For three years, the promise of AI agents — software that doesn’t just answer questions but executes multi-step tasks, makes decisions, and hands off work to other agents — lived mostly in research papers and demo videos. That era ended sometime in early 2026. Enterprises are now deploying autonomous AI agents at scale, and the productivity numbers coming back from early adopters are forcing a serious reckoning with what white-collar work will look like by the end of the decade.
From Chatbot to Colleague
The architectural shift from conversational AI to agentic AI is more than a marketing reframe. A chatbot responds. An agent acts. It can be given a goal — “reconcile last quarter’s vendor invoices against purchase orders and flag discrepancies over £5,000” — and it will plan a sequence of tool calls, access internal systems, reason through ambiguous results, and deliver a structured output, without a human shepherding each step.
Salesforce’s Agentforce platform, launched in October 2024, has become the clearest commercial proof point. The company reported at its March 2026 analyst day that more than 5,000 enterprise customers had deployed Agentforce agents in production environments, handling tasks spanning tier-1 customer support, sales pipeline qualification, and internal IT ticket resolution. In a disclosed case study, Wiley — the publishing group with significant UK operations — deployed agents to handle customer service enquiries during peak enrolment periods and saw a 40% reduction in case escalation volume within the first quarter of deployment.
Microsoft tells a parallel story. As of March 2026, Copilot Studio — its low-code platform for building custom agents — has been used by more than 85,000 organisations to deploy purpose-built agents across finance, legal, HR, and operations functions. Chief Executive Satya Nadella described the trajectory in a February earnings call as “the fastest-growing capability we have ever shipped to enterprise.”
The Numbers That Matter
McKinsey Global Institute’s 2025 labour automation update estimated that generative AI agents — as distinct from earlier robotic process automation tools — could automate or substantially augment 30% of work tasks across professional services industries by 2030. The distinction matters: earlier automation waves targeted repetitive, rules-based tasks. Agents are being deployed for work that previously required judgement: reading contracts for non-standard clauses, synthesising competitive intelligence from earnings calls, drafting regulatory submissions with citations.
The financial impact is beginning to show up in earnings reports. JPMorgan Chase disclosed in its Q4 2025 filing that AI-assisted processes had reduced processing time for routine legal document review by an average of 68%, while Morgan Stanley’s wealth management division reported that AI agents now handle the first-pass preparation of client portfolio review documents for approximately 70% of its adviser base.
Productivity gains are real, but they are not evenly distributed. Organisations with strong data infrastructure — clean CRM records, well-documented internal knowledge bases, structured workflow tooling — are seeing the largest gains. Companies attempting to deploy agents on top of fragmented legacy systems are discovering that the bottleneck is not the model; it is the mess underneath.
Workforce Implications: More Complex Than It Looks
The labour market question hanging over agentic AI is both more urgent and more nuanced than the previous automation debates. A 2025 Brookings Institution study examined early-adopter enterprises and found that headcount reductions in affected roles were, on average, smaller than feared — but workload per remaining employee had increased substantially. The pattern emerging in several sectors resembles what happened to radiologists after AI-assisted diagnostic tools arrived: the technology raised throughput expectations rather than eliminating positions outright, at least in the near term.
That calculus is not stable, however. Anthropic’s economic research team published a white paper in February 2026 noting that agent capability is improving faster than enterprises are deploying it — meaning the economic pressure on certain job categories will intensify before labour markets have fully adapted to current capability levels.
Regulators are beginning to pay attention on both sides of the Atlantic. The EU AI Act’s high-risk classification covers certain automated decision systems in employment contexts, and the EU AI Office issued guidance in March 2026 clarifying that AI agents making consequential HR decisions — including workload allocation and performance assessment — likely fall within scope. In the United Kingdom, the Equality and Human Rights Commission (EHRC) has flagged AI-driven workforce tools as an emerging area of scrutiny under the Equality Act 2010, particularly where automated systems may produce disparate outcomes across protected characteristics. In the United States, the EEOC published a request for comment on AI agent use in hiring and workforce management, signalling that formal guidance is forthcoming.
What Comes Next
The next twelve months will test whether the enterprise agent market develops the winner-take-most dynamics that characterise other enterprise software categories, or whether it fragments across specialised vertical players. Salesforce, Microsoft, ServiceNow, and SAP are all competing for the same budget line, while a cohort of vertical-specific startups — covering legal, finance, healthcare, and logistics — are betting that domain specialisation will prove defensible.
The underlying model providers — Anthropic, OpenAI, Google DeepMind — are watching carefully. Their frontier models are the engines inside most commercial agent platforms, which means the enterprise agent market is, structurally, a massive customer acquisition channel for whoever builds the most capable next-generation reasoning model. That alignment of incentives is accelerating capability development in precisely the directions enterprises need most: reliable tool use, long-horizon planning, and the ability to catch and correct its own mistakes.
The autonomous workforce is not arriving all at once. But it is arriving faster than most organisational planning cycles were built to handle.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.