If you own a product roadmap, deciding which generative AI feature to add to your product is not about being an innovator or a laggard in technology adoption. It is not about accepting or refusing what AI can or cannot do. And it is not about whether that feature should be built in-house or integrated via APIs.

The decision is more strategic. It is about recognizing how a new AI feature fits your existing value proposition, how it protects and consolidates your competitive advantage, and how it can help your product generate more revenue or secure new funding. On a more tactical level, the decision also depends on how long it takes to bring features to market, whether the timing is right for the team to take on something new, and what to do with what is already on the roadmap.

This post gives you a framework to answer those questions, step by step. Before digging into the process, this post describes what makes the decision on whether to add an AI feature to your product so difficult. It also explains the 3 different types of generative AI features that are usually added to products and what you need before starting the decision process for your next generative AI feature.

Avoiding the “Peanut Butter” Trap

Deciding whether to integrate generative AI is more complex than it appears. If your development team produces a shiny proof of concept, competitors appear to be pulling ahead, and internal stakeholders create a sense of urgency, you may feel the timing is right. But the pressure to launch AI product features can easily lead you to a “rush to conclusion” without sufficient strategic clarity.

Conversely, you may find no clear evidence of customers actually requesting any generative AI feature. This should not be surprising. Most customers are experts in their own needs, not in the technical solutions used to address them.

On one hand you do not want to miss the trend or be the next Blockbuster, disrupted by a Netflix because too slow to innovate. On the other hand, while the opportunity potential of generative AI is high, it comes with two risks that you should consider: productization risk and dilution risk.

Productization risk

Generative AI that works as a proof of concept does not immediately work in the real world. Similar to technology that comes straight from the lab, generative AI requires fine-tuning and product validation before it can be released to customers safely. If you ship too early, you risk damaging the trust you have built with your users.

Dilution risk

Adding generative AI can extend the value provided to users and unlock new monetization opportunities. On the flip side, an all-rounder product risks diluting its value proposition and losing competitive advantage. This happened to Yahoo in the nineties. As the leading search engine at the time, they tried to also become a content provider and a tech platform, losing focus when Google was accelerating. Brad Garlinghouse, a Yahoo senior vice president, recognized the problem early and captured it in a memo: Yahoo was spreading its resources too thinly, like peanut butter on a slice of bread.

Types of generative AI features

Before going into detail on how to find the generative AI feature for your product, let’s clarify what generative AI features actually look like.

Product features using generative AI usually fall into 3 categories: copilot, agent, and creator. The following table describes each category.

Category	What it does	Examples
Copilot	Works alongside users in real-time on tasks like answering questions, summarising data, and providing personalised responses	A marketing copy assistant that suggests edits as you type
Agent	Operates autonomously in the background to handle complex, multi-step workflows according to predefined goals	A tax accounting feature that gathers taxpayer data, checks compliance, requests missing documents, and files returns, all without human intervention
Creator	Generates new content based on knowledge and guidelines provided by the user	An image library, like Freepik, that generates images on demand based on style, format, and text description

Regardless of category, generative AI features may require humans in the loop — one or more domain experts who participate at runtime to audit outputs, or to contribute where human judgement is genuinely needed.

Generative AI features do not need to cover entire processes. In many cases, they automate only the parts of a larger workflow where humans are heavily involved, leaving the rest of the process unchanged.

Prerequisites to discover what AI features to build

If you want to avoid rushing your go-to-market, spreading your peanut butter too thin, or arriving too late to the generative AI party, you need a strategy. Think of it as a game plan: a set of logical choices that address your customer needs, your competition, and ultimately your unique value proposition.

Before defining a strategy for an agentic AI feature, your product should already have a product vision and a value proposition description.

Product vision: An aspirational, long-term statement that describes how your product changes the world or solves a relevant problem. The product vision is the North Star that aligns everyone involved with the product. It explains why some product features are more relevant than others.

Value proposition: A foundational statement that identifies your target customer segments, the specific solutions you offer them, and the unique attributes that differentiate your product from the competition. A product’s value proposition is relevant to a generative AI feature because of its nature. A value proposition built upon “delighting customers with the latest innovation” is strategically opposite to one focused on “providing superior customer service through deep customer relationships”. In the first case, integrating agentic AI is a strategic necessity to maintain market relevance. Conversely, a product focused on customer intimacy requires a more rigorous justification to ensure an AI agent actually enhances the customer experience without diluting the human touch.

5 steps to choose which generative AI feature to add your product

Finding the right generative AI feature for your product is a structured decision that starts from your product vision and value proposition, works through what your customers actually need, and ends with a clear assessment of what is worth building. The steps below give you a repeatable way to decide which AI features should belong to your roadmap.

Step 1 – Define a feature discovery matrix

To ensure discovery stays grounded in real value for customers, internal users, and stakeholders, use a feature discovery matrix to structure your thinking before any feature conversation begins.

The vertical axis lists all the ways a product can create value. For this axis, use the six Value Utility levers from the Buyer Utility Map developed by Chan Kim and Renée Mauborgne.

The horizontal axis is specific to your product. It maps all the steps of the user journey from start to finish. It includes not only touchpoints with your product, but also steps the user performs in between. This helps surface feature ideas that address needs beyond what the product currently covers. Depending on the product, five to eight journey steps are generally sufficient. If you have already developed a user storymap for your product, its backbone (the top layer) is a good starting point for this axis. For example, the following picture shows the 6 utility levers on the y axis and 7 steps of a patient journey applicable to medical booking apppointment product on the x axis.

Step 2 – Discover value creation ideas

If you already have evidence of customers spontaneously requesting generative AI features, start there. For new ideas, go directly to customers through interviews or surveys. In some industries such as healthcare, Key Opinion Leaders — clinicians or researchers with broad visibility across the field — are often more effective than individual user interviews, as they can reflect both practitioner needs and attitudes toward new technology.

A complementary source of ideas is inside your own company. Run a structured brainstorming session with representatives from product, engineering, operations, sales, and customer success. The goal is breadth: the more perspectives in the room, the more of the matrix you are likely to cover.

At this stage, capture each idea as a short ticket answering three questions:

• Who — which customer segment would this feature serve?
• What — what would it do for them?
• Why — what impact would it have?

There are no good or bad ideas at this stage, and feasibility is not on the table yet. The only objective is to populate the matrix as fully as possible. A few practical rules:

• A single cell can hold more than one ticket
• Empty cells are fine
• Similar or overlapping ideas should be consolidated into one ticket

Step 3 – Filter ideas based on value potential and feasibility

At the end of the discovery session, you have all the ideas for generative AI product features on different tickets. It is time to select those two or three ideas that hold the most promise to become product features.

To do this, rank each idea according to two separate dimensions: value potential and implementation effort. This ranking approach is similar to the Attractiveness Map developed by Mark Gruber and Sharon Tal as part of their Market Opportunity Navigator framework. Decoupling value potential from implementation effort is particularly important for AI features, for two reasons. First, AI features tend to give users more flexibility, which widens the range of possible outcomes. The value upside can be larger than expected, but so can the downside. Second, considerations specific to AI, such as data quality, inference guardrails, real-world validation, and post-launch monitoring, complicate the development process and make effort harder to estimate accurately.

Hence, ranking ideas involves assesing implementation effort and value potential independently before choosing the right feature.

implementation effort

At this stage, you do not need to estimate effort in FTEs or story points. T-shirt sizing gives a good enough result without spending too much time on estimation. Effort has two components. The first is capability: does your team have the skills to build this feature, or do you need to acquire them through hiring, training, or an external partner? The second is feasibility: some features are technically or regulatorily more complex than others, regardless of your team’s capability. That complexity introduces uncertainty, and uncertainty drives effort up. As a rule, involve technical leads and legal support early when estimating features with meaningful complexity.

VALUE POTENTIAL

Effort alone is not enough to compare features. You also need to assess the value each one could create for the business. Value potential has two components. The first is how directly the feature addresses an unmet customer need: this should already be visible from the discovery session. The second is market pull: how strongly would the feature attract your existing customers, and could it bring in new ones? You do not need precise numbers here. A three- or five-point qualitative scale is sufficient, calibrated to the number of features you are comparing.

With all ideas mapped across both dimensions, the selection logic is straightforward. Start with low-hanging fruit: features that combine high value potential with low implementation effort. If your shortlist allows, you may also consider quick wins, features with low effort despite moderate value potential.

Step 4 – Assess strategic value

With the most promising features selected, it is time to zoom out. Customer value and implementation effort tell what is worth building, but not whether it is the right move for your product in the long run.

AI features can fundamentally shift how a product is perceived in the market. Rushing into development without understanding the broader strategic implications can be costly.

To assess strategic value, work through the “lenses” described in the following table. As a result, don’t be afraid to modify a feature or drop it altogether. Features with unclear or problematic strategic implications should not be carried forward by default.

Strategic Lens	Key question	What to consider
Competition	What competitive edge does this feature enable, and how will competitors respond?	A classic example for competitors response is Google adding the “AI mode” feature to keep Search users and avoid them migrating to platforms like ChatGPT or Claude to perform web search.
Risk of Substitution	Who else is likely to disrupt you if you don’t develop the feature and bring it to market?	Think about Kodak, which invented the first digital camera in 1975 but failed to go to market to protect its film business. The result was that Nikon, Canon and Sony overtook them.
Cannibalization vs Segmentation	Will this feature cannibalize revenue from existing customers, or create new customer segments?	AI can replace existing services at lower cost, pulling customers away from higher-margin offerings. You may accept this if you plan to retire the old service. The alternative is segmentation: position the AI feature for a distinct customer group your current product doesn’t reach or can’t serve economically. In this way, you can protect existing revenue while expanding your addressable market.
Time to revenue	How long will it take after launch till meaningful revenue materializes?	Even if you develop, validate and ship the feature fast, you may face a slow adoption curve before the feature pays off. With AI features, not all users react in the same way. Some users may be sceptical at first, and it may take time and effort until they commit. This is why both feature UX and proper customer communication help shorten the time to revenue.
Value chain gaps	Where in the customer journey are you leaving space that a new entrant could occupy?	The most dangerous disruptions don’t come from direct competitors, but from new entrants that fill the white space around your core product. For example, medical appointment booking platforms own the scheduling layer but often leave the between-visit journey largely unaddressed. That gap is exactly where startups can operate: supporting patients between visits with AI agents that complement clinicians.
Learning advantage	How can this feature learn from data and generate proprietary knowledge that compounds over time?	Even if you rely on third-party AI providers, you can build a durable moat by capturing interaction data to refine your own proprietary models. “Owning the feedback loop” creates a self-reinforcing unfair advantage: as the model improves through continuous usage, the product becomes increasingly difficult for competitors to replicate.
Return on Investment	What additional revenue will the AI feature bring, and at what cost?	Costs go beyond development, infrastructure, maintenance, and evaluation pipelines. They also include API calls and token volume. For AI features, costs scale directly with usage and often in unpredictable ways. On the revenue side, consider what changes when the feature is available to customers, even if it is free. The main revenue levers typically are: new customers acquired, reduced churn, upsell potential, and operational savings.

Step 5 -Decide on execution

Once you have considered the strategic implications of your selected features, you need to decide how to implement each feature. Consider the decision aspects in the following table as guidelines.

Decision	Key Question	What to consider
Build / Buy / Partner	Should you build internally, acquire, or integrate a third-party solution?	For AI features, this is more a question of time to market than ROI. Building on a third-party API is faster but creates dependency risk. Building proprietary models is more durable but requires capabilities you may not have.
Pricing	What price, if any, will be attached to this feature?	Since every interaction consumes API calls and token volume, your cost structure is inherently variable. A flat subscription model may erode margin quickly with heavy users. Usage-based pricing, or usage thresholds within tiers, lets revenue scale more predictably with cost.
Regulatory compliance	What is the legal and regulatory risk of shipping this feature, and what measures are required?	For AI, relevant frameworks include the EU AI Act, GDPR implications for training data, and liability exposure for automated decisions. Ignoring compliance at this stage risks building a feature that gets blocked before launch, or decommissioned shortly after.
Human in the loop	How will you ensure the AI workflow stays accurate and performs well in production?	Depending on the feature, you may need a domain expert to evaluate AI outputs for accuracy, or a customer service representative to review responses triggered by specific conditions. Take extra care when the AI feature replaces what a human previously did. Users will notice the difference.
Incremental release	How will you release the feature in the safest way?	AI outputs are probabilistic: a feature may perform well in testing and fail unpredictably in production. A bad AI interaction tends to leave a stronger negative impression than a conventional bug. A staged release, starting with selected representative users, limits the potential impact and allows for improving the model before a full release.
Product roadmap	What is the right time to implement this feature given the current roadmap?	Having assessed customer value, strategic fit, and implementation complexity, prioritization is the final decision gate before committing to development. The key input is a picture of what is already underway and what is planned. The output is a business statement on whether, and why, this AI feature is more important or more urgent than what else is on the roadmap.

Conclusion

With new technology such as generative AI, the temptation to add a new feature to your product is high, for good reasons. But before planning new AI features for your product, you should be clear on why you are doing it and what strategic outcomes the feature would unlock. To avoid spreading the value proposition of your existing product too thin, you need a compelling and specific motivation for adding the feature, not just the fear of being left behind.

Every product, market, and feature is different. The framework in this post is designed to apply to most products, but the right answer always depends on your specific context. For a conversation on your specific product, feel free to get in touch.