Defensibility in the AI Era
In the age of large language models, defensibility now lives where proprietary workflow data, built-in growth loops, and rapid experimentation intersect.
Defensibility via Data, Distribution, and Iteration
Over the past decade, software defensibility has largely followed a predictable playbook: build something unique, move faster than incumbents, and defend your wedge with proprietary tech or data. But the rise of large language models (LLMs) has upended much of that logic.
Today, we're witnessing the most intense competitive dynamics we’ve seen in over a decade of backing software founders. The landscape is shifting as companies scramble to build end-to-end functionality, drawn by the potential of AI to unlock value through seamless context-sharing across tasks. Meanwhile, the time and cost to replicate features has collapsed, and traditional distribution moats are crumbling as product discovery moves from search and research to a handful of dominant chat interfaces.
In this new reality, defensibility must be rethought. At Defiant, we’ve come to believe that it now hinges on three interlocking forces: data, distribution, and iteration. These aren't new ideas - but LLMs change how they work, where they show up, and how they compound.
Data: Context Is the Key
LLMs are generalists by design. They’ve been trained on vast swaths of public data, which makes them remarkably good out of the box. But what really moves the needle - what creates product-quality output - is context.
Context is the proprietary, real-time information (users’ history, their workflows, domain-specific content) that you inject into a foundational model so it becomes specific, accurate, and actionable. For those building atop LLMs, what separates one product from another is the quality and fidelity of their context.
This is one of the defining shifts of the LLM era: borrowed capabilities (those within foundational models) produce a significant portion of end-user value, which compresses the moats for all the companies building downstream. This means that the founders who control unique data pipelines - especially those embedded inside vertical workflows - aren’t just improving output; they’re building one of the few remaining moats. Every user interaction, every edge case, adds context to the moat.
Crucially, the ability to share context across tasks within a broader workflow unlocks even more value. That’s why startups and incumbents alike are expanding their product surface to own more of the workflow - extending horizontally or vertically not just to add features, but to deepen context and increase lock-in.
Take Metaview, for example. Their core product - AI-generated interview notes - relies entirely on LLMs, which exposes them to the risk of being cannibalised by the foundational models. But by embedding themselves in the interview process and accumulating rich, structured context across transcripts, roles, and hiring team preferences, they’re building a proprietary data layer that the models don’t have. That layer becomes both a product advantage and a moat.
The same dynamic is playing out across verticals - from legal tools that ingest case history to medical platforms fine-tuned on clinical workflows. The winners won’t be those with the best prompt - they’ll be those with the best context.
Distribution: Where the Battlefront Has Shifted
How software is discovered, adopted, and distributed is transforming. Many of the traditional GTM engines - especially content-driven acquisition strategies like SEO - are weakening in the face of LLM-native interfaces.
Users no longer need to search the web or scroll through pages of product reviews to find solutions. Increasingly, they just ask their AI assistant. As more user intent is intercepted at the model layer - whether it’s ChatGPT, Gemini, or a system-level copilot - the value of being “top of funnel” decreases. And the risk of being disintermediated increases.
The control points of software distribution are consolidating - and shifting in unpredictable ways. A small group of hyperscalers now own the foundational models, giving them not just influence over how users discover tools, but growing control over how those tools are used. In an agentic world, where assistants act on behalf of users, these platforms risk becoming both the first and final stop in the user journey. They don't just sit at the top of the funnel - they are the funnel.
At the same time as distribution moats are eroding and the funnel is consolidating, distribution is also increasing in importance. The software development life cycle has been dramatically compressed - thanks to LLMs, new applications can be built faster and cheaper than ever. In this environment, where building is easy and differentiation is fleeting, the hardest problem becomes getting in front of users. This means that those that achieve cut-through in distribution are at a significant advantage.
Some early-stage companies are adapting. Harvey, for example, has focused on deep partnerships with law firms and tightly integrated product experiences, ensuring they don’t get cut out by general-purpose AI tools. Cody AI is another case: by embedding directly inside developer workflows and integrating tightly with repo tools, they’re growing through usage - not through search.
Of course, the best products have always treated distribution as a design problem - not just a marketing function. What’s changing now is the urgency and centrality of that mindset. In an ecosystem mediated by LLM interfaces and increasingly system-level agents, distribution must be built into the product’s core loops from day one. Teams that design for virality, leverage existing ecosystems, or tap into community-driven growth are pulling ahead of those relying on traditional content or paid acquisition tactics. In the LLM era, designing for distribution is more important than ever.
Iteration: Fast Isn’t Good Enough
If LLMs have lowered the floor on how quickly you can build, they’ve also raised the bar on how fast you need to move. Anyone can ship a v1. The challenge is staying ahead once you do.
Foundational models are evolving rapidly, with new capabilities and API changes arriving almost monthly. Model providers are increasingly integrating “app-like” functionality natively, moving up the stack and threatening to subsume entire products. In this environment, shipping is no longer a competitive edge - learning is.
What matters now is how fast your team can understand user needs, test new ideas, and adapt to platform shifts. Organizational learning speed is the new defensibility. Teams need to be built not just to code quickly, but to experiment constantly - on prompts, data pipelines, UX surfaces, and beyond.
Companies like Langfuse show what this looks like in practice. By leaning into open source, they’ve not only expanded their community but also accelerated their learning cycle - external developers are constantly testing, extending, and feeding back into the product. Others like Dust are giving teams the infrastructure to iterate LLM-based workflows rapidly, making fast feedback a first-class product goal.
What This Means for Founders
So how do you build in this new world?
It starts with data. You want to be the system of record, not the thin interface. If you’re not already collecting unique context, design for it now. Every bit of telemetry you own compounds. Where possible, own more of the workflow so that you can unlock more value for users via context sharing.
Next, distribution needs to be embedded from day one. Viral mechanics, integrations, usage-based loops - these are no longer growth hacks; they’re survival strategies. If you’re reliant on users finding you through traditional means, you risk acquisition hitting a wall.
Finally, design your organization to move fast. Structure teams for experimentation. Build in public if it helps. Embrace the fact that what works today won’t work tomorrow - and make that your advantage.
What we are looking for
In the LLM era, defensibility is harder than ever. We’re looking for early-stage teams that sit at the heart of a workflow, continuously generate proprietary data that makes their product measurably smarter, and can protect that data with credible, defensible pipelines. They must own distribution - serving as the first stop for a critical job and growing through built-in loops rather than channels LLMs can disintermediate. Finally, we back founders with deep domain fluency who run fast, instrument learning cycles and aren’t afraid to cannibalise yesterday’s win to stay ahead tomorrow. To these ends, we are asking the following questions of our founders:
| Category | Screening question | Green flag | Amber flag |
|---|---|---|---|
| Data | Does proprietary data materially boost model performance over the foundation model? | Clear lift shown in pilots or A/B tests; metrics improve as proprietary data volume rises. | Little or no delta vs. base model; performance plateaus quickly. |
| Is there a credible pipeline to acquire and ring-fence that proprietary data? | Data generated as a by-product of core workflow, reinforced by network effects or exclusivity. | Data gathering depends on one-off deals, scraping, or easily replicated integrations. | |
| Will user-specific data deepen lock-in over time? | Product gets smarter for each user/org the longer it’s used; switching costs rise. | Output quality stays flat regardless of historical usage; customers can port data easily. | |
| Does the product sit centrally in the workflow, capturing expansive context? | Inserted early and often (system of record / daily cockpit); high context coverage. | Tack-on feature, or point solution used sporadically; peripheral to the main flow of work. | |
| Distribution | Is the product a first-stop destination for getting a job done, not just a referral endpoint? | Users open it proactively; strong daily/weekly active ratios. | Traffic relies on search results, marketplace listings, or other aggregators. |
| Do built-in growth loops drive adoption independent of search or research? | Collaboration, viral sharing, or embedded output exposes new users organically. | Growth plan depends almost entirely on paid acquisition or SEO. | |
| Does the team have a distribution strategy that survives the shift to AI chat interfaces? | Owns proprietary channels (e.g., workflow integrations, partnerships, community). | Betting on query-based discovery that LLMs can intercept or disintermediate. | |
| Iteration | Does the founder have lived experience in the domain they’re serving? | Founder can cite personal pain points, speaks customers’ language, identifies edge cases. | No significant experience in domain; learning proxy via surveys or consultants. |
| Is the org set up for constant learning (tight build-measure-learn loop)? | Weekly ship cadence, user interviews baked into sprint cycle, data-driven decisions. | Long release cycles, siloed research, little direct user contact. | |
| Do they show an anxiousness to stay ahead and to disrupt themselves? | Roadmap includes experiments that could obsolete v1; culture celebrates kill-switching weak features. | Defensive posture around legacy code; “don’t rock the boat” attitude. | |
| Can the team rapidly fine-tune or re-prompt models as data/needs evolve? | In-house ML ops, automated retraining, sandbox for prompt/version testing. | Reliant on external vendors or manual processes; updates ship quarterly at best. |
Defensibility Reimagined
The traditional moats - tech, brand, even data alone - no longer hold up on their own. In the LLM era, defensibility lives where data, distribution, and iteration intersect.
Companies that own proprietary context, reach users through durable and flexible channels, and out-learn their competitors with each release are the ones that will win - not just today, but as the platform layer continues to shift beneath us.
For founders, the challenge is stark - but so is the opportunity. We’re entering a phase of software where execution speed, contextual leverage, and structural advantage converge. Those who understand that early will be the ones still standing when the next wave hits.
Thank you to Brian Balfour and Jamin Ball for influencing our thinking with respect to moats and defensibility in the current environment.
