Survive the AI Knife-Fight: Building Products That Win

Think back over just 45 days of AI product launches. In Brian Balfour’s opening retrospective, Notion moves into territory occupied by Granola, Glean, and ChatGPT; Figma expands toward Canva, Framer, Illustrator, and Lovable. A familiar product category no longer tells you who your competitors will be.

The overlap continues: Atlassian adds meeting notes, search, and Claude integrations; Anthropic expands into enterprise search and integrations; Google and OpenAI push further into coding. These are Balfour’s competitive interpretations of the launches surrounding the talk, with well-funded startups crowding the same categories.

Disruption can move just as quickly. Balfour describes Chegg as declining more than 90% within months; the accompanying slide presents a share-price decline of 90% over nine months. That is a stock-price illustration, not a revenue decline. He points to Stack Overflow as another early casualty after ChatGPT’s launch, without offering a quantified causal analysis.

Product managers are doing more engineering, engineers more product work, and designers more of both. Those changes affect how work gets done. They do not answer the underlying question: What do I build, and why will it win? That has always been product’s central responsibility, even when project management and agile process have obscured it.

Balfour draws on Shaun Clowes, whom he introduces as Confluent’s chief product officer, formerly MuleSoft’s CPO and Atlassian’s first head of growth. Clowes’s formulation is to find an overlooked question that yields an insight other people are not acting on. The last part matters: an insight shared and pursued by every competitor offers little advantage. An insight becomes strategically useful when it opens a course of action others have missed.

A map of Gettysburg supplies Balfour’s analogy for the competitive environment: your product sits among several forces moving at once. His description of strategy becoming ten times harder is rhetorical, but the sources of pressure are concrete.

• Large incumbents: Microsoft, Google, and Meta have scale and are moving quickly.
• Horizontal AI platforms: ChatGPT and Anthropic absorb major use cases that might otherwise support standalone products.
• Technology shifts: Foundational capabilities change on a monthly cadence rather than a yearly one.
• Startup entry: Balfour describes five or six YC startups entering each category with traction in every cohort.

The strategic task is to find an opening among these forces. Infrastructure choices and organizational roles follow that decision; they cannot substitute for it.

Balfour brings roughly 25 years in startups to this question. As Reforge’s founder and CEO, he recalls helping launch HubSpot’s CRM about a decade earlier, when entering that crowded category looked implausible. His suggestion that more than half the audience’s companies might now use it is explicitly a guess, not a survey. He considers the present competitive environment substantially more intense.

At the time of the talk, Balfour reports that Reforge had spent roughly ten years serving thousands of product teams and more than 100,000 professionals. Its knowledge base came from more than 400 experts and over 40 expert-led courses. The company then began encoding that expertise into AI agents: Reforge Insights as a product researcher, and Compass as a project manager for lower-value administrative work. Two further agents were planned for that year.

The strategy framework grew out of several months of work with Ravi Mehta, creator of Reforge’s AI Strategy course. Balfour introduces Mehta as a former Tinder CPO with product leadership experience at Facebook, Microsoft, and TripAdvisor. Their starting point is the pair of traps that sit on either side of a useful AI strategy.

Custom models are not necessary to answer what to build, and a copied chatbot does not answer why the product will win. The useful middle is to treat AI capabilities as Lego pieces: assemble the best available components with your product’s data and functionality.

Competitive advantage comes from what is uniquely yours: your data, your functionality, and your understanding of unmet customer needs. Pretrained models, task execution, audio processing, and image processing create powerful possibilities. But competitors can access those same pieces. Using them is part of building an AI product; their availability alone does not distinguish it.

Data supplies the context that lets an AI model produce an output specific to your customer. Useful context can come from several sources:

• Real-time data: Information that may not be incorporated into model training.
• User-specific data: Context about the particular person using the product.
• Domain-specific data: Specialized material, such as legal or healthcare information.
• Human judgment: Curation and reinforcement data that capture what people consider useful.

Combining these categories can create something more distinctive than any one category alone.

The test is marginal value, not volume: how much does this data add beyond what competitors have and what the large models already know? A large repository is not automatically an advantage. Its strategic value depends on the additional context it contributes to the customer’s result.

Functionality determines how the AI behaves inside the product. Specialized workflows, unique algorithms, business rules, and integrations give the model capabilities it would not have on its own. The advantage comes from connecting these pieces into a system, not merely collecting them.

On the data side, distinctive context informs the AI’s understanding and helps generate a distinctive output. Those outputs can then accumulate into a further repository of useful data, supplying context for subsequent work. That is the proposed feedback loop: context improves output, and output expands the context available later.

On the functionality side, the product controls when it invokes AI and how AI actions fit into the user experience. This is orchestration. The interaction also runs in the other direction: through tool use, AI can call functions exposed by the product. The complete diagram joins both directions of this interaction with the data feedback loop.

Granola entered a market already crowded with AI note-takers, including Fathom, Otter, and Fireflies. From a show of hands, Balfour estimates that roughly 40% of the room had tried or used Granola. This is an informal audience estimate, not market share; his assertion that the response would have been zero a year earlier is a retrospective comparison, not a prior measurement.

He also cites about $50 million in funding without specifying whether he means a round or cumulative funding. For context, Granola’s May 2025 funding announcement reports a $43 million Series B and earlier rounds totaling $24.25 million. The strategic question is how Granola attracted attention despite the crowded field.

Balfour characterizes the existing field—including native capabilities in Zoom and Meet—as trying to replace the whole note-taking job. Granola addressed a different need: help the user take better notes while keeping the user involved.

That distinction changes what the product needs to know and how it should interact with its user.

According to Balfour, Granola started with off-the-shelf capabilities, without unique models or custom training. Deepgram supplied transcription, while Anthropic and OpenAI supplied other AI capabilities. The differentiation lay in the assembly of those services around the customer need.

Granola’s input combines the user’s own notes with the generated meeting transcript. AI uses both to produce enhanced notes. The user’s contribution is therefore part of the context, rather than work the product simply eliminates. Over time, those enhanced notes form a repository that supports additional capabilities: chatting across meetings, project workspaces, and downstream actions.

The desktop functionality makes that context available at the moment it is useful. In Balfour’s account, the Mac app detects meeting starts and accesses system sound for transcription. Calendar integration contributes metadata such as attendees. Meeting detection should not be read as automatic recording: Granola’s current documentation describes capture as manually initiated. The historical example’s central mechanism remains the connection between product functionality, context capture, and the growing notes repository.

Does that mean Granola will survive? Balfour explicitly declines to predict it. An initial combination of data, AI, and functionality is only a starting point. The next task is to use those assets to assemble another distinctive combination, then repeat.

Granola’s project and team workspaces illustrate this sequence: the initial layer of meeting notes enables shared context and new use cases. Balfour then describes downstream CRM connections, including HubSpot, and reports seeing an experiment with a company wiki that updates itself. The wiki is an experiment in his account, not an established general release. Each extension draws on what the preceding layer made possible.

Balfour attributes the principle to Jamin Ball of Altimeter Capital: “The real moat is just a sequence of smaller moats stacked together.” Each advantage buys time; execution and evolution determine what happens during that interval. Ball’s quoted contrast is six to twelve months for an earlier moat versus two to three weeks today—a strategic characterization of shrinking protection, not a measured universal duration. The practical implication is to build a sequence of advantages, rather than assume the first differentiated feature will remain defensible.

The framework turns the next product decision into four questions:

1. What customer problem remains unmet?
2. Which available AI capabilities could solve it in a novel way?
3. What proprietary data would power that solution?
4. What functionality can the product give AI so it can deliver the experience?

The customer problem supplies the direction. The other three answers determine how to assemble the product.

Balfour closes by inviting AI engineers to join Reforge, with the team available outside the session and Reforge offered as the follow-up destination. His recruitment pitch includes the claim that engineers could build products with instant distribution to 300,000 people. That invitation places the strategy in a concrete operating context: applying accumulated expertise and an existing audience to the next set of AI products.