The Real Cost of Agentic AI Is the Data Behind It

On August 12, 2026, MIT Technology Review published a piece on scaling AI agents with trustworthy data, making a point that gets buried under most agentic AI coverage: the shift from answering questions to taking actions means an agent needs data from across the enterprise, structured and unstructured, with the right business context attached. The piece frames this as an infrastructure problem. It is that, but the framing understates how much it costs and why the cost cannot be engineered away by a better model.

The line between answering and acting

A system that answers questions can be wrong and survive it. Ask a chatbot what a hotel's cancellation policy is and get a stale answer, and the guest asks a human, or asks again tomorrow, and the error corrects itself before it costs anything. A system that acts does not get that grace period. An agent that cancels the booking, refunds the deposit, or reallocates the room has already spent the error by the time anyone notices it. The question is not whether the agent had access to data. It almost always did. The question is whether the data it acted on was the same data every other system in the business believed to be true at that moment.

This is the failure mode that shows up first in real deployments, and it shows up as a small, boring contradiction rather than a dramatic one. A property has three places a booking can live: the WhatsApp thread where the guest confirmed, the paper diary at the front desk, and the property management system the owner actually trusts for occupancy reports. An agent wired only to the PMS will confidently double-book a room the front desk already knows is taken, because the front desk's version of the truth never made it into the system the agent reads. The agent did not fail because the model reasoned badly. It failed because nobody had decided, in advance, which of three records was authoritative, and the agent inherited that unresolved argument and acted on it.

The plumbing problem is a governance problem

MIT Technology Review is right that this is an infrastructure demand, but calling it infrastructure makes it sound like a procurement line item, a pipe you buy once and connect. It is closer to a governance exercise that happens to require pipes. Someone in the organization has to decide, system by system, which record wins when two disagree, and then enforce that decision in code so the agent never has to guess. That someone is rarely a data engineer. It is usually the person who runs the front desk, the registrar, the ward clerk, the person who has watched the diary and the PMS disagree for years and simply worked around it by phone. Formalizing that judgment into a rule an agent can execute is the actual cost of agentic AI, and it is a cost measured in meetings and decisions before it is measured in compute.

The regional regulatory conversation is starting to catch up to this. CRASA's Legal and Policy Committee opened a meeting this week on AI and data governance across regional ICT regulation, and CMSWire published a six-point governance framework for enterprise agents on the same day the MIT Technology Review piece ran. Neither is a coincidence. The industry is converging on the same realization from two directions at once: the model side keeps getting cheaper and more capable, and the unresolved question keeps being what data the model is allowed to trust and act on. Governance frameworks are showing up because the technical problem was never really technical.

The objection worth taking seriously

The strongest counterargument is that this is a temporary problem. Give the model a larger context window, better retrieval over unstructured records, and the ability to reason across contradictory sources, and it will simply absorb the mess that used to require a human to sort out first. Under this view, the data foundation work described in the MIT Technology Review piece is a stopgap that better models will make unnecessary within a generation or two of releases.

This does not survive contact with what a contradiction actually is. A larger context window lets a model see the front desk diary and the PMS at the same time. It does not tell the model which one the property owner has decided is authoritative when they disagree, because that decision does not live in either document. It lives in an organizational judgment that has never been written down, and no amount of retrieval finds a fact that was never recorded. A model can be arbitrarily good at reasoning over ambiguous inputs and still be reasoning over the wrong one, confidently, because confidence was never the missing ingredient. Authority was.

What this changes for anyone evaluating an agent, rather than a chatbot with better manners, is the order of the work. The data reconciliation has to happen before the agent is given permission to act, not after a pilot reveals the contradiction the hard way. A vendor who leads with model quality and treats data trustworthiness as a later phase has described a system that answers questions well. The question worth asking on Monday is not which model the agent runs on. It is which system in the business gets to be right when two of them disagree, and whether anyone has actually written that answer down.

Sources

Scaling AI agents with trustworthy data (MIT Technology Review, 2026-08-12)