Every enterprise software company has an AI roadmap now. Most of them are the same document: a list of features where “AI” appears in the title, ordered roughly by how impressive they sound in a board deck.
I prioritise a $100M+ technical SaaS portfolio. I have written that document. It doesn’t survive contact with a CFO, and it shouldn’t.
The problem isn’t ambition. The problem is that AI features break the prioritisation heuristics most product teams have spent a decade internalising.
Why the old filters fail
Traditional feature prioritisation assumes roughly fixed marginal cost. You build a reporting module once; the thousandth customer to use it costs you approximately nothing. Under that assumption, “will users want this?” is very nearly the only question that matters, and frameworks like RICE work fine.
AI features violate that assumption in three ways at once:
- Marginal cost is real and recurring. Every inference has a price. A feature that delights users and gets used constantly can be a margin disaster.
- Quality is probabilistic, not binary. A traditional feature works or has a bug. An AI feature is right 94% of the time, and the value of the remaining 6% depends entirely on what a wrong answer costs the user.
- The cost curve moves under you. Model prices fall, capabilities jump. Something uneconomic this quarter may be obvious next year — which means “no” often needs to be “not yet, and here’s the trigger.”
Prioritising AI work with pre-AI heuristics produces a roadmap of things that demo beautifully and quietly destroy your gross margin.
The four questions
Before an AI item gets roadmap space in my portfolio, it has to answer these. Most don’t.
1. What is the unit of value, and what does it cost to produce?
Not “what does the feature cost to build” — what does one use of it cost, at the volume we expect at scale? If a user triggers this twenty times a day and each trigger is a multi-step agent run, you need that number before you need the mockups.
The answer doesn’t have to be small. It has to be known, and it has to be smaller than what the outcome is worth.
2. What does a wrong answer cost the user?
This is the question that determines whether you need 95% accuracy or 99.9%, and those are wildly different products with wildly different budgets.
An AI feature that drafts a message a human reviews before sending has a cheap failure mode — the human catches it. An AI feature that auto-rejects a job candidate has an expensive one, and in regulated domains it is not merely expensive but legally exposed. Same underlying model, entirely different bar.
If you cannot articulate the cost of being wrong, you cannot specify the feature.
3. Does this deepen the business model, or decorate it?
Some AI features make your existing model more valuable — they increase the switching cost, expand the account, or defend a position competitors can copy in a quarter. Others are decoration: pleasant, demoable, strategically inert.
The test I use: if a competitor shipped this identically tomorrow, what would we have lost? If the answer is “not much,” it’s decoration. Decoration can still be worth building — parity matters — but it should be funded as parity work, not as strategy.
4. Who owns it when it misbehaves?
AI features have a governance surface that traditional features don’t: data provenance, retention, bias exposure, explainability, and an audit trail somebody will eventually ask for.
Driving alignment across Data Science, Engineering, and Legal is not overhead you add at the end. It is a gate at the beginning. A feature with no identified owner for its failure modes is not a prioritised feature; it is a liability with a Jira ticket.
What survives
Run those four questions across an ambitious AI roadmap and it typically loses half its items. That’s not a failure of the roadmap — it’s the filter doing its job. What survives has a property the original list lacked: each item has a defensible answer to “why this, why now, at what cost.”
That is the document a CFO funds. More importantly, it’s the document engineering trusts, because it signals that the person who wrote it understands that shipping an AI feature is a commitment to a recurring cost and an ongoing quality obligation — not a one-time build.
The uncomfortable part
The four questions bias against the flashiest ideas. The AI feature with the best demo is frequently the one with the worst economics, because impressiveness and inference cost are strongly correlated.
Holding that line is a political act. You will be the person in the room arguing against the exciting thing. Do it anyway, and do it with numbers rather than caution — “here’s the cost per user per month at projected volume” is an argument; “I’m worried about scope” is a vibe.
Moving AI products beyond hype and into tangible business value isn’t a slogan. It’s this filter, applied consistently, including on the ideas you personally like.