Agentic AI deployment is outpacing the data needed to train it well
MIT Sloan and BCG reported in November 2025 that 35 percent of surveyed businesses had already deployed AI agents, with another 44 percent planning to do so. That 79 percent share of enterprises moving toward agentic AI is the headline. The less visible story, surfaced in an MIT News Q&A with Phillip Isola, an associate professor in EECS and a CSAIL member, is what that adoption wave is being built on: a small number of foundation models, wrapped in tools and integrations, and a training pipeline that lacks the data needed to make most agents reliable.
Most agent products look different on the surface and converge underneath. Isola describes agents as starting with a generative AI system at the core, then adding wrappers: tools the agent can call, memory layers, retrieval hooks, application-specific interfaces. The wrappers matter for product positioning, but the underlying capability ceiling is shared. A booking agent, a customer service agent, and a coding agent drawing on the same foundation model inherit the same blind spots, the same hallucination modes, and the same limitations on the kinds of actions they can plan and execute. Differentiation in the agent market is therefore likely to live in tool integration, UX, and enterprise plumbing rather than in core model behavior, which makes the procurement question harder than the marketing suggests.
The training data problem is the bottleneck that the adoption numbers do not address. Booking a flight is the kind of task that looks easy and is actually hard: there is no large corpus of step-by-step demonstrations showing how to navigate an airline checkout flow, recover from a payment failure, or negotiate a rebooking. Language models are trained on text, and most digital actions are not represented in text. Agents must learn by trial and error in environments that are hard to model, which is the inverse of the conditions where foundation models have scaled well. Scaling laws and pretraining corpora do not transfer cleanly to action selection, and the source does not describe a near-term data pipeline that would change that.
Coding is the one area where agentic AI has clear traction, and the reason is structural: the model can write code, execute it, observe the test result, and iterate. A clean feedback loop is the prerequisite for the trial-and-error learning Isola describes, and few domains outside software engineering provide one that is this direct. The same architecture applied to a high-stakes process like medical decision support or policy automation does not get the same reward signal, which is why the professor's framing of high-stakes domains as 'not ready' reads as a constraint on the technology, not a temporary lag. The survey number does not distinguish between pilots, internal tools, and customer-facing deployment, and the source does not characterize which of those buckets is driving the 35 percent figure.
The risk that follows from easy deployment is verification failure. Isola's 'vibe coding' framing captures the dynamic: when the agent produces output quickly, the human reviewer tends to check less. Bugs slip in, private data leaks, and the failure mode is not that the model is incompetent but that the operator stops being a competent reviewer. De-skilling is the longer-horizon version of the same problem: if the agent drafts the code, writes the math, and assembles the homework, the human skill used to verify those outputs atrophies before the model becomes good enough to remove the human from the loop entirely. The verification gap is structural rather than temporary, because ease of generation tends to suppress verification effort regardless of how accurate the underlying model becomes.
Isola flags a specific architectural uncertainty at the end of the Q&A: whether the next generation of agents is 'Claude with sensors and actuators' or a fundamentally different architecture trained on modalities beyond text. The wrapper model is productizing today's foundation models and will carry the industry through the next deployment cycle. Whether the 79 percent adoption share translates into reliable autonomy or into a larger surface area for verification debt is a question the source does not directly address.