What 60,000 AI agents teach us about returns

Anyone who has spoken with executives recently about AI agents mostly heard promises. Hard figures on what agents deliver in a standard organisation are scarce. Prosus, the Amsterdam-based tech company behind brands such as Just Eat Takeaway and iFood, has now changed that. Across its entire portfolio, the group's 40,000 employees built over 60,000 agents in eighteen months, and in the report The Coming Age of AI Colleagues, Prosus shares what came of it.

The returns are concentrated in a small top tier

The core of the report is a classic power law. Approximately 2 per cent of the most active agents deliver a disproportionately large share of the business impact. The variance is significant. Of the agents that save time, 82 per cent yield fewer than 20 hours per month—mostly personal assistants. 17 per cent save between 20 and 173 hours per month, or up to just over one FTE. Fewer than 1 per cent do the work of thousands of hours per month.

For agents that affect revenue or costs, the picture looks the same. Most generate less than a million dollars a year; a small group generates much more. One subsidiary ran a new affiliate marketplace on agents for onboarding and communication. This marketplace is expected to generate 83 million dollars in revenue per year.

Five things the top performers have in common

1. They solve a clear, recurring problem. The most striking insight: without being instructed by head office, companies across different sectors, countries, and languages consistently built the same twenty use cases, each providing an immediate return. Examples include reviewing and triaging large volumes of incoming messages, generating bespoke reports from corporate data, and tracking customers at risk of churning.

2. They are shared by teams. Personal assistants are useful, but the biggest wins come from agents used by many colleagues. Prosus therefore deliberately spends no time calculating the value of every personal assistant and only quantifies the returns of shared agents.

3. They are linked to internal systems. According to Prosus, agents without a connection to internal systems yield little. Added to this is a less visible point: in many departments, everyone creates a slightly different version of the same product, so a team must first agree on what the standard output looks like before automation makes sense.

4. They make work profitable that was previously too expensive. A delivery platform has thousands of small restaurants for which a dedicated account manager is not financially viable. An agent supporting this long tail resulted in 119 per cent more orders and 73 per cent better retention. The highest yield is therefore not always found in accelerating existing work. Sometimes an agent opens up a market that was previously too expensive to serve.

5. They are measured and then scaled up. Prosus uses a surprisingly simple test. If the return is difficult to determine, the company asks the business unit leader what would happen to revenue or costs if the agent were permanently deleted overnight. Such a business-driven estimate proves more reliable than a calculation from IT or strategy. What works is then made widely available: the best agents are placed as adaptable templates in an internal marketplace that every employee can search.

The model matters less than you think

Anyone expecting the newest model to make all the difference will be disappointed. According to Prosus, most models are now good enough for almost all agent tasks, and the latest models are only needed for the most complex assignments. The primary risk lies in costs. Users are reluctant to switch models once their agent is working, meaning expensive models sometimes execute tasks that a cheaper model could handle just as well.

How Prosus kickstarted adoption

The second part of the report is a practical manual for adoption, and perhaps the most valuable section. A few lessons stand out.

• Commitment from the top. The CEO and the rest of the board must lead the initiative, and AI adoption should be an OKR that counts towards bonuses and promotions.
• Ambassadors per department. Each department gets its own ambassadors: employees who are trained and combine their knowledge of the department with a knowledge of agents.
• Colleagues leading by example. According to Prosus, this is the real engine. What truly drives adoption is employees seeing their colleagues build and use agents.
• An agent that helps build agents. This worked particularly well. At iFood, more than 1,300 employees used such a co-creation agent to collectively build over 10,000 agents.

Our take

We recognise much of what Prosus describes, and a few lessons deserve extra emphasis for Dutch organisations.

Breadth is the prerequisite for depth. The top 2 per cent only exist because 60,000 agents were built. Those who only approve a handful of use cases in advance will never find their outliers. The lesson, therefore, is to make experimentation cheap and accessible, while keeping a sharp eye on measuring which agents scale up. Prosus even does this with its own scoring model, in which value, adoption, and quality collectively determine which agents have potential.

Start with the twenty, but look for your own outliers. The list of twenty use cases in the report is an excellent starting point for any organisation. However, the greatest return is often found in something specific to your company. Therefore, also look at work you currently leave undone because it is too expensive to execute. That is where the category of the long-tail restaurants and the affiliate marketplace lies.

Integration is the real bottleneck. In practice, an AI programme rarely stalls on the model. It stalls on the question of whether an agent is allowed to access the data, whether it may only read or also write, and who decides on that. Whoever sorts this out at the beginning saves months.

Let colleagues tell the story. Prosus is strikingly honest about external trainers: workshops by consultants or other external parties usually did not lead to high usage among average employees, according to the report. We endorse this, even though it is our own profession. An external party can train the first champions, build the first agents alongside them, and teach them how to bring others on board. After that, the baton must be passed to colleagues who resemble their peers. A good adoption programme thereby gradually makes itself redundant.

Measure from the business perspective. The deletion test is so simple it sounds almost naive, but it forces a manager to make concrete exactly what an agent delivers. That works better than a business case drawn up beforehand in a spreadsheet by the IT department.

The lesson of the electric motor

Prosus concludes with a historical comparison that we are happy to expand upon. Electrification only yielded the greatest productivity gains for factories when they redesigned their entire production process around electricity. Replacing a steam engine or waterwheel with an electric motor within an existing factory design yielded far less. The classic academic reference for this historical comparison is economist Paul David, who demonstrated in 1990 that it took decades before factories made that transition. They redesigned their buildings, providing a dedicated motor for each machine instead of one central drive shaft.

With AI agents, we are at a similar juncture. Many organisations are currently bolting an agent onto their existing processes. That is a logical first step, and the report shows that this step can already yield a lot. However, the greatest gains will only come when you redesign the work itself around what agents can do. Prosus is already experimenting with this, establishing departments structured around a desired outcome, with the necessary human roles only becoming clear afterwards. We will be reading Prosus's next report with particular attention.

The full report, The Coming Age of AI Colleagues, can be downloaded here. The press release is available at prosus.com.