Human-ready is AI-ready

Investment without adoption

Organisations have the instruments. Licences have been bought, pilots are running, policy frameworks have been written or sit on the roadmap. Yet the return does not arrive. The cause rarely lies in the technology, but in the distance between what an organisation rolls out and what its people do with it.

That distance is documented. Cisco’s AI Readiness Index classifies 13% of the organisations surveyed as fully AI-ready, a share that held steady three years running across more than 8,000 respondents in thirty markets (Cisco, 2025). The Enterprise AI Maturity Index measures on a different scale: in 2025 fewer than 1% of respondents scored above 50 out of 100, with an average that fell from 44 to 35; in 2026 that average recovered to 51 (ServiceNow, 2025).

On the governance side the picture is one of limited control. In a survey of 900 CEOs, 96% believe employees use generative AI without approval, and 79% are concerned about legal exposure from AI agents (Dataiku/Harris Poll, 2026). Governance is trailing usage. That is less a control problem than an information problem: people are working with AI at full tilt, just not where anyone is looking.

On the employee side the pressure has risen. The Humans at Work Barometer, covering more than 6,000 employees in ten countries, finds that 51% feel more pressure than a year earlier and 48% end most working days mentally exhausted (Workhuman, 2026). AI is not the only cause of that, but it is the change stacked on top.

The consequence is measurable in abandoned projects. In 2025, 42% of companies dropped the majority of their AI initiatives, against 17% a year earlier (S&P Global, 2025). Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on rising costs, unclear business value or inadequate risk controls (Gartner, 2025).

Four building blocks, one complete picture

In practice, AI readiness is measured as a technical property of the organisation. That is one half of the question, and not the half where it usually goes wrong. A complete framework has two halves: is the employee ready for it, and can the organisation carry it. The first half breaks into three axes, the second into seven dimensions. Four parts in all, each with exactly one primary source.

Part 1 · The employee

1a. Usage

What someone actually does with AI: which tools, how intensively, at what cost. This axis is objective and usually already sits in existing telemetry and metadata, without anyone having to be surveyed. It reacts in weeks, which makes it the only axis that runs ahead of the other two.

1b. Engagement

How someone relates to the work and to the change: understanding of the why, trust, perceived autonomy, willingness to put in effort. Measuring engagement alone no longer suffices in an AI-driven environment (WTW, 2026), but without this axis there is no explanation at all for what the usage data shows. It reacts in months and has to be actively collected.

1c. Performance

What someone actually delivers. This axis is left out of AI readiness measurements almost without exception, even though it determines whether usage and enthusiasm land in any result. It reacts in quarters and comes from the organisation’s own existing review and results cycle.

Part 2 · The employer

Data quality, infrastructure, governance, policy and the chosen use cases: seven dimensions together that determine whether the organisation can carry AI. This half is not the bottleneck, but it is the precondition: where governance is missing, usage becomes invisible rather than absent. It is established from interviews and document analysis, supplemented with the system data the organisation already keeps.

Why the four parts do not run in parallel

That is the analytical starting point of this framework. Whoever uses AI most is not necessarily the strongest performer; nor is whoever is most engaged. Measuring on one axis therefore produces systematically wrong conclusions. Perception alone labels critical but productive employees as checked out. Usage data alone does not explain why someone has stalled. Performance alone says nothing about how long that result is sustainable. And a strong organisational score says nothing about the spread underneath it.

AI anxiety is the factor that connects the axes. It is not a general fear of technology but a complex psychological response to rapid, systemic change (HRD America, 2026). A survey of more than 1,500 full-time employees across five countries, in early 2026, found that 24% reported worsened mental health from information overload and 23% a reduced sense of control over their own future (Spring Health, 2026).

From average to segment

A readiness score is an average, and an average describes a reality nobody in the organisation experiences that way. A score of 68 out of 100, strong on talent and weak on governance, is usable steering information at board level and at the same time unsuitable as a basis for interventions.

The first segmentation layer is organisational. Beneath an average of 68, IT may sit at 95 and Finance at 55. That difference determines where an intervention has effect and where it is waste. Teams are rarely homogeneous, though, so this layer keeps producing averages too.

The second layer is individual and crosses the dimensions against each other. The matrix below sets engagement against performance and yields nine segments, each with its own intervention logic.

The largest group sits in the middle. The engine, 24 employees or 19%, scores average on both axes and forms the load-bearing layer of the organisation. This group needs maintenance, not intervention, and gets routinely skipped in adoption programmes in favour of the extremes.

The retention risk sits on the right of the matrix. High performer, flight risk (12 employees) and Critical flight risk (9) perform at the top while their engagement drains away. This is not ordinary turnover risk: the loss hits productive capacity directly, and the replacement cost of a top performer far exceeds the cost of a timely conversation.

Top left sits the reverse profile. Hidden potential (4 employees) is strongly engaged but not yet performing; Struggling willing (6) wants to but is stuck. The margin is greatest here, and the instrument is skill building and coaching, not motivation.

Usage as a leading indicator

The matrix in figure 1 shows two of the three dimensions. Add usage as a third axis and the matrix stays the same but gains depth: the same nine segments exist in three layers, from low to high AI usage. An employee is no longer in a box but at a coordinate.

Why this axis yields the most

The three axes measure the same story at different moments, and that is the substantive reason to add the third. Performance records what has already happened and moves in quarters. Engagement records the current relationship to the work and moves in months, at a frequency tied to survey moments. Usage is behaviour, is continuously measurable from existing telemetry, and changes in weeks. It is the only axis in the framework that runs ahead of the other two, and therefore the only one that makes intervention possible before the loss has occurred.

Eight profiles that only become visible with the third axis

Each of the profiles below is hidden in figure 1 inside an existing box, together with people who need the opposite approach. The usage axis pulls them apart. Not every combination of the three axes is equally meaningful: the eight below are the profiles that stand out most sharply in practice and matter most for an intervention.

From resolution to choice

Every additional dimension raises the resolution, and with it the number of segments an organisation cannot possibly serve all at once. The follow-up question becomes strategic rather than analytical. Organisations usually pick one of three routes. All three are defensible. None of the three is executable without segmentation.

From symptom to cause

A segment locates a problem; it does not explain it. Two employees can end up in the same box for opposite reasons, and that cause does not sit in the KPIs themselves. Usage, engagement and performance tell you what is moving. To know why, that movement has to be correlated with the drivers behind it.

Classic engagement models do exactly that. Job Demands-Resources correlates job characteristics — workload, autonomy, social support, resources — with outcomes such as work engagement and burnout. Self-Determination Theory correlates the degree to which autonomy, competence and relatedness are met with motivation and wellbeing. Neither measures a score in isolation; both explain a score by tying it to what lies underneath.

The elli approach extends that logic into the AI domain. Usage, engagement and performance are the KPIs; alongside them sits a set of key drivers, job characteristics and job experience specific to AI adoption: perceived usefulness, autonomy over the tool, trust in the policy, workload, leadership support, psychological safety to experiment. By correlating KPIs against drivers, it becomes visible which driver most strongly explains a score in a specific segment, in the same way JD-R and SDT do for classic engagement measurements.

That distinction is what determines the intervention. A low usage score that correlates strongly with trust in the policy calls for communication. The same score, correlated with perceived autonomy, calls for a different kind of tool or a different role. The KPI points at the place; the correlation with the drivers points at the cause.

On the communication side, a systematic review confirms the mechanism: when employees understand why change is happening and how they fit into the future, wellbeing improves and engagement rises (Frontiers in Psychology, 2026). Clear and consistent communication shrinks the space in which uncertainty grows — itself a driver taken into the correlation.

In practice this means the KPI measurement is a starting point and not an end point. The segments follow from usage, engagement and performance; the cause follows from the correlation with the drivers, supplemented with open answers, conversations and context. Only then is an intervention more than a guess.

Methodology in four to six weeks

The elli model, based on the ADKAR model (Awareness, Desire, Knowledge, Ability, Reinforcement), offers a usable ordering of the causes that emerge from the measurement. Someone who does not know why AI is being introduced does not have a training problem but a communication problem. Someone who wants to but cannot does not have a motivation problem but a skills problem. Someone who can but does not runs into workload, prioritisation or missing reinforcement.

The framework becomes operational through three data sources that check each other, followed by analysis, reporting and follow-up.

• Survey. A questionnaire maps maturity, adoption and perception across seven dimensions, broken out by team. Open questions supply the cause information that closed items cannot yield.
• Telemetry. Usage and token data come from metadata that is already available, surfaced via OpenTelemetry. They show actual usage, spend per team and the anti-patterns, with a cost estimate per pattern.
• Interviews and document analysis. Policy, governance and context are captured from existing documents and supplemented with conversations in key positions. This source corrects the survey wherever perception and the formal framework diverge.
• Analysis and scoring. Every dimension gets a score from 0 to 100, weighted by priority, alongside a dimension profile showing strong and weak spots. Five maturity levels translate the composite score into one readable position: waiting, exploring, adopting, scaling or embedded.
• Lead time. Week 1 as-is scan, weeks 2 and 3 the survey, week 4 analysis, weeks 5 and 6 a board-ready readout with recommendations per theme, each with an owner and an intended result.
• Privacy and compliance. Reporting is aggregated only, with a minimum of five respondents per group, GDPR and AI Act compliant. That choice is not merely legal: a measurement that undermines trust worsens exactly the problem it sets out to solve.
• Follow-up. Readiness is not a snapshot. Pulse and impact measurements at fixed intervals show whether an intervention actually moves the intended segment, or merely lifts the average.

Preparation begins with sight

The obstacle to AI adoption is rarely technical. It sits in the difference between what an organisation rolls out and what its people can, want to and are allowed to do with it — and precisely that difference stays out of view in most readiness measurements.

Three dimensions together yield the segments you can steer on. Segmentation exposes the retention risks and the growth margins an average hides. The third axis, usage, then determines how much time is left to do something about it. And the cause analysis determines whether an intervention has to address communication, skill or reinforcement.

For most organisations AI is not the biggest risk. Being unprepared for it is. And preparation does not begin with the technology, but with sight of where the people stand.

elli brings the three axes together in one programme, from measurement to segments to intervention.

References