From Prediction to Preparation: Make the Forecast Cheap to Act OnFrom Prediction to Preparation: Make the Forecast Cheap to Act On

I once sat in a quarterly review where a data scientist presented a genuinely impressive churn model. It predicted, with real accuracy, which accounts were most likely to leave in the next ninety days. The room nodded. Someone said "great work." The slide advanced. And then nothing happened. No one was assigned to the at-risk accounts, no play was triggered, no budget moved. A quarter later, most of the accounts the model flagged had churned, exactly as forecast, and we had watched it happen with perfect foresight and zero action. The prediction was correct. It was also useless, because a forecast that does not change a decision is just a more expensive way to be surprised on schedule.

That gap, between knowing what will happen and doing something about it, is the real frontier of AI in 2026, and the market is finally crossing it. The conversation is moving from prediction to preparation: from systems that forecast an outcome to systems that assemble the move, the evidence and the draft, so the decision itself takes minutes. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that a third of enterprise applications will include agentic AI by then. In customer service alone, Gartner expects agentic AI to autonomously resolve 80% of common issues by 2029, cutting operational costs by 30%. As one CIO put it plainly, predicting the future is the easy part. Deciding what to do is where the value, and the difficulty, actually lives.

Prediction Was Never the Bottleneck

We have been good at forecasting for years. We were just bad at acting on it.

The analytics era gave marketing an abundance of prediction: lead scores, propensity models, churn risk, forecast pipelines. The dashboards multiplied, and the foresight got sharper. What never improved at the same rate was the distance between the insight and the action. A prediction lands on a screen, a human has to notice it, interpret it, decide whether to trust it, figure out the right response, and then actually do the work, usually across three other tools. By the time all of that happens, the window has often closed. The bottleneck in AI was never the quality of the forecast. It was the human relay race between the forecast and the move.

Preparation collapses that relay. Instead of telling you an account is at risk, the system brings you the move: the play it would run, the outreach already drafted, the spend change costed, and the reasoning attached. You approve it or you do not. That is the leap that turns analytics from a reporting function into an operating one, and it does not require handing over the decision.

Preparation Is the Step Most Tools Skipped

It helps to be precise about the ladder, because vendors blur it constantly.

Descriptive analytics tells you what happened. Predictive analytics tells you what is likely to happen next. The missing rung is preparation: assembling the recommended move, the evidence behind it and the thing that would be sent, so a person can approve it in minutes instead of reconstructing it over days. Most marketing tooling has lived on the first two rungs for a decade. The reason the top rung stayed empty is not that nobody wanted it. It is that preparing a good move requires something predicting an outcome does not: a deep, current understanding of your specific business. A forecast can be made from patterns in the data. Good preparation has to know your customers, your constraints, your past decisions, and what actually worked the last three times you faced this, or its advice is generic and you should not follow it.

Preparation Without Context Is Confident Guessing

This is exactly where the wave will break for most companies, and the warning signs are already in the data.

Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear value, and inadequate risk controls. Read that next to the autonomy numbers and the real story emerges: the appetite for systems that act is enormous, and the failure rate is about to be enormous too. The projects that fail will not fail because the models are weak. They will fail because a system was asked to decide and act inside a business it did not actually understand. Preparation is only as good as the context behind it. Strip the context away and "what you should do" is just a confident guess with an automation budget attached, which is more dangerous than no recommendation at all, because now it executes.

I have made the version of this argument about output and about returns. In Why Most Companies Are Getting AI ROI Wrong I argued that value shows up when AI changes a decision, not when it accelerates a task. Preparation is that idea taken to its conclusion: the whole point is to change the decision, and the cheapest way to change a decision is to make the right one easy to reach. But you cannot trust a machine to prepare a decision in your business unless it can see your business, which is the thesis I keep returning to in Context Is the Whole Game. Context is what separates work you can act on from work you have to second-guess, and second-guessing every recommendation puts the human relay race right back in the middle.

Action Raises the Stakes, So the Context Has to Be Real

Prediction is low-risk because a human still decides. Keep that part. What changes is how much work the human has to do before deciding.

When a forecast is wrong, you ignore it. When an autonomous system acts on a bad call, it has already moved the budget, sent the email, or reprioritized the queue before anyone reviews it. That is the argument for stopping at the decision line rather than crossing it. The more work a system does in front of the call, the more it matters that the context underneath it is complete, current, and coherent. This is the operating discipline behind agentic marketing that I described in earlier writing on the orchestrator role: the marketer's job shifts from doing the work to judging work that arrives prepared, and you cannot judge what was assembled from a fragmented, out-of-date picture of your own company. The reward for getting it right is real, the 80% resolution and 30% cost numbers are not small, but they are only available to teams whose context is solid enough to act on. Everyone else gets the cancellation statistic.

What to Do Before You Trust the Preparation

You earn the right to autonomy. You do not buy it.

Before you trust what a system puts in front of you, run two checks. First, audit the context it will reason from: can it actually reach your customer history, your past campaign results, your product detail, your real constraints, in a current and usable form, or is it reasoning from a stale slice. If the context is thin, fix that first, because a confident recommendation built on thin context is a liability. Second, start where a wrong move is cheap to undo. Watch whether what it prepares matches what you would have done, and widen its remit only as it earns that. As I argued in What Is Operational Intelligence, the value was always in connected intelligence driving decisions, not in another report. Preparation is where that finally pays off, but only on a context layer rich enough to deserve the trust.

The Takeaway

The era of prediction taught us to forecast beautifully and act slowly. The gap was never the forecast. It was the days of assembly between knowing and doing, and that is the part worth automating: moving from "here is what will happen" to "here is the move, here is what it is based on, and here it is ready to send."

The dividing line is context. Preparation is only as trustworthy as the understanding behind it, and the work in front of a decision is only as safe as the context that assembled it. The teams that win will not be the ones with the boldest automation. They will be the ones who made the right call cheap to make: the evidence gathered, the draft written, the exception flagged, and a person still saying yes or no.

That last part is not a limitation to engineer away. The churn model was right about the accounts. What it could not know was which of those relationships was worth saving, which renewal was already lost for reasons no system had, and which customer would read an outreach as an insult. Keep the judgment where the context actually is. Automate everything in front of it.

Tracy Thayne is the founder of Expona, an AI consulting practice. We find the recurring decisions that bring work into a business and build the preparation in front of them, stopping where judgment starts. Subscribe to the blog (below) for the weekly take on context, AI and the work in front of the decision.