The Platform Encroachment Test
Build where the platform's mission says it will never go — the general layer is not yours to own.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 93%
Kilpatrick's decision rule for anyone building on top of a foundation-model provider: the platform will relentlessly build the general capability (reasoning, coding, writing, general-purpose agents) because that is its mission. It will not build verticalized, domain-specific products. Your defensibility comes from domain knowledge, custom tuning, and a differentiated interface — not from being a thinner wrapper around the same general capability.
Origin
Logan Kilpatrick, head of developer relations at OpenAI, on Lenny's Podcast (Feb 2024), answering the question every developer building on the OpenAI API was asking after DevDay: 'will you eat my product?' He cites Harvey (legal AI) as the archetype of the safe zone.
Core principles
- 01The platform's mission defines its roadmap — read the mission, not the changelog.
- 02General capability is the platform's territory: reasoning, coding, writing, general assistants and agents.
- 03Verticalized products (an AI sales agent, a legal AI) are explicitly not what the platform is building towards.
- 04A domain-specific model can beat the general model on its own turf, because the general model optimises for breadth.
- 05If you must compete on the general layer, 'a bit better' loses — you need to be radically different.
How to run it
- 1
Read the platform's mission as its roadmap
Write down what the platform says it is building towards. For OpenAI that is AGI via general capability. Anything that sits on the critical path of that mission is territory you will eventually be steamrolled in.
Pro tip Kilpatrick's tell: 'you shouldn't be surprised' when the platform launches the general-purpose version of what you built.
- 2
Classify your product as general or vertical
Is it a general-purpose assistant/agent, or does it solve one domain's problem with domain-specific knowledge? General means you compete head-on with the platform's own engineering and research budget.
Watch out 'General purpose agent for everyone' is the single most crowded and most encroached-upon category.
- 3
If vertical, stack domain assets on top of the general model
Fine-tune on domain data, build custom tooling, and encode workflow knowledge the platform has no interest in acquiring. Let the platform fund your model R&D while you own the customer relationship.
Pro tip This is Harvey's shape: custom models and tools for lawyers and legal firms, riding the general model underneath.
- 4
If general, apply the radical-difference bar
You must be able to name ten specific problems your product solves that the incumbent chat product does not. If you cannot list them, the product is not differentiated enough to pull users off the default.
Pro tip State the ten problems out loud to a prospective user. If they don't say 'wow', you fail the bar.
Watch out Being marginally better than ChatGPT is not a business — the platform is pouring enormous engineering and research effort into that exact surface.
- 5
Differentiate on interface, not just capability
Kilpatrick's separate bet: products that move beyond the chat box (infinite canvas, hardware, domain-native UI) have an advantage precisely because the platform is optimising the chat paradigm.
Pro tip Ask what format the human actually wants the answer in — chat is often a worse container than a canvas or a dashboard-free direct answer.
In the wild
Harvey builds custom models and tools for lawyers and legal firms. Kilpatrick states plainly that OpenAI's models will probably never be as capable as Harvey's on that specific task, because OpenAI's mission is to solve the general use case, leaving the domain-specific ceiling to whoever cares about the domain.
→ Harvey rides OpenAI's model improvements without doing the underlying R&D, and stays out of the platform's path.
Kilpatrick says he talks to many developers building general-purpose assistants and agents. He calls it cool and a good idea, then warns that they will end up competing directly with OpenAI, which is already building that with GPTs.
→ The explicit signal: OpenAI will ship a general-purpose agent product, and those builders should not be surprised when it happens.
TLDraw builds an infinite-canvas interaction surface where an AI fills in details, files, and video references spatially rather than as a chat log. Kilpatrick argues this maps far better to how humans want to consume information than listing things out in chat.
→ A defensible position built on interface paradigm rather than on model capability.
Common mistakes
Building a thin wrapper on the general capability
If your only asset is the platform's model plus a prompt, the platform ships your product as a feature and you have no remaining moat.
Assuming the platform's silence is a promise
Kilpatrick is explicit that they may change their mind — the mission is the constraint, not any specific public statement about roadmap.
Competing on the chat interface itself
Enormous engineering and research effort is being poured into making the incumbent chat product incredible. That surface is the worst place to fight.
Is it for you?
Best for
Founders and PMs building an AI product on top of a foundation-model API who need to decide where their durable differentiation actually sits
Not ideal for
Teams building purely internal AI tooling, where competitive encroachment by the platform is irrelevant
From the transcript
“our models are probably never going to be as capable as as some of the things that that Harvey's doing because like our our goal…”
“you shouldn't be surprised when you know we end up launching some like general purpose agent product”
“like it has to be so radically different like people have to really like be like wow this is solving like these 10 problems that…”
“I think products that like move Beyond this chat interface really are going to have such an”
From the episode
Inside OpenAI
Logan Kilpatrick (head of developer relations)