Why AI Roadmap Features Don't Actually Improve Prioritisation

Generating a ranked list is not the same as reasoning about trade-offs

There is a meaningful difference between an AI that can produce a prioritised list and one that can reason about trade-offs in context. Almost every tool ships the former and calls it the latter.

RICE, MoSCoW, impact/effort — these frameworks are useful when the person applying them understands the strategic constraints behind the numbers. Reach means something different if you are targeting enterprise customers versus SMBs. Impact means something different if your board has told you this is a retention year, not a growth year. Effort means something different when your senior backend engineer just went on parental leave.

An AI that scores your backlog without any of this context is not doing prioritisation. It is doing autocomplete on your data. It produces output that looks plausible enough to get through a weekly planning meeting, but it has no idea what you are actually optimising for.

Shreyas Doshi put it well in a thread on X: "The inputs to good prioritisation are mostly invisible to any tool — they live in the heads of the PM, the CEO, and the three customers who called last week."

That is exactly the problem. The inputs that matter most are not in the tool.

What outcome-oriented prioritisation actually requires

My experience is that the teams who prioritise well do three things no current AI roadmap feature asks for. They define what they are willing to trade away, not just what they want. They make their constraints explicit before scoring anything. And they keep a record of why they said no to the things they declined.

A tool that supported real outcome-oriented prioritisation would need to know your strategic bets for the quarter, your delivery constraints right now, which customer segments you are and are not optimising for, and the reasoning behind your last five deprioritisation decisions. Without those inputs, the AI is working from your backlog data and whatever scoring you have managed to keep up to date. That is a thin foundation.

The usability-capability trap here is real. The tools that have invested most heavily in AI features tend to require the most operational discipline to get any signal out of them. You need clean data, consistent scoring, up-to-date effort estimates. Most teams I have worked with do not have that. So the AI runs on noise and hands you back noise with more confidence.

What accelerating your biases actually looks like

The uncomfortable truth about AI prioritisation is that it does not challenge your thinking. It amplifies whatever patterns already exist in your data.

If your team has historically overweighted requests from your loudest customers, the AI will learn that pattern and surface more of the same. If your effort estimates are systematically optimistic, the model will score on that basis. It is not a second opinion. It is a faster version of your existing judgment, stripped of the moments of doubt that sometimes produce good decisions.

For AI prioritisation to genuinely challenge your reasoning, it would need to be able to say: "You have deprioritised three requests from your mid-market segment this quarter. That segment is your stated growth target. Is this intentional?" That is a different kind of tool. It would need memory, strategic context, and a model of what you said you were trying to do. None of the current tools do this.

What to do instead

I am not arguing you should ignore these tools entirely. But use them for what they are actually good at: reducing the time it takes to get a structured view of a messy backlog. That is genuinely useful. Just do not mistake the output for strategic prioritisation.

Before you run any AI prioritisation feature, write down your three strategic constraints for the quarter. Do it in plain language. Then look at what the AI produces and ask whether those constraints are visible anywhere in the output.

If they are not, you have your answer about what the tool is actually doing.