In 2025, Fortune 500 companies deployed an average of 35 autonomous artificial intelligence agents in their operations, compared to fewer than five two years earlier. This is not a metaphor about the speed of technological change. It is a concrete organizational fact, with consequences for who does what, who decides what, and who bears responsibility for what.

The Stanford Institute’s annual report on artificial intelligence published in 2026 documents this shift with unusual precision for a subject where media noise typically exceeds signal. AI ceases to be a tool that a human uses. It becomes an agent that acts, makes decisions, and interacts with other systems without human validation at each step. This shift is discreet in organizational charts. It is massive in practices.

The Essentials

  • Fortune 500 companies deployed an average of 35 autonomous AI agents in 2025, compared to fewer than 5 in 2023, according to Stanford’s AI Index 2026.
  • The most advanced sectors (finance, logistics, legal services) use these agents for tasks involving repeated decisions: credit validation, route planning, contract review.
  • The governance challenge remains open: in the majority of pioneering companies, the chain of responsibility in case of agent error is not formally defined.
  • Emerging organizational models (human supervisor of a fleet of agents, AI workflow architect) are being tested but not yet stabilized.
  • The next round of sectoral negotiations in Europe, particularly in banking and logistics, should be the first to explicitly address the question of autonomous agents.

Thirty-Five Agents Per Company, and Nobody Really Planned for It

The Stanford figures deserve closer examination. Counting active AI agents in large enterprises is not a metric that human resources departments were tracking. Most deployments occurred through IT departments and operations directors, often outside discussions about social strategy. An agent that automatically processes expense reimbursement requests, plans delivery slots, or scans thousands of contracts to identify risk clauses does not appear on the organizational chart. It has no job description. It is not subject to annual evaluation.

This organizational discretion is partly structural. AI agents are often software components purchased as SaaS subscriptions or built internally by technical teams without explicit mandate to consult employee representatives. The result is a transformation that advances at the pace of software deployment cycles, not at the pace of labor negotiations.

This gap is not unique to agents. It reproduces the pattern observed with automation of junior developers, where effects on hiring emerged before companies had time to establish a framework. The difference with autonomous agents is scale. Thirty-five agents per Fortune 500 company means thirty-five entities that act, not thirty-five tools that a human operates.

What These Agents Actually Do, Sector by Sector

Agentic AI is not homogeneous. Its most advanced uses are concentrated in three sectors where high-frequency repetition of standardized decisions creates fertile ground.

In finance, agents manage entire segments of the business credit pipeline. JPMorgan Chase has deployed agents capable of analyzing balance sheets, cross-referencing sectoral data, and producing a credit recommendation that the human analyst validates or rejects. According to the bank, processing SME files has been significantly accelerated. The analyst does not disappear, but their role refocuses on atypical cases, complex files, decisions that involve client relationships. It is a reconfiguration, not an elimination.

In logistics, DHL and Amazon have deployed planning agents that optimize routes in real time by integrating weather data, traffic constraints, and client priorities. These systems make several thousand micro-decisions per minute. No human validates each package assignment to a delivery driver. The human supervisor intervenes on exceptions, allocation conflicts, out-of-norm situations.

In legal services, firms like Allen & Overy and Linklaters have made public their deployments of agents for contract review. An agent can process a thousand contracts where a junior associate would process ten per day. The impact on entry-level hiring in corporate law is documented by several American bar associations: first-year associate positions have declined for the third consecutive year.

The Governance Question Nobody Wants to Settle

If deployments are advancing fast, responsibility frameworks remain fuzzy. This is where optimism must be clear-eyed.

When an AI agent makes an erroneous decision, who is responsible? The IT department that deployed the system? The supplier of the underlying model? The manager who configured the parameters? The company as a whole? In the majority of pioneering companies, according to a Gartner survey published in 2025, this chain is not formally documented. Companies deployed before governing.

This is not a criticism of their competence. It is an observation about adoption speed. AI governance frameworks in enterprises are typically built two to three years after first large-scale deployments. We saw this with HR scoring algorithms, with customer recommendation systems, with dynamic pricing models. Governance follows use, it does not precede it.

But autonomous agents pose a new problem compared to previous decision-making algorithms. A scoring model produces a recommendation. An agent acts. It sends emails, places orders, modifies parameters in third-party systems. An agent’s error is not an incorrect number in a spreadsheet. It is an action in the real world, with effects that can propagate before a human detects them.

The European Union has tackled this problem head-on in the AI Act, whose provisions on high-risk AI systems now explicitly include certain categories of autonomous agents. But the text defines obligations for transparency and audit, not a responsibility architecture. The question of who pays when an agent makes a costly error remains open in most jurisdictions.

The New Professions Emerging Around Agents

The picture is not uniformly dark. In the most advanced companies, new roles are appearing that did not exist in 2022.

The most documented is that of agent supervisor, sometimes called AI workflow manager. This profile is responsible for supervising a fleet of agents, detecting behavioral drift, adjusting parameters, and escalating abnormal cases to business teams. It is a profession halfway between software engineering and team management. It requires an understanding of agent decision logic, ability to read system logs, but also operational judgment on what constitutes legitimate exception versus what constitutes systematic error.

Other roles emerge upstream: AI workflow architects, who design the processes in which agents operate; operational AI ethics officers, distinct from AI Act compliance lawyers. Downstream: agent audit profiles that verify ex-post that decisions made comply with defined rules.

These professions do not compensate in volume for positions that disappear. But they signal something important: the economy of agents is not an economy without humans. It is an economy where human work shifts toward supervision, design, and control of automated systems. This shift toward roles of judgment and governance corresponds to what Tyler Cowen and other optimistic economists anticipated as the natural repositioning of work in the face of automation. The macroeconomic debate remains lively, but the micro signals are there.

What Collective Bargaining Has Not Yet Integrated

The social front is lagging behind the technical front. In Europe, employee representative bodies have broadly approached AI through the lens of decision-support tools and employee surveillance. French social and economic committees obtained, via the multi-employer agreement on professional transitions, a right of information-consultation on AI projects affecting working conditions. But the notion of autonomous agent—a system that acts rather than recommends—does not yet appear in collective agreement texts.

This void is not inevitable. In Germany, the co-determination model has allowed Betriebsräte to negotiate framework agreements on AI use in several large industrial companies. These agreements define the scope of system autonomy, modalities for human recourse, and training obligations for supervisors. They do not yet explicitly cover agents in the technical sense, but they establish contractual precedent on which future negotiations can build.

The timeline is known. Collective agreements in French banking, German logistics, and British financial services come up for renegotiation between 2026 and 2028. This will be the first negotiation cycle where AI agents will be sufficiently present in operations for their social governance to become an explicit issue. Trade unions that prepare their doctrine in advance will have significant framing advantage. Those arriving at the table without precise technical vocabulary risk negotiating on the employer’s terrain.

The Question the Coming Months Will Force to Be Settled

The fundamental issue is not whether AI agents will be deployed. They are being deployed. The issue is who defines the rules of the game during the stabilization period that is opening, and whether these rules will be negotiated or simply imposed by accomplished technical fact.

Two models clash in practice. The first, dominant today, is deployment first, governance later. It has the advantage of speed and the disadvantage of improvisation: when an incident occurs, companies respond urgently, without pre-established framework. The second, emerging, is coordinated deployment: trade union actors, lawyers, technical teams, and operations management co-construct the rules before systems go into production. This model is slower. It is also more robust in the face of edge cases.

AI’s energy consumption will face increasing regulatory constraints. Governance of autonomous agents will probably follow the same trajectory: first ignored, then endured, then codified under pressure from an incident visible enough to force legislators’ hands. The question is whether companies and social partners will have built a framework before that incident, or after.

The coming rounds of collective bargaining are a concrete opportunity. Not to slow deployments, but to define the conditions under which they take place. That is not the same thing.

Sources

  1. Stanford HAI, AI Index Report 2026 — https://aiindex.stanford.edu
  2. Stanford HAI, AI Index Report 2026 (official link) — https://hai.stanford.edu/ai-index/2026-ai-index-report
  3. Gartner, Survey on AI Agent Governance in Enterprises, 2025 — citation without stable direct URL
  4. Gartner, AI Agent Governance, May 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure
  5. European Parliament, AI Act — Official text — https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence
  6. European Commission, Guidelines for classification of high-risk AI systems, May 2026 — https://digital-strategy.ec.europa.eu/en/library/draft-commission-guidelines-classification-high-risk-ai-systems
  7. JPMorgan Chase, institutional communications on AI deployment, 2024-2025
  8. Allen & Overy / Linklaters, annual reports 2024
  9. A&O Shearman — Agentic AI agents (official, February 2026) — https://www.aoshearman.com/en/news/ao-shearman-and-harvey-to-roll-out-agentic-ai-agents-targeting-complex-legal-workflows
  10. Microsoft Cyber Pulse 2026 — Fortune 500 and AI agents — https://www.microsoft.com/en-us/security/blog/2026/02/10/80-of-fortune-500-use-active-ai-agents-observability-governance-and-security-shape-the-new-frontier/
  11. DHL Freight, AI route planning — https://dhl-freight-connections.com/en/trends/ai-route-planning/
  12. NALP / ABA Journal, Associate recruitment — https://www.abajournal.com/web/article/law-firms-reduced-the-pace-of-associate-hiring-shifting-to-new-talent-model-report-says