AI is a people problem
before it's a technology problem

In January 2026, Roland Berger published "Making AI a people topic — Turning technology into everyday practice", part of its Crafting Tomorrow series.¹ Its finding cuts against the way most enterprises run their AI programmes. Asked what actually blocks AI, the surveyed leaders point not to compute, data pipelines or model quality, but to missing skills and capabilities and insufficient training as the leading roadblocks.¹ The hard part of enterprise AI, in other words, is people.

The temptation is to treat that as a soft, HR-adjacent footnote to the real engineering work. The evidence says it is the work. Roland Berger reports that companies which build people topics into a clear AI strategy are about 2.9× more likely to outperform their peers.¹ Same models, same vendors; the difference is whether the organization made AI something its people actually do. Roland Berger puts the shift plainly:

Read that twice, because it reorders the usual priority list. A licence is easy to buy and easy to celebrate. The capability to use it well is slow, diffuse and cultural — so it quietly never gets built, and the tools sit idle. Most organizations, the report finds, are still stuck in the earliest maturity stages — exploration and initial adoption — not because the technology isn't ready, but because the practice isn't.¹ The rest of this piece is about what closes that gap, what the evidence says works, and what it takes to operationalize it without the usual security and governance traps.

The diagnosisThe barriers are people-shaped, not model-shaped

The headline finding holds up across the wider market: every independent survey that asks what is actually blocking enterprise AI keeps returning people answers — skills, training, culture — not technical ones:

Notice what every one of these is: a property of the workforce and the culture around the tools, not the model weights. The corroborating evidence points the same way. McKinsey reports that the overwhelming majority of organizations now use AI in at least one function, yet only about 1% describe their rollout as "mature" — adoption has outrun capability almost everywhere.³ And when AI does fail in production, the causes are organizational: Gartner predicts over 40% of agentic-AI projects will be cancelled by the end of 2027, citing "escalating costs, unclear business value or inadequate risk controls"² — none of them a model deficiency. The technology is ready before the organization is.

The leverBuild a safe-to-fail sandbox, and make it visible

If the barrier is capability and culture, the highest-leverage fix is not another tool licence — it is giving people a low-stakes way to practise. Roland Berger's illustrative case is an industrial firm that built exactly that: a deliberate "sandbox mindset" with dedicated "AI hours", peer showcases that surfaced both successes and instructive failures, and leadership that recognized creative use regardless of whether the outcome was perfect — producing, by the report's account, a high volume of small experiments and several that scaled into real impact.¹ The report's prescription is explicit:

The mechanism is not magic — it is the well-documented effect of psychological safety. In a study of more than 2,000 managers, MIT Sloan Management Review and BCG found that 83% of executives say a culture prioritizing psychological safety measurably improves the success of AI initiatives.⁷ People try, share and learn far more readily when failure is framed as data, not as a black mark — and that is what turns idle AI licences into everyday practice.

But notice what the mechanism actually requires to work at scale. "Make efforts visible" and "celebrate both successes and learnings" are not slogans — they are a data problem. Someone has to capture every experiment, sort the signal, and put successes and honest failures side by side in front of peers, every week, without it becoming a job nobody wants. Run by hand, the showcase quietly dies; the experiments keep happening but stop being visible, and the culture's flywheel stalls. This is precisely the kind of repetitive, judgement-light, high-cadence work a governed flow is built to carry.

The enablersThe conditions that turn AI into everyday practice

The report generalizes its case studies into a set of enablers for making AI a people topic. They are organizational, not technical — and each one has a clear operational footprint:

• Safe-to-fail zones. Designated sandboxes where people can experiment without consequence — the precondition for everything else.
• Open peer sharing. Showcases that make successes and failures visible, so learning compounds across teams instead of staying siloed.
• Recognition, decoupled from perfect outcomes. Leadership rewards the effort and the learning, not just the wins — which is what keeps people experimenting.
• Tiered fluency and a champion network. Literacy, applied and strategic tracks, plus visible champions who pull others along.

The thread running through all of them is the same: AI scales when it becomes a practice that people see, share and are recognized for — and practice only stays alive if the seeing and sharing keep happening on a cadence. Recognition, crucially, stays a human decision; the job to automate is the capture and the visibility, never the judgement of who deserves credit.

The synthesisAutomate the visibility, never the judgement

Here is the subtle part, and it is where most "let AI run the culture programme" ideas go wrong. The moment you ask a model to read people's experiment write-ups, you have invited two failures. The first is security: those submissions are untrusted text, and a crafted entry can try to steer the model — OWASP ranks prompt injection as the #1 risk for LLM applications, including content that need not be human-visible as long as a model parses it.⁵ The second is more human: the instant a model starts judging whose experiment "succeeded", you have rebuilt the fear of failure the sandbox existed to remove.

The fix is a deliberate division of labour. The flow captures, classifies and makes visible; it never adjudicates and never acts on a person. The outcome shown in the showcase is the submitter's own self-reported label, not a verdict a model talked itself into — so there is no high-stakes judgement to game in the first place. A handful of non-negotiables make it real:

And it runs self-hosted — on-premise, private cloud or air-gapped — so the experiment submissions, names and team data never cross a boundary you don't own. Cisco's 2025 benchmark found 90% of organizations believe local storage of data is inherently safer;⁶ for data about your own people, where it runs is not a preference.

flow8 in practiceA sandbox that stays visible, by itself

We built the report's #1 lever as a concrete flow8 flow — an AI-experiment sandbox tracker that keeps the showcase alive without a coordinator. People submit a micro-experiment by web form or email; the flow scans it, classifies the use case, dedupes it, and writes it to a shared experiment bus — an experiments ledger — that rolls up every week into a single peer-showcase digest of successes and fail-with-learning, plus per-team uptake. The architecture, not the prose:

The takeawayMake AI something your people do

Roland Berger is right that the binding constraint on enterprise AI is no longer the technology — it is whether the organization turned it into everyday practice. The firms that pull ahead will be the ones that gave people a safe-to-fail place to experiment, made successes and learnings visible through peer sharing, and recognized the effort — while keeping the security and governance discipline that lets all of that run on real company data. Get the people topic right and the tools you already pay for finally get used. Skip it, and you will keep buying capability your organization never quite learns to wield.