What We Brought to Geneva: AI Agents and the Future of Work
Last month I was at the Palais des Nations in Geneva for the annual AIFOD Summit—AI for Developing Countries Forum. I chaired an Independent Review Board panel titled “AI Agents and the Future of Work,” alongside three co-chairs: Pavlo Yalovol (VP Innovation, Intetics), Shikhin Agarwal (Founder and CEO, StatsLateral), and Georg Zangl (COO, MudoZangl). The room was full. That’s no small detail: it means the question of how we organize work between people and autonomous agents has already entered the real priorities of decision-makers, not just vendor slide decks.
I wanted to share here, calmly, what we put on the table and why.
The Problem We Wanted to Address
The vocabulary we all use today—generative AI, AI Agent, Agentic AI—describes one thing: how much a system can do on its own. It’s a legitimate axis, and the companies selling it have every right to be evaluated on that. But it doesn’t answer a different need that matters just as much to an organization: those actions, however autonomous, are work only when someone is accountable for them.
For the IRB panel we proposed calling this second dimension accountability: not how much a system can do, but under what conditions its actions count as work for which an organization can be held responsible. The underlying idea is simple, and I believe important: automation is not work. To call something work, there must be someone who takes responsibility for the results. A system can execute excellently and remain pure automation if no one answers for the outcome, if nothing can be verified, if there is no boundary within which it is required to operate.
A Work Operating System is the category of solution that fully covers both dimensions together—capability and accountability—for every digital employee, as a property of the system itself, not as a custom-built arrangement around a single agent. It’s not a better point on the capability scale. It’s what brings every piece of “AI-enabled” work to where it should be on both axes together.
The Comparison: Eight Platforms, Four Categories
We didn’t want to stay in theory. With the IRB we selected eight platforms representative of four categories that today claim to be capable of “doing work”: ecosystem platforms (Microsoft Copilot, Google Gemini Enterprise), agents integrated into ERPs (SAP Joule, Oracle AI Agent Studio), cloud AI infrastructures (Amazon Bedrock/AgentCore, Anthropic Claude Platform), and digital employee platforms (Kore.ai, Dorg.pro).
To keep the comparison honest, we asked Anthropic to collect only public information—official website and documentation released by each vendor—and had the scores independently verified by ChatGPT, applying the same set of eight criteria to all: sovereignty, resilience, embedded environmental and social mitigation, anti-lock-in, self-configuration of competencies, decoupling between governance and competence, adjustable behavior, cost of adoption at scale.
The result wasn’t “who wins.” As we wrote in the conclusions shared with the IRB: none of the eight platforms is the wrong answer. Each in its own way exceeds the minimum threshold that separates simple automation from real work. The choice among them is a matter of organizational fit, not absolute quality. That said, I won’t hide that seeing Dorg.pro position well on nearly all criteria—particularly on embedded mitigation and cost of adoption at scale—was a valuable confirmation of architectural choices we’ve been pursuing for some time, not a surprise we invented for the occasion.
The Point That Really Sparked Discussion
The data point that got the strongest reaction in the room wasn’t about performance, but about the freedom to leave. Two of the most widely adopted enterprise platforms score the lowest possible on vendor independence. And one of the major ERP vendors describes its agents in its own marketing materials as “built in, not bolted on”—an elegant way of saying that changing vendors means losing not just the software, but years of accumulated rules and workflows.
Only three of the eight platforms we examined contractually guarantee portability of accumulated configurations, not just raw data. Hence the question I posed to the panel, which I’m sharing here because I believe it deserves to remain open beyond that room: should we allow public and institutional bodies to sign contracts for AI agents that provide no exit path?
The draft standard we discussed is simple to state and hard to ignore: public procurement of AI-agent platforms should require contractual guarantee that the organization can extract accumulated data, configurations, and workflows—not just data alone—in a usable format, within a defined timeframe, at no additional cost.
Why This Isn’t an Abstract Issue for Us
I say this as president of a foundation that has built its mission on this principle, so I accept the partiality that entails: sovereignty is the freedom to choose and the freedom to leave. If the second is missing, the first is just a word in the contract. It’s why, when we design a digital employee, the portability criterion isn’t a module to add later: it’s part of the architecture from the start.
It’s not a position I expect everyone to share equally—for a business with a single specific need, sovereignty over inference matters relatively little. But for a bank, a hospital, a public administration, or simply for an organization that intends to build a long-term relationship with its technology, it’s the right question to ask before signing, not after.
What We’re Taking Home
The complete panel materials—framework, comparison table, and recommendations shared with the IRB—are available on request, along with the questions we left open to the Board: what criteria are still missing to evaluate a Work Operating System designed for an emerging economy, and whether the “future of work” is today a matter for discussion or already a matter for decision.
Sincere thanks to Pavlo Yalovol, Shikhin Agarwal, and Georg Zangl for sharing the panel with rigor and perspectives that were new and complementary to mine, and to all the delegates who turned an hour-long session into a real discussion, not a sequence of slides.
Flavio Bordignon — President, Dorg Society Foundation
