When a company tells you they've deployed thousands of AI agents, your first instinct should be skepticism. Not because the number is fake, but because deployed doesn't mean working.
Prosus, one of the world's largest builders of AI agents, just published the first real data on what agentic AI actually does at scale. Over eighteen months, their 40,000 employees deployed 60,000 AI agents across a global portfolio. The report is blunt. The headline: only 2% of those agents drive a disproportionate share of business impact.
This isn't a failure of execution. This is how power laws work. But for marketing leaders and executives betting their 2026 roadmap on AI agents, it changes everything about where to focus.
The 20 Use Cases Nobody Can Avoid
The Prosus research found something unexpected: across entirely different industries, geographies, and languages, with no directive from HQ, portfolio companies kept building the same 20 use cases. These aren't sexy agents. They're not customer-facing AI that generates headlines. They're the unglamorous backbone of operations.
Data analytics and market intelligence dominate at 18% of all agents deployed. Operations is second at 15%. Then there's a category nobody talks about openly: 14% of agents live outside any formal department, sitting in employees' personal workflows as private AI assistants. That's not enterprise adoption. That's people solving their own problems because the company hasn't.
The productivity math is brutal. Eighty-two percent of agents deliver under 20 hours saved per month. That's less than half an hour per work week per agent. A small tier saves 20 to 173 hours monthly. And then there are the outliers, less than 1%, that operate at a completely different scale. One agent in the report managed communications and onboarding for a new affiliate marketplace projected to generate $83 million in annual revenue.
That's the 2% right there.
The Cost Trap Nobody Sees Coming
Here's where it gets expensive. Once an AI agent works, nobody wants to switch models. A team builds an agent on Claude or GPT-4. It works. Six months later, a cheaper model becomes viable for that exact task, but the team resists moving because switching creates friction. Training people, updating prompts, testing behavior in production.
So organizations end up paying premium prices for tasks that cheaper alternatives could handle just as well. Multiply that across 60,000 agents, and you're spending millions on model overkill. The report doesn't put a number on it, but the implication is clear: cost creep happens quietly.
This matters to CMOs right now. If your marketing ops team is evaluating AI agents for analytics, campaign optimization, or lead scoring, the natural instinct is to pick the best model, not the cheapest. But that instinct costs you across 2027.
The Conversion Illusion
Prosus breaks agent complexity into four tiers that map directly to human employee seniority. Senior-level agents handle the most complex work. But here's what the data shows: daily usage splits almost evenly between senior and junior agents. The simple stuff gets used every single day. The complex, sophisticated agents get used less.
This is the opposite of how most organizations pitch AI agents internally. They lead with the impressive cases, the transformation stories, the senior-level use cases. Meanwhile, the actual impact is distributed across hundreds of simple agents doing lightweight work consistently.
For marketing teams, this pattern is critical. You don't need an AI agent that writes your entire campaign strategy. You need agents that handle repetitive micro-tasks: pulling historical performance data, flagging anomalies in creative performance, surfacing competitive moves, organizing research. The unsexy stuff. That's where real productivity gains happen.
What Happens Next
Prosus is explicit about where this is heading: autonomous, AI-enabled organizations where sales, operations, and customer support aren't coordinated by org charts anymore, they're coordinated by networks of AI agents built around outcomes.
That's not happening in 2026. The infrastructure isn't there yet. The culture isn't ready. But the companies building the right 20 use cases now, with the discipline to measure which 2% actually matter, will own the next phase.
For you, that means three things:
Stop treating AI agent deployment as a success metric. Treat impact as the metric. If an agent doesn't save measurable time, save money, or improve an outcome, it's a cost center dressed up as innovation.
Build the boring cases first. Data analytics, ops automation, personal assistant workflows. They're not exciting in a board meeting, but they're the 20 cases that work across industries.
Plan for model switching friction. When you pick a model for an agent today, assume you'll have switching costs six months from now. Price that in.
The 2% will emerge over time. But only if you measure. Prosus has the advantage of scale and data infrastructure most companies don't have. You don't need that scale to find your 2%. You just need discipline.