Prosus Report: 60,000 AI Agents and the Coming Age of AI Colleagues

The Coming Age of AI Colleagues: Inside Prosus’ 60,000-Agent Experiment

Europe's quiet tech giant didn't just theorize about agentic AI-they built an army of them. In a landmark May 2026 report, Prosus reveals how deploying over 60,000 AI agents across its global portfolio is fundamentally rewriting the rules of human-machine collaboration.

Key Takeaways

• The 2% Power Law: A tiny fraction of AI agents drives the vast majority of business impact, signaling that leaders should hunt for extreme outliers rather than forcing shallow, company-wide adoption.
• Organic convergence: Without any top-down corporate mandate, disparate global teams naturally converged on 20 "default" agent use cases, from churn tracking to complex contract review.
• AI seniority tiers: Agents behave remarkably like human staff, scaling from single-tool "interns" to multi-tool "seniors." Senior agents manage complex workflows and see the highest user volume across the business.
• The "Delete it tonight" test: ROI is best measured pragmatically by asking a business owner a simple question: What crashes if this agent is switched off overnight?
• Autonomous organizations are next: We are currently layering AI onto legacy corporate structures. The approaching wave involves rebuilding entire departments from scratch around agentic workflows.

---

When we talk about enterprise AI, the discourse is usually dominated by pilot purgatory or vendor hype. Prosus offers a much-needed reality check.

As the fifth-largest publicly listed tech company in Europe, Prosus commands a massive global footprint. Its portfolio includes outright ownership of operations like PayU and Despegar, majority stakes in household delivery names like iFood and Just Eat Takeaway, and formidable minority holdings in giants like Tencent, Swiggy, and Meesho.

To understand how AI actually functions at scale, Prosus handed the keys to its 40,000 employees. Using the company’s internal agent building platform, Toqan, these workers actively developed over 60,000 custom AI agents. What emerged is what Prosus rightfully claims to be the largest independent dataset on agentic AI currently in production.

The resulting insights cut through the noise, offering a definitive blueprint for how businesses will operate for the rest of the decade.

The 2% Power Law and the 20 "Default" Use Cases

If you assume AI value is distributed evenly across an enterprise, you are already losing money. Prosus discovered a brutal, highly concentrated "power law" at play: roughly 2% of the active agents built on Toqan drove a wildly disproportionate share of the business impact.

The lesson here is simple. Once a business finds its 2%, the first operational priority is to ruthlessly double down on those agents.

Fascinatingly, this extreme concentration happened organically. Prosus explicitly avoided a top-down, centralized mandate telling companies what to build. Yet, across borders and drastically different business models, the portfolio companies all converged on the exact same 20 "power law" use cases.

Within a few years, these 20 use cases are highly likely to become the "default AI" footprint in virtually every modern enterprise. They aren't parlor tricks; they are core business functions. Notable examples include:

• Triaging inbound customer messages and reviews.
• Generating hyper-customized internal data reports.
• Tracking customer churn-risk in real-time.
• Generating highly specific, multi-channel marketing content.
• Compiling deep sales prospect briefings.
• Detecting anomalies and financial fraud.
• Reviewing high-volume invoices and legal contracts.

Meet the AI Workforce: From Interns to Seniors

Across those default use cases, Prosus identified 54 distinct AI tasks. When we look at where these tasks live, the data points to clear departmental winners. Data analytics and market intelligence led the pack, accounting for 18% of all tasks. Operations followed at 15%. Interestingly, "personal AI assistants"-which belong to no formal department at all-made up 14% of the ecosystem.

But all AI agents are not created equal. Prosus categorized this digital workforce into four distinct "seniority" tiers that perfectly mirror their human colleagues.

1. The Intern
Intern agents are single-purpose. They have access to 0-3 tools and handle rudimentary, highly constrained tasks. They need supervision and clear prompting, but they grind through busywork flawlessly.

2. The Junior
Stepping up the ladder, junior agents have slightly more autonomy and can chain together a few simple inferences, handling routine data pulls or basic customer triage.

3. The Mid-level
Armed with 4-10 integrated tools, mid-level agents can access live databases, cross-reference company wikis, and draft contextual responses. They begin to handle multi-step workflows with minimal human oversight.

4. The Senior
These are the heavy hitters. Senior agents possess 11 or more tools. They manage highly complex, multi-layered workflows, adapt to ambiguous prompts, and make sophisticated judgment calls.

According to the data, daily usage across the network splits roughly 50/50 between senior and junior agents. However, senior agents boast the highest number of unique human users. When an AI can genuinely do heavy intellectual lifting, word travels fast, and adoption skyrockets.

Pragmatic ROI: The "Delete It Tonight" Test

Calculating the ROI of generative AI has notoriously tied up CFOs in knots. Prosus advocates for a radically simplified approach: focus on empowering teams, avoid rigid top-down budgeting, because costs at this early stage are nearly impossible to accurately forecast.

To measure real value, Prosus divides its top-performing agents into two tracks: Productivity and Value.

Productivity Agents (Measured in time saved)
These agents give employees their lives back. Prosus found that 82% of productivity AI (mostly personal assistants) save smaller increments of time-less than 20 hours a month per user. However, 17% save between 20 and 173 hours a month, effectively replacing one full-time equivalent (FTE) workload. A rare elite tier, representing less than 1% of agents, quietly does the work of thousands of human hours. Across the portfolio, these productivity agents represent well over 1,000 FTE equivalents of labor repurposed to higher-value thinking.

Value Agents (Measured in new revenue)
These agents directly generate hard currency. While the majority pull in modest wins of under $1 million annually, the mid-tier agents consistently generate between $1 million and $10 million natively. And then there are the outliers. At one portfolio company, a completely AI-driven third-party affiliate marketplace was built from scratch and is now projected to generate an astonishing $83 million per year.

But how do you audit these figures without hiring an army of consultants? Prosus relies on a brutally effective internal metric: The "Delete it tonight" test.

You approach the business owner of an AI agent and ask what would happen if the agent completely disappeared at midnight. If the answer is "we would be annoyed," the agent is a luxury. If the answer is "workflows crash, SLAs are breached, and we lose money tomorrow," you have found your true, pragmatic ROI.

The 3-Phase Playbook for Agentic Adoption

Building an internal platform and expecting workers to flock to it is a recipe for failure. Driving the creation of 60,000 agents required a calculated orchestration of corporate psychology. Prosus outlines a three-phase playbook for getting an enterprise onto the AI grid.

Phase 1: Initiative
This is the foundational stage. It requires an aggressive executive sponsor, firm OKRs, and dedicated engineering roles. Critically, internal AI pricing must be free for the end-user. If departments have to pay out of their own budgets to experiment during this phase, innovation dies instantly.

Phase 2: Initial Adopters
Here, you hunt for your early power users. You stimulate the culture via hackathons, prompt engineering training, and viral how-to videos. A massive unlock here is "co-creation AI"-agents that actually help employees build other agents. At iFood, an internal tool called Optimiza AI walked 1,300 non-technical employees through the creation process, resulting in over 10,000 active agents being built almost overnight.

Phase 3: Scale
Once the spark catches, it is time for structural scaling. This requires a regular executive cadence to review agent performance. Gamification is highly effective at this stage; Prosus hosted massive internal contests, rewarding the creators of the highest-impact agents with significant prizes, including flying the winning teams out to China. Finally, you establish an internal "App Store" marketplace where top agents can be downloaded and adapted by other teams.

The Next Wave: Autonomous AI Organizations

The overriding conclusion of the Prosus report is a warning against complacency. What the company achieved with Toqan is entirely categorized by the authors as "the current wave."

Right now, we are layering AI on top of existing corporate structures. To borrow an analogy from the industrial revolution, we are tearing out the steam engines and replacing them with electric motors, but we are leaving the layout of the factory floor exactly the same.

The next wave is the leap to autonomous, AI-enabled organizations. This is the era where entire departments-marketing, customer service, internal auditing-are torn down to the studs and rebuilt from scratch around intelligent, agentic workflows. Humans move strictly to the edges to manage the system, while the core loop of the business runs entirely on silicon. Prosus is already actively testing this operational model inside its portfolio.

What this means for European business leaders

For European executives parsing this data, the tactical takeaways are sharp and unforgiving:

1. Stop dictating AI use from the boardroom. Provide secure sandboxes and empower employees to build their own solutions. The convergence on the highest-value use cases will happen organically.
2. Embrace the 2% rule. Don't try to make every agent a winner. Let a thousand flowers bloom, execute the "delete it tonight" test, aggressively prune the frivolous 98%, and scale the outliers globally.
3. Budget for enablement, not precise software ROI. In 2026, obsessing over exact AI budget forecasts is a fool's errand. Make the compute free internally in Phase 1 and Phase 2 to ensure widespread behavioral change.
4. Restructure your HR mindset. You now manage a hybrid workforce of humans and code. Treat your agents like employees: track their seniority, give them distinct toolsets, and measure their performance in FTE equivalents or hard revenue.
5. Prepare for the autonomous shift. Layering AI over broken legacy processes is a temporary fix. Start war-gaming what your core business units would look like if they were redesigned assuming AI, rather than humans, did the majority of the processing.

For a deeper dive into the methodology and data, read the original report here: The Coming Age of AI Colleagues.

// About the author

Remy Gieling

Mede-oprichter, AI-expert & bestseller-auteur

Tech-expert (1988) gespecialiseerd in kunstmatige intelligentie en mede-oprichter van ai.nl, The Automation Group, Proxies en eBrain.ai. Oud-hoofdredacteur van diverse zakenmerken en daardoor een geoefend verteller op het podium en in de media. Verzorgt jaarlijks 150+ AI-keynotes in binnen- en buitenland en is gastdocent aan Nyenrode. Co-auteur van zeven boeken, waaronder 'Handboek AI Strategie' en 'AI Agents', en bekend als presentator op radio en RTL Z. Reist langs de labs van OpenAI, Nvidia en Tencent en vertaalt de nieuwste doorbraken naar inzichten die leiders direct kunnen toepassen.

LinkedIn