AI product teams still have to answer what to build and why it will win before they worry about implementation speed.
2
Differentiation comes from combining a company's data, functionality, and understanding of unmet customer needs with widely available AI capabilities.
3
Products need to keep stacking new, smaller advantages because the useful life of a competitive moat may now be only two to three weeks.
Summary
Brian Balfour argues that faster prototyping does not solve the hardest product question: what should a company build, and why will it win? The competitive field now includes large incumbents, horizontal AI platforms, rapid technology changes, and many funded startups. He describes two traps: building custom AI infrastructure unnecessarily and copying generic features such as chatbots. His alternative is to treat AI capabilities as Lego blocks and combine them with proprietary data, product functionality, and a clear understanding of unmet customer needs. Granola illustrates the approach. It entered a crowded meeting-notes market by helping people take better notes rather than trying to replace the entire job. Its notes, transcription, integrations, and workflows then created more data and supported further features. Balfour says this process must continue because competitive advantages now expire quickly. Product teams have to keep finding customer problems, assembling new solutions, and extending the system around them.
Product teams still have to decide what deserves to be built
Balfour says the rush to have engineers do product work, product managers write code, and designers build prototypes misses the question that determines whether any of it matters: "What do I build and why will it win?" He presents this as the original responsibility of product leadership, before project management and agile processes took up much of the role. Sean Clowes, Confluent's chief product officer, describes the work as finding an insight that others have not yet acted on. An insight only becomes a competitive advantage when competitors have not acted on it.
The competitive environment has become much harder to read
Balfour compares the current market to a map of Gettysburg, one of the bloodiest battles of the Civil War. Fast-moving incumbents such as Microsoft, Google, and Meta compete with horizontal platforms such as ChatGPT and Anthropic, while the underlying technology changes monthly. He also points to large numbers of funded startups entering categories with traction. Companies can now collapse in months rather than years, as he says happened with Chegg and Stack Overflow after ChatGPT launched. Teams must find a narrow opening among these forces before choosing technology or defining roles.
Teams should avoid both custom AI infrastructure and generic AI features
Balfour identifies two common traps. One is trying to reinvent the AI wheel by building custom models and infrastructure when it is unnecessary. The other is copying basic features such as chatbots into an existing product. He recommends treating AI as a set of Lego blocks. A company can combine available models and capabilities with its own data and product functionality. The resulting advantage comes from what the company uniquely understands and controls, especially its customer problems, data, and workflows.
Proprietary data matters because it gives the model context others do not have
AI capabilities such as pretrained models, audio processing, and image processing are available to everyone. Balfour says differentiation begins with data because data gives a model context that can produce a distinct output. Useful data may be real-time information outside a model's training set, user-specific information, domain-specific information, human curation, or reinforcement data. The amount of data is less important than its marginal value compared with what the large models already contain. A product can then turn its unique outputs into more data, creating a continuing loop.
Product functionality controls how AI behaves in a real product
Functionality is the other major part of Balfour's system. It includes specialized workflows, algorithms, business rules, and integrations already built into the product. These components control when the product calls AI and how the result reaches the user. AI can also call tools inside the product, which gives the model access to product actions. Balfour describes data and functionality as working together: data informs the AI's understanding and creates a unique output, while product functionality controls the AI's actions and shapes the user experience.
Granola found an opening by helping users take notes instead of replacing the whole job
Granola entered a crowded market containing Fathom, Otter, Fireflies, Zoom, Meet, and other AI note-taking products. Those products generally approached the problem as having the system take all of a user's meeting notes. Granola focused on a different need: users who wanted help taking better notes themselves. It used off-the-shelf components, including Deepgram for transcription and Anthropic and OpenAI for other functions. Its distinct result came from assembling those components around the user's own notes, generated transcription, and the specific workflow.
A product can turn its first useful workflow into a data and feature flywheel
Granola combines the user's notes with its transcription to generate enhanced notes. Those notes become a repository that supports chat across meetings, project workspaces, and downstream actions. Its Mac app can detect when a meeting starts and access system sound for transcription at the moment the user needs it. Calendar integration adds meeting metadata such as attendees. Balfour says these choices connect the data and functionality pieces into a system. Granola has extended the initial product with team workspaces, CRM connections such as HubSpot, and experiments with a self-updating company wiki.
Competitive advantage now has to be rebuilt in short sequences
Balfour quotes Jamon Ball of Altimeter Capital saying that the real moat is a sequence of smaller moats stacked together, with each one buying time. Balfour says Granola cannot stop after its first successful combination of data, functionality, and customer insight. It has to use that starting point to create another set of differentiated use cases. He gives a reported contrast between older six-to-twelve-month moats and today's two-to-three-week timeframe. Staying ahead depends on how quickly a team can keep finding and assembling the next set of building blocks.
