How did we get here? Not a day goes by without LinkedIn telling us AI is great, AI is bad, we should ignore it, augment it, use it, it can replace humans, it can’t replace humans. And on and on.
I realised recently when speaking to a client that so many organisations or leaders are caught up on the AI hype machine ‘We need AI features’, etc. But whilst AI has been around since the 50s many people don’t actually really know what it is – we’ve had symbolic systems, machine learning, and neural networks in some form for decades. But something changed in the 2020s.
For the first time, four forces converged: model breakthroughs (like transformers), scalable cloud compute, massive training data, and human-friendly interfaces. This perfect storm turned AI from a research discipline into a product-level reality — fast, visible, and valuable. And that’s what consumers have been seeing in products and of course social media.
Now, every product team is under pressure to “add AI.” But adoption doesn’t equal impact. I wanted to explore how to turn AI strategy into real ROI — not just proof-of-concept demos.
Over the past year, AI has moved from exploration to expectation. Boards are asking what the AI strategy is. Customers are being promised smarter, faster, more personalised experiences (With varying levels of success – Siri anyone?) Internally, teams are under pressure to “add AI” somewhere in the roadmap.
In many companies, the results have been underwhelming.
Features are launched, demos are run — but the impact is often limited. The model works, but usage drops off. Or adoption is high, but the business value is hard to measure. Or, in some cases, the feature never gets beyond internal experimentation.
I built a practical framework to help product teams move beyond AI as a “nice to have” and build features that deliver clear, measurable value.
[A quick tip – for founders or companies wanting to know what AI really is and why it has become so pertinent ‘now’ the crib sheet below might help!]
1. Start With Real Friction, Not a Model
Begin with user or operational pain points — not with a dataset or an algorithm.
Look for:
- Time-consuming manual tasks
- High drop-off areas in core flows
- Repeated user queries
- Delays in internal processes
- Areas where human input is predictable or repetitive
Example:
At Notion, AI wasn’t added as a separate interface. Instead, it was embedded into core editing actions — summarising notes, generating text, translating content. The result: faster task completion, improved retention, and a better user experience — all without changing how users work.
Example:
In a healthtech platform I supported, symptom entry was leading to high abandonment rates. We applied a lightweight AI layer to pre-fill likely answers based on partial input and history. It wasn’t technically complex, but it improved form completion by over 20%.
2. Ensure a Closed Data Loop
To be effective, an AI feature needs a feedback loop — not just raw data. This includes:
- A way to train and refine the model based on user actions
- Clear labels or signals from outcomes
- Sufficient volume and consistency in the data
- Control over noise and edge cases
Spotify built Discover Weekly by not just analysing what users played, but what they skipped, repeated, and saved. Every week, feedback trained the next recommendation cycle. Without that loop, the model wouldn’t have improved — and users would have tuned out.
If you can’t link usage to outcome, and outcome back to model improvement, you won’t get long-term ROI.
3. Prioritise Explainability and Trust
Users and internal stakeholders need to understand what the AI is doing — especially in domains like finance, healthcare, and recruitment.
Good AI features:
- Show why a recommendation is made
- Make confidence levels visible
- Offer user control or overrides
- Provide feedback options (“this wasn’t helpful”)
Duolingo explains why you’re being shown certain words again. Salesforce Einstein includes factors behind lead scores. These small touches reduce friction and increase trust.
Without transparency, adoption drops — even if the model performs well technically.
4. Define and Measure Business Outcomes Early
AI success shouldn’t be defined by usage stats or model accuracy alone. It should link directly to business value.
Examples of meaningful outcomes:
- Time saved per user
- Reduction in support ticket volume
- Conversion uplift in key funnels
- Cost savings per workflow
- Improvements in NPS or CSAT
Intercom, for example, measures how many customer queries are resolved end-to-end by AI — and what that translates to in hours saved and agent cost reduction.
Before you start building, define what success looks like in business terms.
5. Assign Clear Ownership
Too often, AI features sit in limbo: owned by no one, or split between data science and product teams. That rarely works.
A successful AI initiative needs:
- A product owner or strategist who defines the user experience
- A technical lead responsible for model development
- Clear metrics of success tied to broader product goals
- A rollout and iteration plan like any other product feature
LinkedIn’s “Skills Match” feature had a dedicated PM and ML owner who worked together from early prototyping through to scaling. It now supports millions of job matches each month.
AI features are still product features — they require the same discipline and ownership.
Summary
Building successful AI products isn’t about having the most advanced models. It’s about solving real problems, designing for trust, measuring impact, and ensuring the right team is responsible.
If you’re building AI into your roadmap and want to avoid wasted cycles, start with this:
- Target real user friction
- Ensure your data can support a feedback loop
- Prioritise explainability
- Track business outcomes
- Give it a clear owner
AI isn’t magic. It’s just good product practice — applied to modern capabilities.
Founders Guide – What this all means and why AI suddenly comes of age in 2025
1. Model Breakthroughs (like transformers)
These are new kinds of computer “brains” that learn better and faster.
Transformers, for example, are a type of AI model that made tools like ChatGPT possible — they can understand language, patterns, and context far more accurately than older models.
📌 Think of it like going from a basic calculator to a smartphone that can hold a conversation.
2. Scalable Cloud Compute
This just means powerful computers that live in the cloud and can grow with your needs. Instead of buying your own servers, you rent computing power from companies like Amazon or Google.
📌 Imagine you’re renting a supercomputer by the minute, and you can add more power instantly if needed.
3. Massive Training Data
AI learns by looking at examples. The more data it sees, the better it gets.
Thanks to the internet, we now have billions of conversations, photos, documents, and interactions AI can learn from.
📌 It’s like training a chef — the more recipes and dishes they’ve seen, the better they become at creating something new.
4. Human-Friendly Interfaces
These are the tools and apps that make AI feel natural to use. Instead of typing code or giving technical commands, you just chat, click, or drag and drop.
📌 Think of how easy it is to talk to Siri or type into ChatGPT — that’s a human-friendly interface.