When “we need AI” lands on the roadmap: deciding what should change

“We need some AI in the product.”

It’s a request product leaders need to take seriously. It may signal a shift in customer expectations, a new way to solve a difficult problem, or a competitor making progress that deserves attention. Sometimes, though, the request arrives before anyone has established what would improve for the customer or the business.

The roadmap conversation should start by teasing out what’s behind the request. That may not be clear at first, and the reasons can shift as assumptions are challenged and new evidence emerges.

The pressure to adopt AI is a useful prompt to revisit the assumptions behind the roadmap. Changes in priority should follow evidence about customer value, feasibility and economics, including the consequences of delaying other work. The more consequential the commitment in business terms, the stronger that evidence needs to be.

Start with the workflow that needs to improve

An instruction to add AI describes an approach. To evaluate it, we need to understand the work someone is trying to complete and what currently prevents them from doing it well.

Consider a customer onboarding process. Team members may spend too long finding documents, repeatedly entering the same information, waiting for approval or correcting an incomplete hand-off. These are different problems. A tool that accelerates document retrieval may have little effect on overall onboarding time if the real delay is waiting for someone to take ownership. We may simply get the work to the same bottleneck sooner.

Before discussing the interface or model, establish who experiences the problem, how often it occurs, its consequences and what a worthwhile improvement would look like. Observe the workflow as well as asking people about it. A request for a chatbot is evidence of a suggested solution; the underlying difficulty still needs to be understood.

This is established product discipline. Prodify’s guidance on AI product strategy similarly begins with customer outcomes and examines where AI can change the customer journey. The practical challenge is carrying that discipline into the allocation decisions that follow. Prodify: incorporating AI into product strategy

Ask what has changed enough to revisit the plan

Roadmaps are intended to reflect assumptions about what customers need, what can be delivered and what the business can sustain. AI can justify revisiting all three.

A previously impractical task may become feasible. Customers may expect a much shorter time to completion. A service that once required extensive manual effort may become economical to offer to a wider segment. Competitors may be changing the standard against which customers judge the product.

The response needs to be specific. Which assumption has changed? What evidence supports that change? Which part of the customer outcome could improve as a result?

The opposite risk also matters: preserving the existing plan because its evidence is familiar. New opportunities rarely arrive with the same operating history as established products. Requiring certainty before funding any investigation can become a way of avoiding a strategic decision. I’ve seen that reluctance to commit hold back worthwhile opportunities.

I would match the commitment to what is known. Limited evidence may justify a bounded experiment. Stronger evidence may justify changing the delivery plan. A compelling shift in the market may justify a larger strategic investment, provided the assumptions and exposure are explicit.

Compare the ways to improve the outcome

Any comparison should include process changes, conventional automation, purchased capabilities and AI-assisted or agentic approaches where relevant.

An AI option can offer flexibility where inputs vary or the work involves interpreting unstructured material. That benefit needs to be assessed alongside latency, review effort, reliability and operating cost. Anthropic’s engineering guidance recommends starting with the simplest approach that meets the need and adding agentic complexity when the task warrants it. Anthropic: building effective agents

For roadmap purposes, the unit of value should be successful completion of the customer’s work. A faster draft is useful only to the extent that it improves the completed workflow. If reviewing and correcting it absorbs the saved time, the case for reprioritisation becomes weaker.

That does not mean every benefit must become an immediate cash saving. Reduced effort can improve capacity, responsiveness or the customer experience. It may also free people to tackle work that previously went unattended. Be clear about the benefits supporting the initial decision, and track any additional gains as they emerge.

Make the allocation decision visible

Consider this fictional example. A B2B software provider has 20 team-weeks available for a planning period, after routine support and operational reserve. A team-week means one team’s working capacity for one week. This is a simplified allocation across teams over the period, rather than a measure of elapsed delivery time; it assumes the teams have comparable capacity and the relevant skills. Its customers are struggling to complete onboarding. Leadership proposes an AI assistant, while the team already has work planned to improve that journey.

Assume the existing allocation is:

The assistant is a plausible option for the retrieval problem. It might also uncover an opportunity that the current plan underestimates. The team therefore chooses to investigate it without assuming a production launch.

The experiment is funded by deferring two team-weeks of reporting work and two of the original retrieval scope. Those consequences should be visible to stakeholders. The example assumes the retrieval work is divisible and the required skills are available; if either assumption is false, the allocation must change.

The numbers are illustrative. There is no universally correct percentage of a roadmap to allocate to AI. The useful discipline is showing where capacity comes from, what is deferred and why the new opportunity deserves that trade-off.

Decide what evidence would change the next commitment

The experiment needs a decision at its end. “Explore the potential of AI” is too open-ended to guide allocation.

For this scenario, I would compare the current retrieval workflow, the planned conventional improvement and the assistant on representative onboarding tasks. Measure completion time including human review, correctness, unresolved cases and cost per successfully completed task. Include awkward cases and permission boundaries, rather than selecting only the material on which the prototype performs well.

Set the acceptable quality and improvement thresholds before the trial. They should reflect the consequences of an incorrect result and the value of the time saved. Define who judges the outcomes, the amount of evidence needed and when an inconclusive result warrants another test rather than a launch.

There are several legitimate outcomes:

• The assistant adds little value: complete the conventional improvement and retain the learning.
• The assistant helps a bounded task: plan a limited release with clear scope, review and support.
• The assistant materially changes the workflow: reassess a larger part of the roadmap, including hand-offs and user responsibilities.
• The evidence is incomplete: decide whether the expected value of further learning merits another bounded investment.

This gives the AI opportunity a fair test. It also prevents a successful demonstration from silently becoming a delivery commitment.

Plan for the service that follows the demonstration

A production commitment includes integration, permissions, evaluation, support, recovery and ongoing ownership. These requirements belong in the estimate and sequence of work. They affect whether the promised customer outcome can be delivered consistently.

If AI changes who does the work, who checks it or who owns the outcome, the roadmap may need to include changes to roles, hand-offs and service delivery. Those changes need discovery and planning too.

Some enabling work will remain useful across several options. Reliable data access, clear workflow ownership and reusable evaluation cases may support both an AI approach and a conventional one. That can make them sensible early investments, but their value still needs to be explained; “foundation” should not become a label that exempts work from scrutiny.

The plan should also state when the decision will be reconsidered. A model change, a different operating cost, a new customer requirement or disappointing task performance may alter the preferred approach. Reopening the decision should be tied to those material changes rather than every new announcement.

A decision record for the next roadmap discussion

A product leader’s response to “we need AI” should make the opportunity easier to evaluate. That means being willing to change direction when the evidence supports it, and being precise about the customer outcome and trade-offs involved.

The test of the revised roadmap is whether it gives the organisation a better route to outcomes worth delivering.