AI Capability · Whitepaper · 2026

Building an Organisation That Can Actually Use AI

Why lasting AI capability depends on more than tools, and how leadership, skills and operating practices need to develop together.

Read Whitepaper

Executive Summary

Eighty-eight percent of employees use AI at work. Five percent are actually using it to transform how they work.

EY's 2025 Work Reimagined Survey, gathering responses from 15,000 employees and 1,500 employers across 29 countries, found that while 88% of employees now use AI at work, they mostly confine it to basic tasks such as search and summarisation. Only 28% of organisations have built what EY calls a "Talent Advantage", the effective integration of people and technology needed to unlock AI's full value, and just 5% of employees report they are maximising AI to genuinely transform their work. A separate EY analysis found that organisations are missing out on up to 40% of potential AI productivity gains specifically because of gaps in talent strategy, not technology.

Workera's 2026 AI Skills Enterprise Benchmark Report, drawing on 88,753 verified skills assessments across Fortune 500 companies and the US federal government, found that only 13% of employees are "Accomplished" in the skills needed to work with AI agents before any upskilling, the weakest of fourteen capabilities measured. More troubling still, only 11% of employees can accurately judge their own skill level, and nearly seven in ten either overestimate or underestimate what they can actually do, even though 85% of learning and development leaders report confidence in that same self-reported data.

What happens when an organisation scales the tool faster than it builds the capability around it was demonstrated publicly by Klarna. In early 2024 the Swedish fintech announced its AI customer service assistant was doing the work of 700 human agents and cut staff accordingly. By May 2025, its own CEO told Bloomberg the result was, in his words, "lower quality", and the company spent the following year quietly rebuilding the human capacity and hybrid operating practices it had removed.

This whitepaper sets out why lasting AI capability is not a technology purchase but the compounding product of three things developing together, leadership attention, accurately measured skill, and operating practices built around the tool, and why an organisation strong in one or two of these while weak in the third does not yet have AI capability at all.

01 · Why Tools Alone Don't Create Capability

The 83-point gap between who uses AI and who transforms their work with it is the actual battlefield.

EY's 2025 Work Reimagined Survey is one of the largest studies of its kind, 15,000 employees and 1,500 employers surveyed across 29 countries, and its central finding cuts against the prevailing narrative of AI adoption. Usage is nearly universal: 88% of employees now use AI in their day-to-day work. But that usage is shallow, concentrated in search, summarisation and other basic tasks, and 64% of the same employees report a perceived increase in their overall workload over the past year despite it. Only 28% of organisations have achieved what EY terms a "Talent Advantage", genuinely integrating people and technology rather than layering one on top of the other, and just 5% of employees say they are maximising AI to actually transform how they work.

EY's own analysis of the same data draws a direct causal line: organisations are missing out on up to 40% of the productivity gains AI could realistically deliver, and the cause is a gap in talent strategy, not a gap in the technology itself. This is a materially different diagnosis than the one most AI investment currently assumes. Licences, seats and model access have never been the scarce resource. The capability to use them well has been scarce throughout.

The 83-point gap between the 88% who use AI and the 5% who are using it transformatively is not a rounding error in an adoption curve. It is the clearest single measurement available of how much of current AI investment is producing genuine capability, and how much is simply producing usage.

02 · The Klarna Lesson: What Happens When the Tool Outruns the Capability

Klarna did not lack an AI tool. It lacked the organisation built around it, and rebuilding that took over a year, in public.

In early 2024, Klarna announced that its AI customer service assistant was performing the equivalent work of 700 human agents, and the Swedish fintech reduced staffing on the strength of that claim. By May 2025, the picture had changed materially. Klarna's own chief executive, Sebastian Siemiatkowski, told Bloomberg plainly that the result of the shift had been, in his words, "lower quality." Customers described the automated responses handling complex issues as generic, repetitive and insufficiently nuanced, and satisfaction scores fell on those specific interaction types even as aggregate volume-based metrics continued to look healthy, the aggregate numbers were, in effect, masking exactly where the quality was breaking down.

Rather than reversing course entirely, Klarna spent the second half of 2025 and into 2026 quietly rebuilding its human customer service capacity, moving to a deliberately hybrid operating model: AI handling routine, high-volume queries, and human agents handling escalations, complex cases and high-value customer interactions. The company began hiring again, specifically targeting flexible remote agents, students, parents, and workers in rural areas, rather than simply restoring the roles it had cut.

Klarna's AI assistant worked. The organisation around it did not yet exist: the escalation practices to catch quality decline before it showed up in noisy aggregate metrics, the operating model to blend AI and human capacity deliberately rather than as an all-or-nothing bet, and the judgement, held somewhere in the organisation, by someone accountable, about which interactions genuinely belonged to the AI and which did not. Building that took over a year, after the fact, in full public view. Building it in advance would have taken a fraction of the cost and none of the reputational exposure.

03 · The Skills Paradox: Why Most Organisations Don't Know What Their People Can Do

Employees and their managers get the skills question wrong roughly seven times out of ten, and most training is built on top of that error.

Workera's 2026 AI Skills Enterprise Benchmark Report is drawn from 88,753 individually verified assessments covering 32,422 people across fourteen AI and data capabilities, spanning Fortune 500 companies in professional services, pharmaceuticals, financial services and consumer goods, alongside the US federal government. The pattern it uncovers is consistent: workforces are strongest wherever familiar human skills, communication, ethics, reasoning, meet a new tool, and weakest wherever the technology itself is genuinely new. The strongest benchmarks were Data Storytelling Essentials (231 out of 300), AI and Data Communication (230) and Responsible AI Essentials (229). The weakest were Deep Learning Fundamentals (163) and the capability Workera labels "Beyond LLMs: Prompts, Agents, and Retrieval-Augmented Generation," where only 13% of employees were rated "Accomplished" before any upskilling took place, the single lowest score across all fourteen capabilities measured.

The more consequential finding sits underneath those numbers. Only 11% of employees can accurately assess their own skill level, and Workera's research found that nearly seven in ten either overestimate or underestimate what they can actually do, a pattern that puts fair promotion decisions, targeted learning investment and genuine AI readiness all at risk simultaneously. And yet 85% of learning and development leaders report confidence in exactly this kind of self-reported skills data. Training programmes, upskilling budgets and deployment decisions across most organisations are currently being built on a foundation that is wrong roughly seven times out of ten.

The genuinely encouraging finding in the same report is that targeted upskilling works, dramatically, once it is aimed at an accurately measured gap rather than a self-reported one. Employees who received targeted upskilling in Data Visualisation and Storytelling improved by 72% on average; Generative AI Essentials scores rose 53%; and in Responsible AI specifically, the share of employees rated "Accomplished" rose from just 25% before upskilling to 81% afterward. The capability gap closes quickly. It only fails to close when organisations do not first measure where it actually is.

The Capability Gap

Using AI and being able to use it well are measurably different things.

Independent 2025 and 2026 research consistently shows a wide gap between AI usage and verified organisational capability to use it effectively.

Maximising AI to Transform Work

Of employees say they are maximising AI to genuinely transform their work, despite 88% using it daily.

EY, 2025 Work Reimagined Survey

Accomplished in Agentic AI Skills

Of employees are verified as "Accomplished" in working with AI agents before any targeted upskilling.

Workera, 2026 AI Skills Enterprise Benchmark

Have a Mature Upskilling Programme

Of organisations have a mature, organisation-wide AI upskilling programme, despite 82% offering some training.

DataCamp, 2026 State of Data & AI Literacy Report

Together, these figures show that AI usage, verified skill and structured upskilling are three separate gaps, and closing any one of them does not close the others.

04 · Leadership: The Missing Ingredient Skills Data Can't Supply

Training exists almost everywhere. Leadership attention to whether it actually works exists almost nowhere.

DataCamp's 2026 State of Data & AI Literacy Report, surveying more than 500 enterprise leaders across the US and UK in partnership with YouGov, captures the paradox directly: 82% of enterprise leaders say their organisation provides some form of AI training, yet 59% of those same leaders report their organisation still has an AI skills gap. Only 35% describe their organisation's upskilling programme as mature and organisation-wide; for the rest, training tends to be fragmented, optional, and disconnected from the actual tasks employees are expected to perform with AI day to day.

IDC's broader workforce research, summarised by Workera, estimates that unresolved skills shortages could cost the global economy up to $5.5 trillion through product delays, quality issues, missed revenue and impaired competitiveness, with more than 90% of enterprises projected to face critical skills shortages. The barriers organisations cite are telling in themselves: lack of talent (46%), data privacy concerns (43%), poor data quality (40%), high implementation cost (40%) and unclear return on investment (26%), a list dominated by organisational and human factors, not technical limitations of the AI itself.

None of this is a training-content problem waiting on a better course library. It is a leadership-attention problem. Training remains fragmented and optional in most organisations precisely because it has not yet been treated with the same seriousness, the same budget discipline and the same executive ownership as the technology purchase that came before it, and EY's finding that talent-strategy gaps alone account for up to 40% of missed AI productivity gains suggests that closing this attention gap may be the single highest-leverage move most organisations have not yet made.

05 · Building Capability as One System, Not Three Separate Projects

Leadership, skills and operating practices have to move together, because each one exposes the weakness of whichever of the other two is neglected.

Treat capability building with the same governance seriousness as the technology deployment itself: name an accountable owner for organisational readiness specifically, separate from whoever owns the AI rollout, so that "did the organisation actually get better at this" has as clear a line of ownership as "did the tool get installed." Measure skill before training rather than after: with employees and their managers wrong about actual capability roughly seven times out of ten, any training plan built on self-report alone is, on the evidence, aimed at the wrong target from the outset.

Build the escalation and quality-monitoring practices before scaling the tool, not after a Klarna-style correction forces the issue, and track quality metrics granular enough to catch decline on specific interaction types, since aggregate volume metrics demonstrably can and do mask exactly this kind of deterioration. Expand AI's role in any given workflow only as fast as verified skill and the operating practice around it can keep pace, matching autonomy to demonstrated readiness rather than to enthusiasm or a public announcement already made.

And measure the right outcome. Workera's data shows upskilling produces real, large improvements, a 72% average gain in one capability, a rise from 25% to 81% "Accomplished" in another, when it is targeted at an accurately measured gap. The organisations pulling ahead are treating the closing of that gap as the actual result to manage toward, not the initial announcement that a tool has been deployed.

06 · What Building an Organisation That Can Actually Use AI Demands of Leadership

The evidence points toward five practical changes in how organisations should build lasting AI capability.

Name an owner for capability, not just for the rollout. With up to 40% of AI productivity gains lost to talent-strategy gaps rather than technology gaps, organisational readiness deserves the same single point of accountability as the deployment itself.

Measure skill before you train it. With only 11% of employees able to accurately judge their own capability, replace self-reported confidence with verified assessment as the starting point for any upskilling investment.

Build the operating practices before you need them. Escalation paths, quality monitoring granular enough to catch decline masked by aggregate metrics, and clear rules for what belongs to AI and what does not, are cheaper to build in advance than to rebuild in public, as Klarna's experience showed.

Match autonomy to demonstrated readiness. Expand what AI is trusted to do in a given workflow at the pace verified skill and operating practice can actually support, not at the pace of ambition.

Manage toward the gap closing, not the tool launching. Upskilling works dramatically when targeted at a measured gap; make that measurable improvement the metric leadership actually reviews, not the initial deployment announcement.

Conclusion

The organisation was always the hard part. The tool was never going to be.

Eighty-eight percent of employees use AI; five percent use it transformatively. Eighty-two percent of organisations offer AI training; fifty-nine percent still report a skills gap. Thirteen percent of employees are verified as capable of working with AI agents before anyone measures and closes that gap properly. And a company that publicly staked its customer service strategy on AI alone spent more than a year rebuilding, in full view of its customers and its market, the human capability and operating practices it had assumed it no longer needed. Four different studies, and one very public case, all describing the same condition from different angles: the technology was never the constraint.

What closes the gap is treating leadership attention, accurately measured skill, and operating practice as one system that has to mature together, rather than three separate initiatives running on three separate timelines. Leadership sets the accountability and the pace; measurement replaces the seven-in-ten error rate of self-assessment with an accurate starting point; and operating practices, the escalation paths, the quality checks, the clear line between what AI owns and what a person owns, are the thing that turns a capable workforce and a working tool into an organisation that can actually be trusted to run on both.

Building an organisation that can actually use AI was never a purchase decision. It is a compounding capability, built deliberately, in the order that leadership, skill and practice can genuinely support each other, and the organisations that build it before they are forced to, rather than after, are the ones that will still be standing when the next Klarna moment happens to someone else.

References & Sources

Research and evidence informing this whitepaper.

This whitepaper draws on large-scale workforce surveys, verified skills benchmarking data and a widely documented case study. Statistics and case details are attributed to their originating source wherever directly cited.

EY. 2025 Work Reimagined Survey: Can AI Advance Toward Value if Workforce Tensions Linger?

Referenced for findings on AI usage versus transformative impact, the "Talent Advantage" gap, and missed productivity gains linked to talent-strategy shortfalls.

View source

Workera. 2026 AI Skills Enterprise Benchmark Report. May 2026.

Referenced for verified skills benchmarking across enterprise capabilities, the self-assessment accuracy gap, and the measured impact of targeted upskilling.

View source

DataCamp. 2026 State of Data & AI Literacy Report.

Referenced for findings on the gap between AI training availability and reported skills gaps, and the low prevalence of mature upskilling programmes.

View source

Workera, citing IDC research. The $5.5 Trillion Skills Gap: What IDC's New Report Reveals About AI Workforce Readiness.

Referenced for the estimated global economic cost of unresolved skills shortages and the leading barriers organisations cite to AI readiness.

View source

Forbes. Klarna Reverses AI Push, Says Customers Prefer Human Support. May 2025.

Referenced for the account of Klarna's AI customer service reversal, its CEO's public comments, and the company's subsequent hybrid operating model.

View source

Research sources were reviewed for relevance to the themes examined in this whitepaper. Survey findings reflect the populations, methodologies and reporting periods of their respective source publications and should be verified against primary sources before being relied on for workforce planning or investment decisions. This whitepaper is provided for general informational purposes and does not constitute legal advice.

Scroll