Executive Summary
AI does not make a product sticky. Repeat value does.
This article draws on a 37-minute conversation between Karthik I. Gunasekaran, Vice President of AI and Software Products at the Project Management Institute (PMI), and Net Solutions CTO Joseph Jude. Their examples from SaaS, education, and ecommerce show how to tell activity from meaningful engagement, define success before building, and use the seven patterns of AI to find experiences worth testing. The on-demand webinar links to the webinar page carries the full conversation.
- Define the value event before design. A click, a session, or a minute matters only when it shows the user reached the outcome the feature exists for.
- Product-market fit comes first. Engagement cannot rescue a product that does not solve a recurring user problem.
- Name the return trigger. Retention improves when the product meets the user’s next real moment of need. Teams reach for a notification schedule when they haven’t found one.
- Start with the problem, then the AI. A model capability looking for a home becomes a feature nobody asked for.
- Use the seven patterns against one real problem. The chosen pattern tells you which data, instrumentation, and controls the test needs.
The Engagement Metric That Lied
An app’s analytics showed a healthy jump in repeat opens. Investigation traced many of those sessions to accidental taps on an Instagram ad. The app launched for a few seconds, then the user left. Every number on the dashboard was real, and none of it was valuable.
Karthik’s example captures the central problem with engagement programs. Digital products record sessions, clicks, time on screen, feature use, and funnel steps. None of those signals, on its own, says whether the experience worked. Someone who spends ten minutes on a screen may be absorbed. They may also be stuck.
AI makes the mistake easier to scale. It can produce more recommendations, messages, nudges, and rewards. All of them can raise opens or time spent without improving the outcome that brought the user in.
A more useful unit of measurement is the value event: an observable action that shows the user reached the outcome a feature was built to produce. For a language learner, that could be completing a speaking scenario against a defined objective. For a parts buyer, it could be confirming fitment and completing checkout with a compatible part. What matters is defining the event before designing the feature.
Product-Market Fit Still Comes First
Once activity and value are separated, the next question is whether the product creates enough value to be worth engaging with.
Joseph opens the conversation by asking whether teams should think about engagement at the MVP stage or wait until growth. Karthik says teams shouldn’t tune for engagement and retention until they have evidence the product solves a real user problem. Are we solving a real user problem? How deeply? Until we have answers to those two questions, the metrics are premature.
Fit is also not permanent. Expectations move, competitors improve, and the feature that won the first cohort is table stakes for the next. Product teams keep earning fit, and engagement and retention help them see whether they still have it.
Karthik describes early users as including plenty of “app tourists”: people who arrive, explore, and leave. That is fine. A broad early audience lets you learn which users have a recurring problem, what they do when they find value, and what separates them from the people who leave.
Engagement and retention signal product-market fit. Personalization and gamification layered on a weak proposition create more activity around it and nothing else.
Define the Value Event Before Designing the Feature
Product teams often add analytics after a feature ships. The sequence should run the other way. Before design begins, name the user outcome, the value event that will show it happened, and the next natural moment when the user would need the product again.
Three questions to swap. Replace “How much time are users spending?” with “What are they accomplishing in that time?” Replace “How many people used this feature?” with “Did the feature get them to the intended outcome?” Replace “How do we increase engagement?” with “Which behaviors come before repeat value, conversion, or retention?”
Karthik’s rule is blunt: if a team cannot measure a feature, it should not build it. The working test is whether the team can explain how it will know the feature worked. If it cannot, the idea is not ready for the roadmap.
Quantitative data shows where users go, what they use, and where they stop. Qualitative investigation explains why. An unexpected pattern may need journey analysis, session-level evidence, or a conversation with users. The Instagram case would never have been solved with the aggregate dashboard alone.
Our view comes from the products we inherited, updated, and enhanced. Most engagement dashboards fail for a duller reason than the wrong metric. Instrumentation was bolted on after launch, so the events that would answer the real question were never captured. Retrofitting missing events creates avoidable engineering work. Leaving them missing weakens the roadmap decisions that follow.
Build around a defined problem and instrument the value event from the start. Once the feature is live, observe what users do, investigate the surprises, and feed the evidence back into the product. Shipping is when the product starts producing evidence.
AI Should Solve a Problem Users Already Have
Duolingo is the example Karthik uses. Learning a language requires practice. A digital product can teach vocabulary, grammar, and concepts. Learners still need someone to speak with. AI creates a way to meet that need when a human practice partner is unavailable.
The point is the sequence: user problem, experience opportunity, then AI-enabled solution. Beginning with a model capability and hunting for somewhere to place it reverses the logic. The result may demo well and still solve nothing.
There is also a commercial reason to be disciplined. Generative AI adds a running inference cost, or in Karthik’s words, every token costs money. A chatbot or recommendation engine that does not improve a measurable user or business outcome is a cost line with a demo attached.
Gamification Should Reward Progress
Streaks, leaderboards, and progress markers help users build momentum. AI can make those mechanisms more personal. It can also make them better at driving the wrong behavior. In education, a daily return only matters if the person is learning. In health or financial products, more time in the product may be the opposite of success. The standard is whether the mechanism helps the user reach the outcome that brought them there. Where the line between a useful nudge and manipulation sits deserves its own article.
Retention Is Repeated Value Across the Lifecycle
A value event explains why a session mattered. A return trigger explains why a user comes back: the next natural moment when the product can solve another problem in their routine, job, or lifecycle. For a buyer, that is the expected run-out date of daily replenishment items like soap or lotion. For a student, it is advancing from high school test prep to college entrance and professional licensing exams within a unified platform. For a B2B user, it is an upcoming corporate evaluation deadline where workplace bonuses depend on course completion. The trigger comes from the user’s real-world workflow. The notification calendar is what teams reach for when they can’t find one.
“The product is just one touchpoint. Retention is an orchestrated effort across the lifecycle.”
Joseph Jude, CTO, Net Solutions
Lifecycle email, push notifications, subscription nudges, and in-product messaging can run as one system. AI can personalize and time those interactions. The purpose stays the same: bring the user back when another useful outcome is available.
Karthik uses Amazon to show how the relationship extends beyond a single purchase. Prime creates an ongoing relationship, subscriptions turn repeat purchases into predictable behavior, and the wider ecosystem gives customers more than one reason to stay. The lesson is a question rather than a feature: where can the product create enough recurring value that returning becomes natural?
Retention measurement should distinguish a user who returns and reaches the value event from one who opens the product again. Supporting measures vary by product: task success, conversion, expansion, reduced churn. The metric stays tied to the job the user came to do.
AI Personalization Needs One View of the User
Personalization and cross-channel retention break down when the same person appears as several unrelated records. There is the web account, the app install, the email subscriber, the support ticket, and the identity inherited from an acquired product.
Karthik describes the desired state as a customer data cube. One axis carries identity. The others carry interactions across products, devices, and time. Stitched together, fragmented journeys become one story, and the product can respond with context instead of treating every channel as a new relationship.
Every AI pattern in the next section runs on data. If what you know about a user is split across five systems, hyper-personalization works from an incomplete view.
This is where most of the engineering work sits, and where we spend most of our time with clients. The AI feature is rarely the hard part. Identity resolution, the event taxonomy, consent, and the pipeline that makes a signal from Tuesday available to a model on Wednesday decide whether the experience can work. Greenfield products can design that foundation from the beginning. Products with history need the data investment first.
Seven Patterns for Finding Useful AI Opportunities
Toward the end of the conversation, Karthik introduces the Seven Patterns of AI. Cognilytica developed the framework, and it now forms part of PMI’s CPMAI methodology after PMI acquired Cognilytica in September 2024. Rather than starting from individual tools, a product team explores one real problem through seven recurring patterns.
Use the Patterns as Prompts
Do not start the workshop by asking “What can we do with GenAI?” Take one customer problem and ask whether one of the seven patterns could produce a materially better outcome. Karthik’s method is to pick a pattern, run a focused design sprint, and come out with the few features most likely to move a meaningful engagement or retention measure. Then the next pattern.
A pattern is a lens. It narrows four questions the team still has to answer: What problem would this solve? Do we have the required data? Would the resulting experience be better? How would we measure it?
Here is how a feature brief changes when you apply those questions. The example is illustrative, a language-learning product rather than a client, and the pattern is conversational and human interaction.
Once you choose the problem and pattern, the technical foundation becomes clearer. Hyperpersonalization usually requires unified identity and consented behavioral data. Domain-specific recognition depends on reliable inputs and may require labeled examples. Goal-driven or autonomous systems that act through enterprise workflows need governed access to tools, policies, and business rules. Every production use case needs instrumentation and evaluation, with permissions and controls proportionate to its risk. Build the foundation the selected use case requires, and no more.
Five Questions Before Approving an AI Feature
- What recurring user problem are we solving?
- What observable value event will show that the experience worked?
- What will bring the user back without a notification doing the work?
- Which of the seven patterns materially improves the outcome?
- What data, instrumentation, permissions, and controls does the test require?
These five turn a product vision into a testable proposition. A concept that cannot answer them is still an AI capability looking for somewhere to live. The examples behind each answer, and the exchanges between Karthik and Joseph, are in the recording.
FAQs
Behavior that shows users are getting value from the product: completing a useful task, making progress, converting, or returning for another outcome. Raw activity such as sessions or time spent does not qualify on its own.
An observable action that shows the user reached the outcome a product or feature was built to produce. It connects behavior to user value, which makes it more useful than a click, a session, or a minute on its own.
A unified view of one user across every product, device, channel, and time period. A person who appears under several identities in separate systems is treated as one continuous relationship. Personalization and cross-channel retention depend on that foundation.
Hyperpersonalization, conversational and human interaction, recognition, pattern and anomaly detection, predictive analytics and decision support, goal-driven systems, and autonomous systems. Cognilytica developed the framework, and it now forms part of the Project Management Institute’s CPMAI methodology.
No. AI belongs where it solves a problem users already have, and the improvement can be measured. Generative AI carries a running cost on every request, so a feature that cannot show a measurable return is an expense with a demo attached.
Is Your Product Ready for the AI Features on Your Roadmap?
AI features fail when the data, instrumentation, or system access behind them is not ready. The assessment examines your data foundation and the systems around it against the use cases on your roadmap, and shows what to build first and what to fix before you build.
