Lindy AI founder on the agent moment: B2B is ready, consumers aren't, and p(doom) is terrifying

Flo Crivello's core argument is that the agent moment is already here for businesses and nowhere near here for consumers — and that conflating the two is a strategic mistake.

Lindy AI is a no-code platform for building AI agents that automate business workflows: sales, customer support, operations. Starting at $50/month, it targets SMBs and enterprise customers through a combination of self-serve and a partner network of implementation specialists.

B2B is ready; consumers aren't

Crivello is bearish on consumer agents. His reasoning is structural: consumers are time-rich and money-poor, which means the incentive to delegate routine tasks is weak. Google Assistant and Alexa failed not because the technology wasn't good enough, but because the transactions consumers perform are already frictionless — businesses have spent years making them that way. The flight booking experience, he says, is already fine.

Where Crivello sees transformational potential is in the enterprise: a "drop-in replacement for a human worker" that he expects to arrive in 12 to 18 months.

Enterprise adoption vs. enterprise usage

Crivello draws a sharp line between AI spend and AI adoption. His read on the enterprise market is that board-level pressure is unlocking large budgets — "hundreds of millions of dollars" — but that a lot of what's being purchased is effectively shelfware, a dynamic he says echoes the 1990s Oracle and Microsoft enterprise software era. Revenue at some high-profile AI startups is growing fast while internal adoption remains thin, and satisfaction is low even where deployment has happened. SMBs, by contrast, are the ones running agents on mission-critical workflows today.

“I broadly buy the meme that 2025 is the year of agents. The holy grail of AI agents is what we call the drop-in replacement for a human worker, and I think that's coming in twelve to eighteen months. We are seeing costs fall by roughly 40 to 100x every year. Llama four — the evals are good and the vibes are quite bad. There are rumors they injected some data from the evals into the training set at the last minute.”

The multi-product platform argument

Crivello frames Lindy as a single product with many use cases, built from a small number of low-level primitives that combine nonlinearly. His argument is that roughly 95% of the work in any vertical agent — phone agents, recruiting, customer support — is shared scaffolding: team collaboration, orchestration, SaaS infrastructure, go-to-market. That shared base means adding a new capability costs far less than building it from scratch. The agent swarm demo he shows, which deploys three parallel agents to book a flight, find an Airbnb, and reserve a car simultaneously, emerged from combining two independently built primitives: agent swarms (released the prior week) and computer use (still in development at the time).

The longer-term case for a unified platform is agent collaboration. Crivello says Lindy already supports agents that hand off to each other — a sourcing agent feeds candidates to an outreach agent, for example. He argues that as agents develop non-human-readable communication between themselves, being on the same platform will matter more, not less. His near-term product direction is an "AI chief of staff" that designs and manages the topology of an organization's entire agent network, recursively.

Inference costs and model quality

Crivello says inference costs are falling 40 to 100x per year, driven primarily by algorithmic improvements and distillation rather than hardware gains. He points to GPT-4's estimated 1.7 trillion parameters versus GPT-4o's roughly 50 billion as an illustration of how much compression has already happened. Lindy's unit economics are thin today, but he and his investors are treating that as a short-term condition.

On Llama 4, Crivello is direct: the evals look good, the vibes are bad. He cites rumors that Meta injected eval data into the training set at the last minute, and notes that Meta's VP of AI research and several core researchers appear to have resigned in protest over what they considered unethical practices. He attributes the pressure to Zuckerberg raising internal temperature aggressively and threatening to restructure or fire teams that miss targets.

How Lindy uses itself

Crivello's own weekly agent stack is a concrete illustration of the cron-job use case he sees as underappreciated. He runs agents that summarize his favorite podcasts, compare his calendar against his stated priorities, surface warm contacts from his personal CRM, and compile meeting digests. Company-wide, Lindy ingests external meetings, support tickets, and internal communications daily and distributes a Slack digest to the whole team. His view is that rapid data ingestion and summarization is one of the strongest current agent use cases.

P(doom)

Crivello says he's terrified. His p(doom) fluctuates with wide error bars, but his point is that even 10% — the optimist end of what he heard serious researchers cite on the Dwarkesh podcast — is an unacceptably high probability of civilizational extinction within a decade. He finds it under-discussed relative to how seriously the people closest to the technology take it.

Lindy is currently hiring engineers, designers, and salespeople. The company has a partner network at lindy.ai/partners for implementation support.