The Adoption Curve Is the Finding
What the MIT Sloan/BCG agentic enterprise report reveals about the gap between deployment and readiness.
What the agentic enterprise report reveals about the gap between deployment and readiness
Source: S. Ransbotham, D. Kiron, S. Khodabandeh, S. Iyer, and A. Das, "The Emerging Agentic Enterprise: How Leaders Must Navigate a New Age of AI," MIT Sloan Management Review and Boston Consulting Group, November 2025. Based on a spring 2025 survey of 2,102 respondents across more than 21 industries and 116 countries, plus interviews with 11 executives.
Some management research rewards a second reading: once for what it argues, once for what its data quietly says about the organizations it surveyed. The ninth annual AI and business strategy report from MIT Sloan Management Review and Boston Consulting Group is one of those. Read the first way, it argues that agentic AI dissolves the categories long used to manage technology. Read the second way, it measures the distance between how fast organizations are deploying and how ready they are to manage what they have deployed.
A tidal wave of adoption, a trickle of strategy
That phrase is the report's own section heading, and the figures earn it. Traditional AI adoption has climbed to 72% over the past eight years. Generative AI reached 70% in three. Agentic AI has reached 35% in two, with a further 44% of organizations planning to deploy soon.
Much of this is distribution rather than demand. Vendors are embedding agentic capabilities as features in software organizations already own. Chevron's chief data and analytics officer, Margery Connor, describes the result: with the company standardized on a single vendor's platform, "more than half of the workforce has access to AI tools, and, by extension, access to agentic AI." Nobody decided to adopt. The capability arrived with the licence.
The authors state the consequence directly. Agentic AI, they write, is reaching enterprises faster than leaders can redesign processes, assign decision rights, or rethink workforce models.
One caveat belongs here, because it is this analysis rather than the authors'. Eight years, three years, two years looks like a compression curve, and it is tempting to read it as one. The report makes no such claim. Its own adoption chart shows traditional AI falling from 57% in 2020 to 50% in 2023 before climbing sharply, so the eight-year figure describes a recovery rather than a steady ascent.
What the authors do argue is that the speed is structural. Drawing on diffusion-of-innovations theory, they conclude that the rapid spread is no accident: the technology is built to minimize adoption friction, and it does. That distinction has teeth. A curve driven by frictionless distribution behaves differently from one driven by deliberate choice, and only the second kind slows down when someone decides to slow it.
Presentation on agentic AI during the AWS Summit Mumbai 2026. Photograph by Zaidusyy, 28 May 2026. Wikimedia Commons, licensed under Creative Commons Attribution-ShareAlike 3.0 (CC BY-SA 3.0).
When the categories stop routing
The report's most-quoted finding is that 76% of respondents say they view agentic AI as more like a coworker than a tool.
When three quarters of leadership reaches for the vocabulary of employment rather than the vocabulary of software, the existing management instrumentation is already partly obsolete. The authors put it more formally: the tool-coworker duality, they argue, breaks a management logic that has always assumed technology either substitutes or complements, automates or augments, counts as labor or as capital — never all at once. What organizations are left holding is a single system requiring both human resource approaches and asset management techniques.
Those distinctions were never merely descriptive. In most organizations they function as routing instructions, determining which committee reviews a deployment, which budget line funds it, and which policy framework applies. The report does not trace that mechanism, but it does document the standoff it produces. IT leaders want predictable, scalable systems with clear technical specifications. CFOs need investment models with measurable returns and depreciation schedules. HR executives require performance management frameworks and supervision protocols. One system, three incompatible sets of demands, none of them unreasonable.
The same collapse takes out the boundary between technology and strategy. That division survived several decades, with technology executives handling pilot, vendor, and infrastructure decisions while strategic executives handled markets, competition, and people. Agentic AI makes it untenable, because a single deployment reshapes process design, role structure, decision rights, and the culture of accountability at the same time. None of the tensions the authors identify, they conclude, can be resolved by any one function acting alone.
Four tensions, and the numbers underneath
Coverage of this report tends to flatten it into a governance story. The authors actually build their argument on four operational tensions: scalability versus adaptability, experience versus expediency, supervision versus autonomy, and retrofit versus reengineer.
Two of them are underdiscussed and worth pulling out. The investment tension turns on an anomaly in how these systems age — they depreciate through model drift while appreciating through fine-tuning, which means conventional depreciation schedules systematically undervalue them. The scope tension is a timing bet: reengineering a process around agentic AI takes long enough that the underlying technology may move before the project ships, which makes deliberately skipping a generation a defensible strategy rather than a failure of nerve.
The supervision tension is where governance discussion has concentrated, and the report frames it more sharply than its secondary coverage does. How do you supervise something designed to work autonomously? Traditional oversight models assume either full human control or complete automation — not systems that need some human control and varying degrees of independence. The sequencing advice that follows is unambiguous: build centralized governance infrastructure before deploying autonomous agents. And the authors move the question itself from guardrails to authority, asking not how to constrain a tool but "How do we assign decision rights, accountability, and oversight to actors we own but don't fully control?"
Several figures do more work than the headline adoption numbers. Expectations of AI decision-making authority are growing 250%, and 58% of leading adopters expect governance structures to change within three years. Among extensive adopters, 45% expect middle management layers to thin. The expectation of a changed operating model and redefined roles runs 42% among organizations with no adoption plans, against 66% among those already using the technology extensively.
One figure cuts against the autonomy narrative and deserves wider circulation. Seventy-nine percent of extensive adopters report investing in AI that generates insights for a human decision maker. Fully autonomous arrangements, where the system both decides and implements, peak at 54% in that same group. Today's dominant investment is augmentation, not delegation — a useful corrective for anyone reading the 250% growth figure as evidence that autonomous decision-making has already arrived.
A management problem in technical clothing
The report's conclusion is explicit: the challenge is organizational, not technological. Who owns AI agents, who supervises them, what authority they hold, how accountability is assigned — none of these are engineering questions, and no engineering team, however capable, can answer them alone.
Vibhor Rastogi of Citi Ventures puts the practical implication plainly: "we still feel that these AI agents should be treated like coworkers who need to be trained, coached, supervised."
Some organizations have already restructured on that logic. Moderna merged its technology and HR functions, a move the authors read as recognition that agents belong to the workforce rather than to the infrastructure. Their general recommendation follows the same shape: an HR function for agents, responsible for onboarding, evaluating, retraining, and eventually retiring them.
They are equally blunt about what speed proves, which is nothing. Adoption speed, they write, is not a measure of progress — and the alternative gets named for what it is. Adopting now and strategizing later is a high-stakes gamble that avoids the harder work of deciding what agentic AI is actually for.
The temptation at this point is to reach for a list of controls. The more useful move is a reframing, and the report's own is the sharpest available: the era of managing technology inside the IT department is finished, and governance has become a cross-functional obligation shared by IT, HR, finance, and operations.
Which returns us to the curve. Whether or not the compression continues is an inference rather than a finding, but the question facing most organizations doesn't depend on it. It is not whether their governance frameworks fit agentic AI as it exists today. It is whether those frameworks can be revised faster than the next wave of embedded capability arrives, and on the evidence assembled here, most cannot.
The report closes on the question worth sitting with: "Are we simply adding a new tool to our business, or are we introducing a new, nonhuman actor into our organization?"