Key Takeaways
- Treating artificial intelligence as an end-to-end digital co-worker mirrors the early automotive mistake of designing cars like horse-drawn carriages.
- True industrial automation isolates and locks in narrow mechanical tasks, such as twisting bottle caps on an assembly line, rather than replicating an entire human role.
- Startups like Harvey and Cognition claim proprietary model intelligence primarily to counter the investor perception that they are defenseless wrappers on foundation models.
- Incumbent software companies hold the structural advantage in enterprise automation because they already control customer workflows and can fund heavy software engineering.
- Real technological moats come from building deterministic software infrastructure around general foundation models rather than training private, bespoke models.
The Horseless Carriage Trap
When early engineers built the first automobiles, they did not rethink transit from scratch. They attached internal combustion engines to wooden buggies and called them horseless carriages. Some inventors even built mechanical horses to pull carts.
Diogo Almeida sees the current wave of enterprise AI startups repeating that exact mistake. Companies pitch autonomous digital colleagues that mimic human lawyers, engineers, and support reps. Almeida rejects that framing entirely: “I believe that this is going to be like the mechanical horses or the horseless carriages of the AI era just like the industrial era people were like let's just make it look like a carriage but without the horse. I believe that the AI co-workers are a form of that.” (play from 31:52)
Real factories never replaced human workers with humanoid robots doing every task at once. They broke workflows down into tiny, repeatable motions. “You don't need to wait for a whole coworker like a humanoid robot to do tasks,” (play from 32:20) Almeida explains. “You just find the extremely narrow thing like twist bottle cap on bottle and then you automate it and it just gets solved forever. So I think that's how work will generally go.” (play from 32:25)
Why Startups Claim Bespoke Intelligence
Every enterprise pitch deck claims proprietary intelligence. AI legal assistants and automated software engineers insist they run on custom fine-tuned weights or private model architectures. Almeida views this claim as defensive marketing rather than technical necessity.
Training and maintaining bespoke frontier models inside vertical enterprise applications is expensive and fragile. Every time OpenAI or Anthropic releases a smarter general model, narrow fine-tuned variants fall behind. The actual durable value in business software is not custom weight training. It is the deep, defensive systems engineering wrapped around frontier foundation models to guarantee reliability.
SaaS Incumbents Hold the High Ground
Because automation requires precise task isolation, established SaaS platforms maintain an advantage that pure AI upstarts struggle to overcome. Legacy platforms already own the user interface, the system of record, and the precise mechanical context of the work.
Almeida emphasizes that automation requires boring, heavy software development rather than purely generative prompts: “SaaS has a massive advantage over AI companies in that they're actually familiar with the workflows. They have the scale in order to spend the investment up front to do the engineering because it takes real engineering to automate stuff.” (play from 30:01)
Once a specific workflow step is solved and locked in with stable software, enterprises have zero interest in constant, unpredictable model changes. Proprietary value belongs to the company that controls the mechanical integration, not the company trying to simulate a human employee.
What to Do With This
Audit your product roadmap and strip out any feature designed around an open-ended conversational co-worker. Pick the single highest-friction mechanical step in your customer's weekly routine, such as pulling contract renewal dates into billing tables. Build a deterministic software loop around an existing foundation model to solve that exact step with zero human intervention.