Tom Blomfield, a general partner at Y Combinator, recently framed one of the most important questions for founders in the AI era: what happens when a company can observe its own work, reason about what is happening, change its tools, and learn from the result?
The answer is bigger than a faster team. For years, software companies treated AI as an assistant placed beside existing workflows. An engineer asks for code. A marketer asks for copy. A support agent asks for a draft reply. That is useful, yet it leaves the company shape unchanged. The org chart still routes information upward through meetings, managers, dashboards, and status reports. Decisions still move downward through plans, tickets, and follow ups.
Blomfield argues that AI pushes founders toward a different company design. The company becomes a set of learning loops. Every loop has sensors, memory, rules, tools, checks, and feedback. Product analytics notice friction. Customer conversations reveal repeated confusion. Sales calls surface objections. Engineering logs expose failures. AI systems read those signals, propose changes, test them, pass them through quality gates, and write the result back into the company memory.
The phrase that matters is legible to AI. A company cannot improve itself if its knowledge lives only in private context and hallway conversation. Emails, calls, meetings, tickets, code reviews, decisions, experiments, and customer notes need to become searchable artifacts. The asset is the living context of the company. Software can be rebuilt. Reports can be regenerated. Playbooks can be rewritten. The durable advantage is a company brain that understands what the team has learned and can apply it again.
This has immediate consequences for founders. The old question was how many people are needed to operate a function. The new question is how much of this function can become a closed learning loop. A product loop can find where users drop off, generate a test, ship a variant, measure the result, and update the roadmap. A support loop can cluster complaints, detect missing documentation, draft fixes, and route sensitive cases to humans. A research loop can collect papers, extract equations and charts, summarize evidence, and turn messy ideas into reusable assets. In that kind of workflow, tools such as Miss Formula help convert formula images into usable math, Editable Figure turns AI generated paper figures into editable vector graphics, while ChatGPT and Gemini can help reason across notes, drafts, code, and decisions.
The provocative management implication is that coordination changes shape. Middle management historically existed because information was expensive to transmit and interpret. AI lowers that cost. A founder can ask the company brain what customers are struggling with this week, which experiments failed, which support topics are rising, and which policies conflict with the product roadmap. A strong individual contributor can operate with far more leverage because the system captures context, remembers decisions, and suggests the next move.
Human judgment becomes more precise. People remain essential at the edge of the system: unusual customer situations, ethical calls, creative taste, final accountability, and moments where trust matters more than speed. The best human role shifts from moving information around to designing loops, setting standards, inspecting outputs, and handling ambiguity.
There is also a measurement trap. Burning tokens can become more important than adding headcount, because tokens are the fuel of machine intelligence. Yet token usage alone is a poor trophy. A company can spend a huge number of tokens and learn nothing. The useful metric is whether each loop produces a better artifact: clearer documentation, better conversion, fewer support repeats, faster research, safer deployments, or stronger customer understanding.
The path starts small. Pick one recurring workflow with clear inputs and visible outcomes. Capture every artifact. Define the policy. Give the AI tools it can use. Add quality gates. Measure the result. Feed the learning back into the system. Then repeat. The first version will feel modest. The tenth version may feel like a department that keeps improving while the team sleeps.
The message from Blomfield lands because it changes the ambition. An AI company becomes an organization designed to see itself, remember itself, and revise itself. Founders who understand that shift will build teams that are smaller, faster, and stranger in the best possible way: companies whose real product is their own capacity to learn.
Top comments (0)