AI Strategy & Transformation5 min read·By , Independent AI Advisor & Fractional AI Lead·

Your AI roadmap is probably missing the only column that matters

An AI roadmap can track every milestone and still fail to tell a company when to stop investing.

A roadmap is good at telling an organization what happens next. It is less reliable at explaining why the next thing deserves to happen at all.

Consider a familiar planning exercise. A team lists the AI opportunities it wants to pursue. Each gets an owner, a budget, a delivery date and perhaps a confidence rating. The portfolio looks manageable. Someone can report progress every month.

Then the first ends. Its output is promising. The team wants to expand it. The roadmap tells everyone which quarter expansion belongs to. It may not tell them what evidence would make expansion a better decision than stopping.

That is the missing column I would look for: What would change our decision?

This is not another score assigned to a use case. It is a commitment made before the result is known. What outcome are we trying to improve? What happens today without the system? Which costs count? Who is allowed to say that the experiment did not work? And what finding would justify more investment?

An AI roadmap without these answers can become a record of activity rather than a means of allocating resources.

A working demo is not a business result

Technical matters. A system that produces unreliable answers, mishandles sensitive information or fails under ordinary conditions should not pass into production. But technical performance is only one part of the decision.

Suppose a team demonstrates that a model can draft a customer-support response in seconds. That observation does not yet establish that customers receive better service, that employees finish their work sooner, or that the company spends less money. The draft may require review. An may take longer than the original process. The model may work on the easy cases while pushing the difficult ones back onto the team.

Those are possibilities to test, not reasons to assume the technology fails.

Research publicized by MIT Sloan in September 2026 describes the substantial, often invisible work employees do to turn AI experiments into organizational solutions: trying approaches, checking outputs with colleagues and adapting the system as models change. The researchers conducted a two-year field study in two organizations. Their findings are specific to those settings, not a universal AI success rate. But the investment lesson is practical: if experimentation, review and adaptation are absent from the cost and , the apparent success of a pilot can mislead. MIT Sloan research.

The measurement question belongs in the roadmap before the pilot starts, not in a celebratory report afterward.

Make the decision rule visible

For each initiative, I would write a short statement that can be challenged:

We will continue this initiative if, against the existing process, it improves the agreed outcome by a meaningful amount without exceeding our limits for quality, risk, total cost and human oversight. Otherwise we will change the design or stop.

The specific outcome depends on the work. A support operation may care about resolved issues and recontact rates. A research team may care about whether its findings change a decision. A creative team may care about usable approved work rather than the number of generated alternatives.

There is no honest universal metric. That is precisely why the decision rule has to be written by the people accountable for the work.

A useful record separates the baseline, the expected change, the method of comparison, the total cost of operation, the unacceptable failure conditions, the decision owner and the date of review. If a threshold is uncertain, say so and define what the exploratory stage must learn before a scale decision is possible.

This does not mean every experiment needs a financial forecast. Exploration can be valuable because it reduces uncertainty. But even an exploratory initiative needs a decision at its end. What uncertainty did we resolve? What did it cost to learn? What would make a second experiment worthwhile?

The cost that arrives after the pilot

The economics of AI systems are not fixed at the moment the first output appears. Monitoring, exception handling, security, integration, vendor dependence and staff capability all affect what the organization can sustain.

An October 2026 conceptual paper in the Journal of the Academy of Marketing Science proposes a Return on AI framework for marketing organizations that considers incremental benefits, value retained by the firm, durability of evidence and full costs over time. It also raises the possibility that delegating judgment may weaken the human expertise required to oversee the system. Its seven predictions are propositions for future testing, not observed causal effects in every organization. Kanuri, Pattabhiramaiah and Sridhar, 2026.

The distinction matters. The point is not that AI makes people less capable by default. It is that transferring work to a system also changes where people practice judgment and where responsibility sits. That change deserves to be evaluated rather than assumed away.

A roadmap can accommodate this without becoming a compliance manual. It needs an accountable owner who can revisit the economics when the operating conditions change.

A roadmap should make stopping possible

Organizations often treat a stopped initiative as a failure of delivery. It may instead be evidence that the investment process worked.

A small experiment that rules out a costly implementation has done useful work. A pilot that uncovers a hidden dependency has improved the next decision. And a system that works technically but creates too much review overhead may be worth redesigning rather than scaling.

There is an important counterpoint. In early exploration, demanding a precise return before anyone understands the opportunity can suppress valuable learning. The answer is not to pretend that uncertain benefits are certain. It is to distinguish an experiment designed to learn from a deployment expected to perform, with different evidence and funding decisions for each.

The best roadmap is not the one with the most completed milestones. It is the one that helps an organization spend its next unit of attention, money and trust more intelligently.

Before adding another AI project, add one column to the ones already there:

What evidence would make us change course?

Where this connects

This is the portfolio-level companion to How to tell if an AI pilot deserves to survive and Most AI strategies are software procurement plans with better typography. A roadmap should make room for the possibility that the best decision is not to deploy.

Get new essays as they publish.

One email per essay. No noise between.

Berk Bayri

Independent AI Advisor & Fractional AI Lead

I help leadership teams decide where AI belongs, test it before it scales and build the teams that run it, drawing on 25+ years of building digital products, experiences and capabilities.