Intense pressure to integrate AI capabilities pulls product teams away from what actually matters: solving real user problems. The main difference between AI projects that succeed and those that become costly experiments lies in a relentless focus on the end-user.
Tey Bannerman, AI strategy and product design leader, draws on lessons from the field to offer a practical methodology helping product teams build AI products users genuinely trust and adopt.
“The biggest risk in AI is building something impressive that nobody actually wants.” — Tey Bannerman
The problem: when AI is an answer in search of a problem
The majority of AI features fail because they are "answers in search of a problem".
Too often, teams start with the capability of the model, asking what the technology can do, rather than what users need. They may prioritise developing chat functionality or auto-generated summaries because the models allow it, without validating whether these features actually enable a valuable outcome.
“If you start with the model, you often end up solving a problem that no one really has.”
This results in a disconnect between corporate mandates and customer reality. While major corporations like Slack, Microsoft, and even Walmart, are rushing to announce new AI features, this urgency is not always reflected in customer desire. For instance, in 2024, 18% of consumers cited AI integrations as a motivator for upgrading their phone; by 2025, that figure had dropped to 11%.
The lesson is clear: If the technology doesn't clearly resolve an existing, painful friction point, users won't integrate it into their daily workflows, regardless of how advanced the model is.
Lesson 1: Match technology to validated problems
The foundation of successful AI-native product development mirrors great product work: start with a deep understanding of the user and their context. Only then should you determine if AI adds value.
“AI should never be step one. Step one is understanding what people struggle with.” — Tey Bannerman
Look for user friction where AI can help:
- Which tasks take disproportionately long relative to their value? If a user spends 15 minutes searching for a product when the sole value is adding it to a basket, that's a disproportionate time sink.
- Where do users “soft abandon” and come back later? These are friction points annoying enough to cause users to delay a task, even if they don't leave the platform entirely.
- What information do users search for multiple times? Repeated searches indicate the system is failing to meet intent.
- What processes happen outside your product that reference data inside it? Users might be taking data out of your product to complete a task elsewhere (e.g., creating a shopping list) that could potentially be solved within your platform.
Case study: The grocery search problem
Traditional grocery search breaks because it relies on literal keyword matching. When customers used natural language or conceptual queries, results often became irrelevant.
- When users searched for "purple vegetable", they often received irrelevant items like purple-colored shampoo, socks, kitchen strainers, or unrelated products because the search engine only found the word "purple" in non-grocery descriptions.
- Similarly, searches for concepts like "quick breakfast for kids" either returned single, non-contextual items (like a chocolate milk sipper or a single biscuit brand) or nothing at all.
To address this pain, the approach avoided adding a highly visible "AI feature" like a chatbot, which many customers "absolutely hated". Instead, the solution was an invisible intervention:
- Creating an intent layer: The team used Large Language Models (LLMs) to create an intent layer within the existing product structure. This layer's strength lies in its ability to interpret natural language queries, translating the customer’s conceptual intent (e.g., "purple vegetable") into multiple executable database commands against the existing product catalog.
- Seamless categorization: This system allowed the LLM to create seamless categories, like "Purple Vegetables," on the fly. Products like aubergine and red cabbage were pulled into this category based on intent, even though the category did not exist in the hard-coded product catalog.
- Contextual understanding: If a customer searched for "bolognese," the system recognized the intent to cook, providing a category listing the necessary ingredients (mince, chopped tomatoes, spaghetti, etc.).
The key to its success was that customers was unaware that a large language model or GPT was working behind the scenes. The product simply worked better, making it easier for them to find and purchase items, leading to a "massive uplift" in relevant business metrics.
Lesson 2: Understand practical realities
Once a problem is validated, the next step involves understanding what AI can sustainably and cost-effectively achieve.
It is vital to bridge the gap between user value (context, constraints, value) and technical reality (data, performance, maintenance), two worlds that rarely communicate in many organisations.
“Just because a model can do it doesn’t mean you can run it reliably, legally, or affordably.” — Tey Bannerman
What can AI do (and what are the risks)?
Beyond these functional risks, teams must consider Performance and Maintenance. Performance is not just speed, it's cost. If making multiple large language model calls incurs high token costs or creates unacceptable latency (e.g., 10-15 seconds), the solution may not be sustainable at scale.
Similarly, maintenance is a serious concern: keeping AI capabilities updated as new models are launched requires a dedicated team.
Lesson 3: Make AI disappear into the user experience
“The best AI products don’t feel futuristic, they feel obvious.” - Tey Bannerman
Making AI invisible changes how users relate to the product. When AI blends into familiar interactions, people don’t have to assess whether they trust the system, they simply experience a smoother path to value.
Visible AI works differently: when an AI feature stands out (whether through a chatbot or an “Ask AI” button), it disrupts how people normally understand the interface. Because the decision-making is opaque, users begin evaluating whether the system is fair, predictable, and safe. Even if the model is technically accurate, adoption depends entirely on perceived trustworthiness.
“Even a perfectly accurate model will fail if the user doesn’t trust it.” — Tey Bannerman
Invisible AI removes this burden: users focus on the improvement, not the mechanism, and adoption follows naturally.
Teams building for global audiences must take one more layer seriously: cultural intelligence. AI models often reflect the norms of the data they were trained on. A product may support multiple languages yet still feel culturally narrow. Subtle differences in tone, formality, humour, or directness can affect trust in different markets.
Lesson 4: Validate value, not adoption
One of the most useful reframes in the AI product space is the distinction between usage and adoption.
- Usage is simple activity.
- Adoption is when a feature becomes a durable part of a user’s actual workflow or mental model.
The goal should be to create features where users would notice if they disappeared.
To measure true value, shift focus away from vanity metrics like "AI usage" or "engagement" toward measurable outcomes that reflect solved user problems.
Measure what matters:
- Did the feature reduce the time it took users to complete a critical task? (e.g., adding eggs to a basket).
- Did it increase conversion rates for specific cohorts?
- Did it reduce customer service queries related to search friction?
If teams align their measurement of success with the specific, valuable problems they set out to solve for customers, the business metrics will follow.
The Duolingo blueprint
Duolingo is a strong example of a company using AI to strengthen its core value rather than chasing novelty. Their features serve real learning challenges, not generic “AI moments.”
Duolingo Max introduced Explain My Answer to give tailored feedback and Roleplay to help learners practise real dialogue with an AI characteur. Both features integrate seamlessly into how learners already study, and feel like natural extensions of the product.
Duolingo is also transparent about its development process. They publicly share instances where their initial attempts with generative AI failed.
This transparency validates a simple principle: Scale what works, kill what doesn’t. Don't be afraid to decommission an AI feature if it doesn't deliver measurable value, even if significant resources were invested.
Conclusion
Building AI products that users actually adopt is less about technological brilliance and more about disciplined product management. In an era of technical temptation, the core message remains: stay product-led.
Your success will come from combining an empathetic, fresh-eyed understanding of the user’s real challenges (context, constraints, value) with an honest assessment of AI’s practical realities (data, performance, maintenance).
By prioritising clarity over chasing hype, validating value over mere activity, and making the AI subtly disappear into a better user experience, you move AI from a buzzword to a fundamental driver of meaningful outcomes. The goal is not to show that you have AI features, but to demonstrate how much easier you’ve made the user’s life.
