Most companies are still early.
Much of the global conversation around AI remains fixated on model breakthroughs and timelines to AGI. Despite the headlines, most organisations are still relatively early in their adoption journey.
We set out to share our experience as a large global technology company that has transformed its business into an AI-first organisation — with over 40,000 employees across a wholly-owned global portfolio, who have together built more than 60,000 AI agents.
We believe this places us among the largest operators of AI agents globally whose core business is not the sale of AI services to other companies.
The classic power law holds.
Approximately 2% of active AI agents drive a disproportionate share of business impact. The implication is clear — the first priority for every organisation is to double down on that 2%.
Organisations should also use data-driven analysis to identify — and then nurture — the agents that are trending toward the power-law group. Without this signal, high-potential agents quietly stay underused while attention scatters across the long tail.
The same 20 use cases, everywhere.
Across industries, geographies and languages — with no central mandate — portfolio companies independently converged on the same set of high-ROI agentic use cases.
While not a comprehensive list, these are a strong starting point that all organisations should implement. They also offer a window into the future: tasks that will likely become “default AI” within a few years. Three examples below; the full list is in the report.
Agents map to seniority levels.
AI agent complexity falls into four tiers that map closely to human employee seniority levels.
Senior-level agents have the largest total number of users. When looking at daily usage, however, it splits almost evenly between senior and junior agents — revealing that simpler agents carry significant weight in daily tasks.
Departmental breakdown.
We identified 54 distinct AI agent tasks across corporate functions. A full breakdown by department and task is included in the downloadable report.
Data analytics and market intelligence claimed the largest share at 18%. Operations follows at 15%, for tasks such as forecasting supply or managing inventory. Notably, the third largest share — 14% — sits outside any formal department: employees' personal AI assistants.
Three bands of hours saved.
Productivity agents — where ROI is measured in hours saved — divide cleanly into three distinct bands.
The majority — 82% deliver under 20 hours saved per month, typically personal assistants. The middle tier (17%) saves 20–173 hours monthly — roughly up to one FTE. At the top, less than 1% operate at a different scale entirely, delivering the equivalent of thousands of hours of monthly work.
Less than 1% operate at a different scale entirely.
Delivering the equivalent of thousands of hours of monthly work — a handful of agents carrying outsized impact.
A small group of outliers.
Value agents — measurable in revenue growth or cost reduction — follow the same shape. A handful deliver outsized gains.
Under $1M in annual value
For example, reducing annual audit costs.
$1M–$10M in annual value
For example, a customer-facing assistant answering niche vacation-rental questions.
Tens of millions of dollars annually
One portfolio company used agents to manage communications and onboarding for a new third-party affiliate marketplace. Projected annual revenue: $83M.
Most models are now good enough.
Across-the-board improvements mean nearly every modern model can handle nearly every agent task. The challenge is no longer capability — it's cost discipline.
Our internal agentic AI platform, Toqan, offers ten different AI models to our 40,000+ employees. The latest cutting-edge models are only needed for the most complex tasks.
AI costs are volatile and hard to predict. We've found it more effective to focus on optimisation and empower individual business lines to make their own cost decisions. Our solution is a two-tier approach:
Free under 200/hr.
Production above.
Anyone can use AI freely under 200 requests per hour. Above that threshold, users need departmental approval — the business line itself is best placed to weigh benefit against cost.
A three-step adoption framework.
We have developed a three-step process for driving adoption of agentic AI at large organisations. The need is real — most large organisations struggle to move from experimentation to scale.
From AI colleagues to AI-led organisations.
After scaling to 60,000 AI agents, we now have a clear view of what an AI-driven organisation looks like in practice — and what comes next.
What we observe across our portfolio is transformation rather than total disruption. We see three trends:
Electrification only paid off once factories were rebuilt around it.
The same framing applies to AI. The biggest gains come when companies re-build core functions from the ground up with AI in mind — not when they layer agents onto existing structures.
It is only when companies re-build core functions from the ground up with AI in mind that they will experience the biggest gains.