unusable rights or a low-value product.

The sequence is practical. Start with a recurring outcome and an eligible population. Trace representative use episodes, including abstention, escalation and recovery. Map the twelve conditions and preserve cohort denominators. Record friction as observed behaviour with rivals. Test whether a minimum system exists. Change the smallest plausible binding condition for a bounded period. Read population, episode, work, outcome and system health together. At expiry, retain, adapt, narrow, replace or stop. Revalidate when the configuration changes.

This reframes leadership's role. Communications may explain a changed practice, and learning teams may build capability, but neither owns adoption alone. Product owners must repair workflow and integration. Operations leaders must fund control work. Accountable executives must make authority usable. Workforce and policy owners must protect voice, privacy, job quality and recourse. Risk teams must preserve necessary friction. Outcome owners must reject activity that does not survive comparison. The organisation adopts only when these contributions meet inside ordinary work.

The discipline is deliberately less flattering than a rising login chart. It can reveal that a celebrated launch excludes a shift, that apparent time savings moved into review, that non-use protected a customer, or that the rational decision is to retire the capability. That is precisely why it is useful. Adoption is not compliance with a product metric. It is the repeatable capacity of a real work system to produce an intended outcome under its actual constraints and obligations.