AI Lessons from a fCPO

Two years on, the LLM revolution is well underway. Here are a few dispatches from the front lines.

As a long-time product guy who has spent his career in the SaaS world, the two years since the public release of ChatGPT have been an exciting time. Having lived through game-changing, assumption-upending tech revolutions before (world wide web, mobile, social media, streaming media, to name just a few), I’m familiar with the idea of a technology concept that suddenly changes everything. The potential opportunities for building new and greater things in ways that had heretofore been inconceivable dazzle the mind. Imagination runs wild. I’m reminded of those old images portraying the Oklahoma Land Rush of 1889 – a bunch of eager people lined up to claim space on a frontier.

Two years later, having served as a fractional product leader for four different companies with AI elements, I’ve learned a few things that go beyond and behind the headlines.

1. AI is Not the Main Character

The value of large language models cannot be denied. Their ability to create results that resembles human intuition and reasoning is astonishing. Their ability to communicate with humans – both inbound and outbound – in conversational, human-readable language brings huge potential.

That said, unless you’re building a large language model yourself, AI is almost certainly PART of your product, not the whole thing. A customer’s needs still matter, the value proposition your product makes still matters, the customer experience still matters, finding product-market-fit still matters. The LLMs can play one – often critical – element of a broader product concept and strategy, but it’s not the whole thing. ChatGPT or Claude or Gemini can deliver open-ended, prompt-driven chatbot interfaces quite well, so any product designed for a narrower or niche purpose needs to be more and better. Nest AI snugly in your broader product and business concept.

2. LLMs are Smart But Dumb

LLMs trained on general knowledge will lag on specialty knowledge for a while. Most amateurs are amazed by how advanced a general-purpose AI seems on any given subject, but the more you know about an area the more you realize how the default responses from LLMs fall short. One simple example is in simple prose writing: ask an LLM to write an essay or a fiction story and the result is notably dull and simplistic. Even with spicy prompting you can tell when something is AI written. In fact, many English teachers are already able to spot AI-written compositions after a few sentences.

So consider the LLM basics – human-readable output, intuitive understanding of instructions – the easy part. The subject matter expertise, customer focus, and overall business execution you supply is the hard part. At a product level, this means using savvy RAG/prompting and memory management to provide tailored results from a large model that customers love.

3. AI and Chatbots are Not The Same Thing

Chatbots are just one form (and a very early form) of the AI user experience. One of the most basic forms of interactions humans have is conversational, so it’s natural that the creators of early AI interfaces gravitated to that as a construct. ChatGPT has “chat” right in the name. And because that’s the AI construct that has received the most attention and traction in the past two years, most people have come to think of a chatbot experience whenever they think about AI.

But a chat experience is actually just a UX built on top of the far more intensive and interesting guts and capabilities of an LLM. LLMs can take input in virtually unlimited ways, and can structure their output as data structures, code, images, further prompts, or any number of other outputs, which can then be used as part of the product concept.

In addition, the chat UX is inefficient for most purposes that aren’t simple conversations. Since ChatGPT’s release, a lot of attention has been paid to “prompt engineering” as a job function, largely because getting the desired outcome from a LLM often takes sophisticated knowledge of how LLMs interpret the requests they receive. In short, what you include in your prompt matters. And the more sophisticated your desired response, the harder it is to consciously think about how to phrase a prompt. The one cardinal rule or user experience dating back to the nineties is “don’t make me think.” Once you get past the surface of customer support, general knowledge queries, and AI therapy sessions, the more the chatbot experience breaks down.

As a product designer, think more broadly about how your startup can wrap LLM responses in more holistic and intuitive experiences: dashboards with controls that generate prompts so people don’t have to type, graphical interfaces that let users tune inputs and outputs, or ways to set (or infer) preferences or reusable guidelines during a session.

4. The Only Constant is Change

It’s clear that AI will get significantly better very quickly. If you’re using RAG or clever prompts to elicit tuned responses from an LLM, count on the fact that the LLMs will get smarter and faster and be able to do more soon. Design your product and your business with those future capabilities in mind. For example, don’t over-invest in a detailed prompt library that is hard to alter or change. Build in the ability to workshop different strategies to achieve your product’s purpose over time, and build structures to constantly be in the same R&D mode you went through during prototyping or MVP development. Stay nimble. Think of AI as an independently-developing component of your product that you should be able to adapt around over time. Otherwise you’ll risk your product stagnating or being eclipsed by competitors or substitutes – or by the big LLM companies themselves.

5. Don’t Fear Big LLM

In many past industries I’ve been involved in (online education, web services, fintech), one constant worry of general-purpose investors and industry outsiders was that Google (and before that Microsoft or Yahoo or AOL) “could step in and own this whole industry overnight.” The theory was that they had so many resources and advantages that they could reproduce my startup’s efforts with the snap of their fingers and put me out of business. I’m hearing the same thing about OpenAI and other Big Tech LLMs.

While not technically untrue, in the past these fears proved almost always baseless. The reason was simple: the larger the Big Tech company, the bigger and more assured the market opportunity has to be to capture their attention. Why bother chasing a hundred $50 million opportunities when you can pursue a single $1B opportunity? This is a corollary of Clayton Christensen’s Innovator’s Dilemma and one that plays out all the time.

Related to this is the perceived risk of from-scratch innovation: if a big tech company does get interested in your product-market-fit combination, for a company with means it’s almost always cheaper to acquire an existing company than create one from scratch. Most would prefer to buy an intact solution rather than build one out. In short, if your space becomes attractive and you’re successful in it, it’s far more likely that you’ll be acquired than replaced.

Don’t worry about whether the general-purpose LLMs will come after your niche application or market. They have plenty of opportunities to create value upstream and plenty of markets where their scale and cost structure will yield a higher return for them. They need an ecosystem of downstream players innovating around their core functionality, and that’s you.

To summarize, AI is a tool in a big toolkit. Work to make it one element of your business and product strategy, but don’t let it limit you.