Patrick Feeley: Evaluating AI in Public Companies Should Focus on Processes Rather Than Headcount
Investor Patrick Feeley commented on Y Combinator President Gary Tan's speech at Startup School, stating that Tan was not arguing for the rise of one-person companies but rather describing operational models that can already generate numbers. Most valuations still treat headcount as a scarce input, while what is truly scarce are tightly written processes that can be executed daily by intelligent agents, and the routing that allocates work to the correct processes.
Feeley provided an example: Emergent, founded by Mukund Jha and Madhav Jha, was noted to have achieved approximately $15 million in annualized revenue with about 15 people, reaching nine-figure revenue about eight months after going public; in July, they raised $130 million in Series C funding, with a post-money valuation of $1.5 billion, annualized revenue of about $120 million, and over 200,000 paying users, with around 200 employees at that time. The voice AI company Retell has an annualized revenue of about $60 million with a team of about 40 people. Revenue per employee exceeds the old software and industrial benchmarks.
Tan mentioned in his speech that the next generation of startup teams will be smaller, as personal-level general intelligent agents can now compound knowledge on their own infrastructure; a skills manual is equivalent to one employee, and in the winter batch, 95% of the code from a quarter of the companies was generated by AI. Feeley added a caveat: untrimmed memory is like a cluttered drawer with search capabilities, where outdated facts can be retrieved with certainty, and bad skill files will forever disrupt processes. Every fact must have a source, new information must be verified against old information, and someone must delete content that no longer belongs to the system.
Software is cheap enough to create dedicated tools for a single user. The old advice was to scratch your own itch before hoping for a market; the new version is that costs are nearly zero, so you create first and turn it into a company if someone shows interest.
In market mechanisms, those paying the bills are rewriting per capita output into growth stocks and venture capital, while those selling are valuing old software based on linear extrapolation of headcount. The events are driven by speeches and commentary, with capital flowing to companies that can write processes into intelligent agent skills; benefiting are high-revenue, low-headcount AI-native teams, while pressured are public company models that still explain growth through headcount.
Source: Public Information
ABAB AI Insight
Headcount used to represent capacity. Now, capacity is written in reusable process documents. Tan speaks of one person plus an intelligent agent already forming an organization, and Feeley translates this to mean that sell-side research should no longer treat headcount as a constraint. Emergent's narrative from fifteen people to a two-hundred-person unicorn illustrates that intelligent agents first raise the slope, with headcount following. Retell's forty people generating sixty million shows that call minutes have turned into usage-based billing, with headcount no longer increasing linearly.
The capital path is to treat skill documents as employees and routing as middle management. Money is shifting from seat-based software subscriptions to task and minute-based intelligent agent billing. In the Y Combinator batch, "completing a year's work in a day" compresses product cycles and also compresses models that estimate revenue based on headcount. If memory is not trimmed, intelligent agents will make decisions based on outdated facts, and auditing costs will erode the myth of per capita productivity.
The benchmarks are Salesforce on a per-seat basis, and oil and gas railroads based on capital expenditures and headcount. Software is at a stage where "the cost of writing tools is approaching zero, and the cost of maintaining memory has become the new bottleneck": those who can delete outdated knowledge will have auditable intelligent agents. Turning self-made small tools into companies is due to the collapse of distribution and marginal costs simultaneously.
Structurally, this belongs to technological substitution. The mechanism is: when processes are written tightly enough, intelligent agents can execute daily, shifting scarcity from hiring to process ownership. If valuations still multiply by headcount, they will systematically underestimate companies that have already changed their clocks. Cluttered drawer memory sounds omnipotent, but what it delivers is un-auditable confidence.
ABAB News · Cognitive Laws
Headcount is no longer capacity; written processes are.
Completing a year's work in a day means those still running on old clocks are already lagging behind.
Untrimmed memory will make decisions based on outdated facts that sound very certain.