Reconciling AI and Human Intelligence

In small and medium-sized enterprises, enthusiasm is evident: more than eight out of ten executives believe that artificial intelligence will revolutionize their business as early as next year, but fewer than four out of ten feel truly capable of deploying it across their organization. The gap is even more striking when looking at the maturity of its use: 92% of teams have already tested generative tools, yet only 19% dare to automate agentic tasks; the former capture 83% of the perceived value, the latter barely 1%.

This paradox reflects a broader trend: in advanced economies, a handful of companies (around 2%) capture two-thirds of the productivity gains. In other words, access to algorithms is not enough; it is the combination of trust, culture, skills, and disciplined execution that separates the pioneers from the pack.

The data suggests a four-step progression: first, establish mutuality; then, evolve the culture; develop human capabilities; and finally, strategically integrate AI. In this sequential approach, the order is crucial: skipping a step often condemns the project to failure or, worse, triggers adverse effects (bias, rejection, loss of trust).

Mutuality, understood as a sharing of power and meaning, requires making models explainable, involving stakeholders very early on, and providing clear avenues for appeal. A large-scale eye screening platform was able to be adopted in regions with a shortage of specialists because ophthalmologists participated in its design, validated the metrics, and retained the possibility of a clinical veto. Conversely, a well-known system for assisting in the selection of cancer treatments was rejected by physicians: trained on theoretical cases, it sometimes proposed non-compliant protocols and remained opaque regarding its reasoning.

Once this foundational trust is established, the culture becomes fertile ground for widespread adoption. Organizations that value continuous learning, the right to make mistakes, and cross-functional collaboration move more quickly from prototype to production. One American hospital distributed AI-powered ultrasound scanners to all its residents; management held numerous experience-sharing sessions, rewarded early successes, and openly discussed technical limitations. The result: a measurable increase in diagnostic accuracy and internal enthusiasm that fuels other initiatives.

Next comes the question of skills. The required competencies are not limited to "prompt engineering." It's necessary to be able to interpret the uncertainty of a model, orchestrate multiple agents, detect data drift, but also exercise ethical judgment, link algorithmic outputs to economic levers, and, above all, hold two opposing ideas together without succumbing to a binary reflex. The figures speak for themselves: more than half of the trained employees still report dreading implementation, indicating that training must be accompanied by mentoring and practical exercises.

Once these building blocks are in place, integration can begin. The most successful projects often follow a "think big, start small, scale up fast" approach. They link each use case to a specific lever: reducing supplier costs, improving cycle time, or increasing equipment efficiency. They include a runbook: model refresh rates, bias audits, escalation procedures, and dashboards that display both operational metrics (resolution rate, OEE, unit cost) and confidence indicators (frequency of overrides, consultation of explanations, talent stability).

This discipline is not a "soft factor" disconnected from profitability: companies that simultaneously invest in skills retention and AI governance double on average their chances of outperforming in productivity and shareholder value creation.

Beyond the initial production deployment, the loop must remain open. The authors emphasize the usefulness of a Plan-Do-Check-Act cycle: plan around a common objective, experiment in a secure environment, measure the technical and human impact, and adjust or expand based on the lessons learned. A major cloud computing company is already applying this principle to the energy management of its data centers: an initial "recommendation" mode was used to build trust; the algorithm now controls the air conditioning in real time, but under the supervision of an operator who can take over at any time—savings have reached 30% and continue to improve.

The macroeconomic stakes justify this rigor: if we do not resolve the alignment between humans and machines, nearly a third of the potential value of AI — more than $4.6 trillion by 2030 — risks remaining out of reach.

For a decision-maker, the emerging roadmap is clear: listen to stakeholders before buying any license, anchor the experiment in a collective narrative, measure psychological safety as seriously as ROI, map skills gaps by role and by business ritual, finance pilots with exemplary governance and then instrument each flow to demonstrate the correlation between AI and productivity.

Ultimately, the most advanced technology is only valuable if it enhances human intelligence. This intelligence is expressed in the creativity of teams, the quality of decisions, and the trust that customers—and employees—place in the company. Where these conditions are met, AI ceases to be a mere gadget and becomes a multiplier of meaning and performance. Where they are lacking, it is simply another line item in the IT budget.

It is entirely up to us to choose the scenario. The tools are available, the benefits proven; the difference will lie in our ability to orchestrate trust, culture, skills, and execution. In other words, to give as much attention to the intelligence of men and women as to that of machines.

To learn more: discover our analysis on orchestrating humans and AI and our article on amplify one's abilities without surrendering one's free will.