A Correct Prediction Is Not the Same as the Right Action

On October 5, MIT Technology Review reported that the argument over whether predictive models beat statistical forecasts is settled. The question enterprises are actually wrestling with now is different: how to let a predictive system act on its own conclusions without drifting from what the business intended. Vishal Gupta, a partner at Everest Group, put it plainly in the piece: "Enterprises are done with a backward-looking point of view; they want to be more forward-thinking." The frontier has moved from prediction to autonomous decision making, and the gap between the companies that have figured this out and the ones that haven't is widening.

That sentence, drift from business intent, is doing more work than it looks like it's doing. A predictive model can be correct and still take the wrong action, because correctness is measured against the data and intent is measured against something the data was never asked to represent. A model forecasting cash flow can be right about the number and still recommend a payment timing that breaches a relationship the finance director cares about and never told anyone to encode as a rule. The model has no way to know that rule exists until it breaks it.

This is the actual shape of the demo-to-deployment gap for agentic systems, and it is worth being precise about why demos hide it so well. A demo tests whether the prediction is accurate. You feed it historical data, you backtest, you show the forecast line tracking the actual line, and everyone in the room nods. None of that tests whether the action the system would take, given that forecast, matches what the business would actually want done. Accuracy is a property of the model. Alignment with intent is a property of the organisation, and organisations are bad at writing their intent down because most of it lives as shared understanding among the people who have been doing the job for years, not as a document a system can read.

The intent nobody wrote down

Ask any manager what the rule is for approving a vendor payment above a certain threshold, or for rebooking a guest after a cancellation, or for flagging a grant application for manual review, and you will get an answer immediately. Ask them to write that answer down as a precise, testable constraint, the kind a system has to check before it acts, and the answer gets vaguer. There are exceptions nobody mentioned. There is a relationship with a particular supplier that changes the rule. There is a seasonal pattern the manager holds in their head and has never had to articulate because no system was ever going to act on it without asking first. A predictive system that only recommends leaves that gap for a human to close at the point of action. An agentic system that acts closes the gap, or exposes it, at the point of deployment.

You can see the shape of this in how the better-built agentic finance tools are handling it. Bujeti's newly launched Brain platform gives finance teams AI teammates that find invoices, automate vendor payments, and handle budgeting, but the payment automation runs through multi-level approvals rather than acting unchecked. That approval layer is not a concession to caution bolted on afterward. It is the mechanism that catches the gap between what the model calculated and what the business actually wanted, before the money moves instead of after someone notices it moved wrong. A system that can act but cannot be stopped mid-action has pushed all of the intent-checking work back onto a human who now has to audit outcomes instead of preventing them.

The strongest objection to this

The obvious response is that this is nothing new. Every automation project has always needed guardrails and sign-off thresholds, and calling it "business intent drift" is just a new name for old change management. There is something to that. Organisations have been writing approval rules into ERP systems for decades, and nobody needed the word agentic to justify a second signature on a large invoice.

The part that is actually new is the speed and the reversibility of the error. A legacy system that misapplies a rule produces a wrong report that a human reads before acting on it, and the human is the last line of defence. An agentic system that acts on a prediction produces a completed action: a payment sent, a booking confirmed, a stock order placed. The error is no longer caught at review time because there is no review before the act, only after. That is the entire point of giving a system autonomy, and it is also the entire cost of getting the intent wrong. The old guardrails were designed to stop a human from making a mistake at the moment they were about to make it. The new guardrails have to stop a system from completing an action in the seconds before anyone is in a position to notice.

What this means for anyone about to grant a system authority to act rather than just to recommend is that the test before launch has to change. The question is not "how accurate was the model on held-out data." The question is "show me every action this system is now allowed to take, and show me the written constraint it checks before taking each one, and show me who wrote that constraint and whether they actually hold the authority the model is now exercising on their behalf." If a vendor cannot answer that for every action, not every prediction, the system is not ready to act unsupervised, whatever the accuracy numbers say. The accuracy was never the thing that was going to fail. The sentence nobody wrote down was.

Sources

Bringing predictive analytics to the agentic AI era (MIT Technology Review, 2026-10-05)

Rwanda’s Entelon is building AI for organisations still running on paper (Disrupt Africa, 2026-10-09)

YC-backed Bujeti launches AI “teammates” to take manual work off African finance teams (Disrupt Africa, 2026-10-08)

How Pakistan’s KangarooCare is using AI to improve neonatal care (Disrupt Africa, 2026-10-09)