Agentic AI Transformation?

This article is part of our “How we work” series. You can also read this article in Spanish and Italian.
If you would like to receive updates on this and other topics around the changing landscape of the world of work, subscribe to our newsletter here: https://cocoon-pro.com/#subscribe
If you would like to learn more, or bring this approach to your organisation, please contact us here: https://cocoon-pro.com/get-in-touch/

You are about to authorize significant investments in agentic AI. But who, exactly, should lead this transformation in your organization?

Research shows a recurring pattern: the same technologies can produce opposite results depending on how they are integrated into teams. This article summarizes what actually works when AI agents enter decision-making processes, and proposes an AI transformation framework to avoid the most costly mistakes.

What happens when AI stops being just a tool?

In recent years, generative artificial intelligence has definitively left the laboratory and entered daily workflows. Now we’re taking a further step: the arrival of AI agents, systems capable not only of generating content, but of observing, deciding and acting with relative autonomy within business processes, orchestrating tasks, speaking with customers, taking operational initiatives.

This is what many analyses call agentic AI: an extension of generative AI that behaves less and less like a tool and more and more like a digital colleague. It is not science fiction, it is already happening: global estimates show that adoption of generative and agentic AI is growing rapidly across all sectors, with measurable impacts on productivity, work structure and talent strategies (Stanford HAI, 2024; Microsoft, 2025).

In our dedicated report we observed how this epochal shift is now at the center of organizational decision-making. It appears on the agendas of boards of directors, public leaders, fund boards and institutions. And it is already reshaping our very idea of work.

At this point, the real question at the center of many important decisions is no longer whether to adopt agentic AI, buthow to do it in a way that truly improves work, instead of creating new constraints and fragilities.

What we already know—beyond the hype

In 2024, at Cocoon Pro, we began exploring in the field what happens when AI agents enter teams. In 2025 we launched an action-research project with the University of Verona, MIK in Bilbao, the Belbin research center in the United Kingdom and some pioneering organizations like CDP. Real experiments: teams working on real challenges, with AI agents integrated into their workflows and decision-making moments.

In parallel, international research is accumulating increasingly robust evidence. Some of the key findings include:

In controlled experiments, access to generative AI tools reduces the time to complete professional writing tasks by about 40% and increases the perceived quality of the result by approximately 18% (Noy & Zhang, 2023)
In hybrid-team experiments, individuals supported by generative AI achieve performance comparable to that of small human teams, while human + AI teams are three times more likely to produce solutions in the top decile of quality (Dell’Acqua et al., 2025)

Recent reviews confirm that the effects of AI on performance are almost always heterogeneous: they benefit especially the less experienced, but depend critically on how human–AI collaboration is organized (OECD, 2025; Schmutz et al., 2024).

The synthesis is simple—and uncomfortable: the technology is mature enough to make a difference, but positive outcomes are far from guaranteed. And this is where organizational choices come in.

From both field experience and the academic literature, three key lessons emerge.

1. AI should not be managed as an IT issue

Many organizations experience the introduction of AI as a natural task for IT: architectures, models, vendors, and security. All necessary, but not sufficient.

AI agents do not simply change tools. They change:

who does what,
with whom,
and with what degree of autonomy.

They shift operational decisions from humans to agents. They redesign boundaries between tasks, roles, teams, and functions. They reopen questions about responsibility, trust, control, and incentives.

Research on human–AI teams shows that performance depends on elements like trust, transparency of AI decisions, “voting” methods between humans and agents, management of judgment conflicts, more than on the raw accuracy of the algorithm (Hu et al., 2024; Carter et al., 2024; Afroogh et al., 2024).

In a recent study, for example, researchers observed what happens when human members of a mixed team are systematically “outvoted” by AI. The result? There’s no revolt, but subtle changes emerge in trust, willingness to contribute, and how people perceive themselves within the team (Hu et al., 2024).

It is not a system bug: it is a socio-technical effect.

For decision-makers in today’s organizations, the conclusion is straightforward: if agentic AI is managed as only an IT issue, the risk is not technical—it is strategic. You risk introducing a new “digital colleague” into teams without rethinking the social system that must accommodate it.

2. Adoption matters more than technology

Studies now agree on one point: comparable AI tools, in different contexts, produce very different results.

In some cases, we see net increases in productivity, quality, and satisfaction. In others, no measurable effect. In still others, deterioration due to over-reliance on AI, loss of coordination, or reduction in human learning (OECD, 2025; Li, 2024).

The same technology, therefore, can become either a lever or an obstacle, depending on how it is adopted. What makes the difference?

The concrete design of workflows involving the agent.
The explicit role assigned to AI within the team (advisor, co-author, autonomous executor, etc.)
How people are supported: training, experimentation spaces, and shared rules of engagement.
The metrics used to evaluate the new setup not only “hours saved” but also quality, equity, learning, and wellbeing).

A growing body of work insists on a point that’s often missing from slides: AI adoption is a portfolio of organizational experiments, not a single top-down decision (Berndt et al., 2026). Companies that see lasting impacts are those that design small high-quality experiments, measure well, learn quickly and scale only what works.IT and HR must literally sit at the same table: one brings infrastructure and constraints, the other brings knowledge of real work, skills, team dynamics.

3. Efficiency cannot be the main driver

Almost all business cases on AI start from an efficiency argument: reducing time, automating repetitive activities, cutting costs. This is understandable—and dangerous if it remains the only lens.

The most recent literature shows that AI not only increases output and quality, but also changes the composition of work: it shifts time from repetitive standard activities toward coordination, judgment, relationships, problem solving (Brynjolfsson et al., 2025; Noy & Zhang, 2023; Dell’Acqua et al., 2025).

If this shift is not thought through, something happens that you’ve probably already glimpsed:

People experience AI as a threat or a source of additional control, rather than as an ally.
Freed-up time is filled with low-value activities—additional reporting, control meetings, micromanagement—instead of learning, innovation, or care for vital relationships.

In parallel, important warning signs also emerge about organizational wellbeing: some studies link poorly managed AI use to greater isolation, loss of collaboration and impact on psychological wellbeing (Financial Times, 2025; Stanford HAI, 2024).

The real question, therefore, is not “how much time do we free up with AI”, but: what do we want to free that time for?

If this question remains implicit, it will be the existing system that answers for us.

What do we need to know to lead our organizations in this new era?

Read our report dedicated to this important transition by clicking here.

An old mining story, that speaks to AI

This tension between technical efficiency and team life is not new.

In the 1950s, Eric Trist and colleagues at the Tavistock Institute studied the introduction of the longwall technology in British coal mines. From a technical standpoint, it was a success: more linear flows, higher output, and reduced cycle times.

The problem? In order to fully exploit the new technology, work had been redesigned in a way that broke up the cohesive units that characterized miner teams. Groups lost autonomy, self-regulation capacity, sense of shared responsibility at the “mining face”. The counterintuitive result was an overall decline in performance, accompanied by increased conflict and absenteeism (Trist et al., 1963).

From that research the socio-technical approach was born, which then grew globally for 50 years: there is no single optimal organization for a given technology. There are organizational choices that can turn the same technology into an enabling or a destructive force.

This is precisely the kind of choice organizations now face with the introduction of agentic AI. We can design teams where AI agents are silent protagonists that rigidify already obsolete processes; or augmented teams where humans and agents empower each other, with more autonomy and more adaptive capacity.

This is why an explicitly socio-technical transformation framework is needed, one that holds together:

operational evolution (processes, roles, metrics);
behavioral evolution (trust, skills, team dynamics);
epistemological evolution (how the organization knows, learns, decides).

A transformation model in three pillars: Ecology × Literacy × Technology

Our proposal, born from the interweaving of practice and research, is to read the socio-technical success of agentic AI transformation through three interdependent pillars:

AI Gain = Ecology × Literacy × Technology

The formula is intentionally multiplicative: if one of the three factors tends toward zero, so will the overall gain. We can have the best agentic AI platform, but if the ecology is toxic or literacy is poor, the net effect will be limited or even negative.

1. Ecology: the vital relationships that hold the system together

By Ecology we mean the system of vital relationships that this change involves. Within this field we find at least three dimensions:

Leadership, understood as a relational process, not as boxes on the org chart. How is this relational space managed? How, and how much do connections between people orient the common direction and trigger movement and reciprocal learning?
Governance, that is the set of rules, guiding principles and safety constraints that translate organizational identity into concrete decisions embedded in systems. Which rules are embedded in agents? Which are deliberately kept outside?
Ecosystem: internal people, key partners, customers, suppliers, regulators. Agentic AI doesn’t stop at the formal boundaries of the company; it enters supply chains, networks, platforms.

The ecology of such a profound transformation focuses attention on its dynamic, flow aspects, rather than on positional or structural aspects of org charts, or technological elements. The point must be the organization’s vital relationships, and the dynamics that amplify them, put them in resonance, and make them generative. How do we shift human potential toward these aspects, and ensure that AI frees up room for maneuver for people and opens new possibilities?

For decision-makers in our organizations, the key questions about ecology are of the type:

Which vital relationships of our business are touched first by AI agents?
Which non-negotiable principles do we intend to explicitly incorporate into automated decisions?
How do we plant to distribute decision-making power between humans and agents at critical moments?

2. Literacy: the ability to participate in change without suffering it

Literacy is the ability to access, evaluate and create use cases in different contexts. It is not just “knowing how to write prompts” or use a couple of tools.

It includes at least:

understanding what an agent can do, what it cannot do, and where it is dangerous for it to decide alone;
knowing how to break down a process into delegable steps, maintaining control over critical passages;
knowing how to measure, at least at the team level, whether a new use of AI is truly improving work and results.

Studies on AI’s effects on work confirm that without an explicit investment in skills the risk is creating a few “islands of excellence” and a large majority that suffers the change or remains on the margins (Stanford HAI, 2024; Microsoft, 2025; OECD, 2025).

At the same time, research on training and GenAI shows that people with less experience can rapidly close performance gaps when supported by well-designed AI (Noy & Zhang, 2023; Cui et al., 2025). The point is that this doesn’t happen “by itself”: it requires instructional design, tutoring, feedback.

For HR and people leaders, the key questions are very concrete:

what is the real level of AI literacy of our key figures today, beyond declarations?
how are we creating safe experimentation spaces, where teams can try, fail, learn?
who has license to propose new use cases and decide what gets scaled?

3. Technology: the right application suite, not the “toy of the moment”

Finally, Technology concerns the choice of the application suite that best fits

the technologies and data already present in the organization;
the evolutionary needs of the ecology (leadership, governance, vital relationships);
the development trajectory of culture and skills (literacy).

In this scope, several practical themes come into play:

agentic AI platforms and workflow orchestration;
integration with legacy systems and real company data;
security, privacy, compliance and decision traceability;
tools to observe, audit, and correct agent behavior over time.

Recent analyses converge on two points:

the near future is not a single “super-agent”, but portfolios of specialized agents that cooperate with one another and with human teams (IBM Institute for Business Value, 2025; PwC, 2024);
technological choice is not neutral: it determines who can experiment, how costly it is to change course, and how easy it is to deactivate an agent that is not producing the desired effects.

For CIOs and COOs, some critical questions include:

What feedback loops exist between those affected by agents’ decisions (customers, frontline staff) and those who design them?
How easy is it for a business team to compose or adapt an agent without having to open a formal IT project every time?
What visibility do we have into what agents are actually doing in production?

So, Board of Directors, who should lead this transformation?

If we put these three pillars together, the picture becomes clear: agentic AI is not an IT project, it is a system-wide socio-technical transformation.

IT brings indispensable expertise on infrastructure, data, security, vendors. But who has the responsibility to:

Read the implications on work, roles, career paths?
Rethink policies, performance systems and incentives in teams where part of the work is done by agents?
Care for wellbeing, trust and sense of equity in a context of reallocation of tasks and decision-making power?

This is the middle ground of the HR function. So, a clear-sighted move today could be this:

1. Build a mixed steering committee (HR + IT + business + risk), with explicit mandate not only on tools, but on the design of augmented teams and the rules of the game.

2. Invest in disciplined experimentation, with few well-designed experiments, clear metrics and rapid learning cycles, rather than massive “big bang” rollouts (Berndt et al., 2026).

3. Anchor each AI project to simple but demanding questions:

what part of our ecology are we touching?
what literacy leaps does it require of the people involved?
how does this technology make work more human, not less?

An invitation for you

Evidence from international research and from work with real teams point in the same direction: when well designed, augmented teams can generate substantial gains in productivity, quality and innovation capacity (Brynjolfsson et al., 2025; Dell’Acqua et al., 2025; Stanford HAI, 2024).

However, there’s no “secret recipe” to copy. There’s a path to build, step by step, in the specific conditions of each organization.

For this reason, we invite you:

to participate in the ongoing research, contributing with your teams to strengthen the evidence base on how human–AI teams actually work;
or to contextualize these insights within your own organizational reality, exploring where it makes sense to begin, with which experiments, and under what safety conditions.

If you are part of the strategy-making leadership of your organization, or in an HR or IT function, or lead a key team, you are in a rare window of opportunity: to decide now what kind of relationship you want your people to have with their future artificial colleagues.

Get in touch

This is not simply about adopting a new technology. It is about rewriting the rules of work anew, so that humans and AI agents can shape the next generation of organizations together.

Essential References

Angus, D. C. (2025). AI, Health, and Health Care Today and Tomorrow. JAMA, 334(12), 1151–1160.

Berndt, J., Englmaier, F., Sadun, R., Tamayo, J., & von Hesler, N. (2026). A Systematic Approach to Experimenting with Gen AI. Harvard Business Review, 104(1–2), 76–87.

Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942.

Carter, W. R., et al. (2024). Toward a theory of AI–human team effectiveness. European Management Review. Advance online publication.

Cui, Z. K., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Software Development. Working paper.

Dell’Acqua, F., Ayoubi, C., Lifshitz-Assaf, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., et al. (2025). The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. NBER Working Paper No. 33641.

Hu, M., Zhang, G., Chong, L., Cagan, J., & Goucher-Lambert, K. (2024). How Being Outvoted by AI Teammates Impacts Human–AI Collaboration. International Journal of Human–Computer Interaction, 41(7), 4049–4066.

IBM Institute for Business Value. (2025). Orchestrating Agentic AI for Intelligent Business Operations.

Li, N. (2024). Generative AI Enhances Team Performance and Reduces Coordination Costs. arXiv preprint arXiv:2405.17924.

Microsoft. (2025). The 2025 Annual Work Trend Index: The Frontier Firm is Born.

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 1121–1124.

OECD. (2025). The Effects of Generative AI on Productivity, Innovation and Entrepreneurship.

PwC Middle East. (2024). Agentic AI – The New Frontier in GenAI: An Executive Playbook.

Schmutz, J. B., et al. (2024). AI-teaming: Redefining collaboration in the digital era. Journal of Safety Science and Resilience.

Stanford Institute for Human-Centered AI (HAI). (2024). AI Index Report 2024: Measuring Trends in AI.Trist, E. L., Higgin, G. W., Murray, H., & Pollock, A. B. (1963). Organizational Choice: Capabilities of Groups at the Coal Face under Changing Technologies. London: Tavistock.