There is a distinction in AI that the consulting firms are now quantifying — and it changes the investment thesis. The distinction is between AI as a tool (you ask, it answers) and AI as a worker (you assign, it executes). The first category — chatbots, copilots, search assistants — is where most enterprise AI spending sits today. The second category — autonomous agents that plan, execute, and adapt across multi-step workflows — is where the value curve is bending.
BCG's AI Radar 2025 estimates that agentic AI accounts for roughly 17% of total enterprise AI value today and projects that share rising to about 29% by 2028 — narrowing the gap on generative and predictive AI as the value mix rebalances. McKinsey's foundational work on generative AI sized the broader prize at $2.6 to $4.4 trillion a year across 63 use cases and 16 business functions; agents are the mechanism by which a growing slice of that gets captured rather than merely modelled. And the plumbing is arriving fast: the Model Context Protocol, the connective standard that lets agents call tools and reach other systems, passed more than 10,000 active public servers within its first year before Anthropic donated it to the Linux Foundation's new Agentic AI Foundation in December 2025 — a build-out that signals where the platform layer is heading, even if almost no Mittelstand yet has the operating model to exploit it.
The trajectory is clear. The readiness is not. McKinsey's 2025 State of AI survey is blunt on this point: in no single business function do more than 10% of organisations report scaling agents. The curve is bending. Almost nobody is standing where it bends.
What agentic AI actually means in practice
The shift from tool to worker is not metaphorical. A copilot suggests edits to a document. An agent drafts the document, checks it against compliance rules, routes it for approval, follows up when approval stalls, and revises based on feedback — without returning to the user between steps. A copilot answers a supply chain question. An agent monitors inventory levels, identifies reorder triggers, generates purchase orders, negotiates against contracted terms, and escalates exceptions to a human buyer only when the negotiation falls outside preset parameters.
This is the difference between augmentation and autonomy. In the Three Levels framework, copilots sit at Level 1 (assistance) — they enhance individual productivity within existing workflows. Agents with human oversight sit at Level 2 (augmentation) — they execute workflow segments while humans monitor outputs and handle exceptions. Fully autonomous agent-to-agent orchestration sits at Level 3 (autonomy) — multiple agents coordinate to run entire processes with minimal human intervention.
The economic distinction matters because the value curves differ. Level 1 produces linear gains: one copilot, one user, one productivity improvement. Level 2 produces workflow-level gains: an agent handles a process end-to-end, improving throughput for the entire function. Level 3 produces systemic gains: interconnected agents optimise across functions, creating compounding effects that no individual tool deployment can match.
Where the Big 3 see agent value concentrating
Deloitte's analysis identifies five domains where agentic AI is expected to have the highest impact: customer support, supply chain management, R&D acceleration, knowledge management, and cybersecurity. These share common characteristics: high transaction volume, multi-step processes with decision points, structured escalation paths, and measurable outcomes. They are, not coincidentally, the same domains where traditional workflow automation has historically produced the strongest ROI.
McKinsey's data shows adoption is accelerating but shallow. In the 2025 State of AI survey, 23% of respondents report scaling an agentic AI system somewhere in their organisation, and a further 39% have begun experimenting with agents. But in no single business function do more than 10% of organisations report scaling agents. The pattern is familiar: broad experimentation, narrow production deployment. The same scaling gap that plagued generative AI adoption is now reappearing one layer up.
Bain's analysis adds the economic backstory — and a caution about reading it too eagerly. In 2023 and 2024, Bain reports, tech-forward enterprises that broke through the pilot phase by scaling information retrieval and single-task AI achieved 10 to 25% EBITDA gains. That is the precondition, not the agentic payoff: the leaders banking those gains are the ones now turning to agents, having already built the data discipline and operating muscle that single-task AI demanded. The honest reading is sobering for anyone hoping agents are a shortcut. The firms widening the gap did the unglamorous work first. Agentic AI compounds an existing advantage; it does not manufacture one from a standing start.
The prerequisites the firms agree on
All three firms converge on one point: the bottleneck is not the technology. The models are capable. The infrastructure is maturing. The agent frameworks exist. What separates organisations that capture agent value from those that deploy agents that fail expensively is organisational readiness — and the specific dimensions of that readiness are well documented.
Bain outlines five critical actions for agent deployment. First, set top-down goals with measurable targets tied to business outcomes, not technology adoption metrics. Second, charge general managers — not CIOs — with ownership of AI transformation, because the value sits in business processes, not in IT infrastructure. Third, redesign entire workflows rather than inserting agents into existing process steps. Fourth, pursue pragmatic data curation — not perfect data lakes, but targeted data quality improvements in the specific workflows agents will operate on. Fifth, make deliberate build-buy-partner decisions based on where competitive advantage actually resides.
McKinsey's survey reveals the dominant barrier. Nearly two-thirds of respondents cite security and risk management as the top obstacle to scaling agentic AI. This is significant because it is not a technology barrier — it is a governance barrier. Agents that operate autonomously need guardrails: what decisions can they make, what thresholds trigger human review, what happens when they encounter edge cases, and who is accountable when they make errors. Most enterprises lack the governance frameworks to answer these questions, which means they lack the prerequisites to deploy agents safely.
BCG's perspective adds the maturity dimension. Its September 2025 analysis found that AI leaders are pulling away from laggards — double the revenue growth and markedly higher cost savings — precisely because they have already built the baseline: functioning data pipelines, established governance, experienced teams, and production-grade monitoring. Attempting to leapfrog from no AI straight to agentic AI is the enterprise equivalent of trying to run before walking. The organisations now capturing value from agents are, in the main, the same ones that spent the previous two years building foundational capabilities — not newcomers who bought an agent platform last quarter.
Why agents fail: the anti-patterns
The most expensive failure mode is deploying agents without workflow redesign. An agent inserted into an existing process inherits the inefficiencies of that process. A claims-processing agent that follows the same steps a human adjuster follows — just faster — captures a fraction of the available value. An agent deployed into a redesigned claims workflow, where classification, routing, investigation, and settlement recommendation happen as coordinated autonomous steps, captures the full value. Bolting an agent onto the old process is the default move in most organisations, because it is the cheapest to authorise and the easiest to demo — and it is also why so many pilots stall before they ever justify their cost. The evidence on workflow redesign is unambiguous about the consequences.
The second failure mode is insufficient data architecture. Agents consume and produce data continuously. They need real-time access to operational data, they need to record their decisions and reasoning for audit trails, and they need feedback loops that flag when their outputs diverge from expected parameters. Most enterprise data architectures were designed for human consumption — dashboards, reports, batch exports. Agent-ready data architecture requires machine-readable, low-latency, well-governed data pipelines. This is unglamorous infrastructure work, and it is non-negotiable.
The third failure mode is governance vacuum. An agent without clear operating boundaries is a liability. It will make decisions in edge cases it was never designed for. It will take actions that seem locally optimal but are globally harmful. It will accumulate errors that no one catches because no one is monitoring. This is not a hypothetical concern: in McKinsey's survey, nearly two-thirds of respondents name security and risk as the chief obstacle to scaling agents, and Deloitte finds only one in five organisations has a mature model for governing autonomous agents at all. The leaders banking real returns are the ones that defined decision boundaries, built monitoring, established escalation protocols, and assigned accountability before they let an agent act unsupervised — not after.
Mapping agent readiness to the Three Levels
The Three Levels framework provides a direct readiness assessment for agentic AI. If an organisation is operating at Level 1 — using AI as individual tools, chatbots, and copilots — it is not ready for agents. The foundations are not in place: workflows have not been redesigned, data pipelines are not production-grade, and governance frameworks do not exist.
Level 2 is the agent deployment zone. At this level, organisations have redesigned workflows around human-AI collaboration, established monitoring and governance, and developed the operational muscle to manage AI systems in production. Agents operate with human oversight — they execute autonomously but within defined boundaries, with humans reviewing outputs and intervening on exceptions. This is where the agent value BCG measures is overwhelmingly being generated today.
Level 3 is the frontier. Agent-to-agent orchestration, where multiple autonomous systems coordinate across functions, is where the value curve steepens — and where the governance requirements become exponentially more complex. The rapid build-out of the Model Context Protocol — past 10,000 public servers in its first year, now stewarded by the Linux Foundation — shows the connective infrastructure for Level 3 is arriving. The operating models to use it safely are not. Very few enterprises, and vanishingly few in the Mittelstand, are ready to inhabit this level today.
What this means under European rules
For a DACH Mittelstand, the governance question is not optional — it is statutory. An autonomous agent that screens credit applications, sorts CVs, prices insurance, or makes decisions in critical infrastructure can fall squarely within the high-risk categories the EU AI Act regulates, with obligations covering risk management, human oversight, logging, and transparency. Governance rules for general-purpose AI models have applied since 2 August 2025, and the high-risk obligations follow on a staged timeline now being adjusted through the Commission's Digital Omnibus. The practical implication is the same regardless of the exact date: the audit trails, decision boundaries, and human-in-the-loop controls that make an agent legally defensible are the same controls that make it operationally trustworthy. Build them once and they serve both masters. Skip them and the agent becomes a compliance liability the moment it acts on a real customer.
The strategic implication
The consulting firms agree: agentic AI is not a feature to enable — it is an operating model to build toward. The $2.6 to $4.4 trillion value opportunity McKinsey identifies is not a pool of money waiting to be collected by buying the right tools. It is a value layer that emerges from redesigned workflows, mature data architecture, robust governance, and organisational readiness to delegate decisions to autonomous systems.
The 17% of AI value that agents represent today is growing toward 29%. The question for enterprise leaders is not whether to invest in agentic AI, but whether their organisation has the foundations to capture that value — or whether they will join the majority that deploys agents prematurely and learns that an autonomous system without guardrails is more expensive than no system at all.
A Fit Call assesses where your organisation sits on the Three Levels and what it takes to reach the level where agentic AI creates value rather than risk. No pitch. Just an honest readiness assessment.
References: BCG, "AI Radar 2025: From Potential to Profit — Closing the AI Impact Gap," January 2025 (bcg.com/publications/2025/closing-the-ai-impact-gap); BCG, "AI Leaders Outpace Laggards," September 2025 (bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings); McKinsey & Company, "The State of AI in 2025: Agents, Innovation, and Transformation," November 2025 (mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai); McKinsey Global Institute, "The Economic Potential of Generative AI," June 2023 (mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier); Bain & Company, "Technology Report 2025," September 2025 (bain.com/insights/topics/technology-report); Deloitte, "State of AI in the Enterprise, 2026," October 2025 (deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html); Anthropic, "Donating the Model Context Protocol and Establishing the Agentic AI Foundation," December 2025 (anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation); European Commission, "AI Act — Regulatory Framework," 2025 (digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai).