From Under 20% to 71%: What an Asset Management Firm's Failed Training Programmes Actually Taught Us

Two AI training programmes. Six hundred employees. Less than one in five completing either of them.

That was the starting point when a mid-sized asset management firm first approached me about a third attempt. They had executive sponsorship, adequate budget, access to established learning platforms, and internal communications support. On paper, the organisation looked ready. In practice, almost nobody meaningfully engaged.

By most conventional L&D measures they had done the right things. There was senior sponsorship, budget allocation, launch announcements, completion dashboards and managerial visibility. The problem was that the training itself had almost no relationship to how employees actually worked.

What followed was a complete redesign. Twelve months later completion had risen to 71%, and 58% of participants had demonstrated measurable workflow change after three months. More importantly, the internal conversation about AI shifted from anxiety and compliance into practical application.

One of the most damaging assumptions in enterprise AI adoption is that resistance comes from employees being unwilling to learn. Most professionals are perfectly willing to adopt systems that clearly improve their work, and most corporate AI training never demonstrates that improvement concretely enough.

The asset management firm had intelligent, high-performing employees operating in a highly regulated environment where precision mattered. These were analysts, operations specialists, client services professionals and portfolio teams working under significant cognitive load. They weren't anti-technology, they were anti-irrelevance.

The first programme focused heavily on platform capability demonstrations, with walkthroughs, feature explanations and templated exercises. The second shifted towards broader AI literacy delivered through the existing learning management system. Both suffered from the same structural flaw, because neither answered the question employees were actually asking, which was how any of it would help them do their specific job better tomorrow morning.

Without that connection, training stays informational. People complete modules because they're required to, and never integrate the behaviours into their workflows.

The first programme was a software rollout disguised as capability development, which is extremely common. Vendors frequently position product exposure as workforce readiness, teaching interface navigation, features, prompt examples and generic use cases on the assumption that exposure leads to adoption. It rarely does.

Software training isn't useless, but the sequencing is wrong. If employees don't first understand where AI helps in their role, where it introduces risk, where judgment remains essential, how to evaluate outputs critically and how the technology fits inside existing workflows, then tool demonstrations stay disconnected from operational reality.

The second programme tried to solve this with broader AI literacy, and generic literacy creates its own problem, because it becomes too abstract. Employees learn broad concepts about AI transformation, industry disruption and future-of-work narratives, then leave without practical integration strategies connected to their day-to-day responsibilities.

In this organisation analysts sat through the same material as operations staff, and client services teams received the same examples as middle-office functions. Scenarios were generic, exercises were synthetic, and none of it resembled the pressure, ambiguity and workflow complexity of actual asset management work.

Completion collapsed because the training generated cognitive overhead rather than practical relief, and employees subconsciously categorised it as additional work. The most successful AI adoption programmes aren't perceived internally as learning initiatives at all. They're perceived as ways of reducing friction.

The broader enterprise market is repeating this mistake at scale, mostly through one of four ineffective models.

Vendor-led demonstrations optimise for platform familiarity rather than capability, so employees leave understanding buttons rather than judgment. Generic awareness workshops create temporary excitement but weak behavioural transfer, leaving people feeling informed while remaining operationally unchanged. Compliance-first AI literacy focuses almost entirely on governance, policy and risk, which is necessary but insufficient, because employees learn what not to do without learning what productive usage looks like. Self-directed e-learning assumes employees will independently map abstract concepts onto their workflows, which high performers occasionally manage and most people don't.

The deeper issue is that enterprise AI capability is behavioural. It's workflow redesign rather than knowledge acquisition, and that requires training built around repetition, reinforcement, role specificity, accountability, operational relevance and visible application. Most programmes optimise for content delivery instead, which is why adoption collapses after the initial enthusiasm.

The third programme started from a different question. Instead of asking what employees should learn about AI, the redesign began by asking which decisions and workflows consumed the most cognitive energy across the organisation.

That shift changed everything. The redesign was built around four principles that now underpin most of the enterprise AI training work I deliver, which I refer to broadly as the CORE framework. It doesn't start with software. It starts with thinking.

The original programmes grouped employees together too broadly, which happens constantly in enterprise environments because one workshop, one deck, one facilitator and one rollout looks operationally efficient. The problem is that AI capability is highly workflow dependent, and two employees at similar seniority may need completely different integration strategies depending on how their work is structured.

In the redesigned programme, analysts focused on research synthesis, interpretation and reporting. Operations teams focused on document review and workflow triage, client services teams on communication acceleration and summarisation, and management functions on decision support and synthesis. Every exercise mapped directly to a recurring operational task.

That cut the abstraction sharply. Employees no longer had to imagine how AI might fit into their work, because they could see it, and engagement rose accordingly.

Most enterprise programmes introduce tools too early, which creates shallow capability. Before participants touched software, the redesigned programme focused on judgment. When should AI outputs be trusted? What signals indicate hallucination risk? Which tasks require human review, where does automation become dangerous, and how do you interrogate AI-generated reasoning?

This stage mattered because confidence without evaluation capability becomes organisational risk. One of the most underestimated realities of AI adoption is that poor AI users often become overconfident very quickly, so enterprise readiness depends less on technical mastery than on calibrated scepticism.

Participants needed to understand AI strengths, limitations, reliability boundaries, workflow suitability and verification requirements before any practical implementation. That sequencing improved adoption quality, because employees became more willing to experiment once they understood the boundaries.

Most corporate learning is vertically enforced. Managers monitor completion, employees complete modules, and L&D tracks participation, which produces compliance behaviour rather than capability.

The redesigned programme used cohort-based accountability instead. Small groups progressed together, discussed use cases collectively, shared experiments, surfaced failures and exchanged workflow adaptations. That reduced the perceived evaluation pressure, which matters because in highly professional environments employees often avoid experimenting publicly, since visible uncertainty carries reputational cost.

Horizontal accountability changes that dynamic. Rather than being evaluated, participants feel like they're collectively solving operational problems, and participation quality rose sharply as a result.

One of the biggest flaws in enterprise learning is reliance on self-reported confidence, which is a weak proxy for behavioural change. People frequently report feeling informed without altering their operational behaviour at all.

The redesigned programme therefore required visible workflow application. Each module concluded with a real task, a documented AI-assisted workflow, a before-and-after comparison and a practical reflection on usefulness. These examples were shared within cohorts rather than escalated upward, which stopped the experimentation from feeling performative and let employees test applications without fear of managerial scrutiny.

That moved the programme psychologically from training into collaborative optimisation. People resist education less when it feels directly tied to operational relief.

The headline figure was 71% completion, and compared with the previous sub-20% outcomes leadership understandably focused on it. The more meaningful outcome was behavioural adoption.

After three months, 58% of participants had measurably integrated AI into recurring workflows. Teams reported reduced cognitive load on repetitive synthesis tasks, document turnaround times improved in several operational areas, internal resistance to experimentation declined, and AI conversations became more practical and less abstract.

Completion measures exposure. Behavioural integration measures transfer. Most enterprise AI programmes optimise heavily for the first while barely measuring the second, which is why organisations frequently overestimate their capability maturity and mistake attendance for adoption.

Although this case comes from asset management, the same dynamics apply across consulting firms, media organisations, legal environments, government institutions, healthcare systems, enterprise operations teams and corporate strategy functions. The workflows differ, the behavioural mechanics don't.

AI adoption succeeds when employees can identify where it genuinely helps, trust the process safely, integrate the behaviours practically, reduce friction meaningfully and retain human judgment where it belongs. Most organisations still treat AI capability as a technology implementation problem when it's much closer to behavioural systems design.

Three shifts emerge clearly from this engagement.

Diagnose before designing. Don't start with platforms, vendors or workshops. Start with workflows, and map the repetitive cognitive tasks, synthesis-heavy processes, operational bottlenecks, communication friction and reporting overhead. Capability strategy should emerge from workflow analysis rather than software availability.

Stop measuring satisfaction first. High satisfaction scores frequently correlate with low behavioural transfer, because people enjoy polished workshops and that doesn't mean organisational capability changed. Measure workflow adoption, behavioural persistence, repeated usage, operational integration, time reduction and quality consistency instead.

Treat AI capability as organisational infrastructure. AI readiness compounds, so once teams begin integrating it effectively, experimentation accelerates, workflow redesign expands, peer learning increases, adoption barriers fall and organisational confidence grows. The reverse is also true, because poor early rollouts create scepticism that's difficult to unwind later, which is why the initial programme architecture matters disproportionately.

Most companies will eventually have access to the same AI tools, so the meaningful divide won't be access. It will be between organisations that operationalised AI effectively and organisations that accumulated fragmented experimentation without behavioural integration.

That gap is already emerging, and the companies seeing real gains aren't necessarily the ones spending the most. They're the ones treating AI capability development as workflow transformation rather than software exposure. In this case the breakthrough came from better behavioural design, not better technology.