Building an Autonomous Enterprise for Real-World Services with Netic Founder Melisa Tokmak

Melisa Tokmak, founder and CEO of Netic, went on No Priors with Elad Gil to make a direct case: her voice and text agents now handle the first customer interaction for over 70% of her essential-services customers, across HVAC, plumbing, pet care, and hospitality. The platform has, by her account, generated $600 million in revenue for those customers. Dispatch, scheduling, triage in real time. Not demos.

The 70% figure is the one to trust. It is a deployment fact: the agent answers the phone first. The $600 million is a last-touch attribution claim, and it has exactly the problem that marketing-mix models spend years trying to fix. If a customer calls to book an emergency furnace repair and the agent takes the booking, the broken furnace generated that revenue. Tokmak's team counted every dollar that flowed through a conversation the agent touched. That is not causation, and no vendor voluntarily replaces a large flattering number with a smaller honest one. Expect the $600M figure to persist until a buyer or auditor demands a controlled holdout.

The technical argument underneath is more interesting, and it holds up. Tokmak says the bottleneck is not model capability; it is the orchestration layer and domain-specific harness sitting above the model. That matches the applied-AI literature. Enterprise deployments stall not because the model is weak but because nobody can verify its output fast enough to ship it. The hard engineering is the plumbing that checks the model and connects it to your dispatch system, not the weights themselves.

Her argument also implies that the model is a commodity input, which has a direct consequence for the competitive picture. If the moat lives in the orchestration and domain data, a well-funded competitor can stand up a rival harness without needing a proprietary model. That is the same fact from two directions: it is what makes Netic's moat real today and what makes it replicable tomorrow. Tokmak did not resolve that tension on the podcast.

Voice also carries a cost structure that text agents don't. Every turn is speech-to-text, inference, and text-to-speech, all under a hard latency ceiling because a human is waiting. At a million calls a month on frontier API pricing, gross margin collapses exactly when the business succeeds. The smart operational move, routing cheap models for easy turns and reserving expensive inference for the hard ones, is almost certainly where Netic's real engineering effort sits. Model-swappability is not a philosophical position here; it is survival economics.

For anyone building in this space, the thing to de-risk is not model quality. It is the eval harness for false confidence, the calls where the agent thinks it handled the situation and did not. Build the evaluation layer before the agent: score transcripts on whether it booked correctly, escalated when it should have, and hallucinated no appointments. That confidence-gated handoff to a human is the entire game.

Elad Gil's aside is worth sitting with. He suggested founders are now too worried about what the labs are doing and are avoiding verticals they would have entered without hesitation before. If that is right, the labs' biggest competitive advantage is the fear they generate before they ship anything, not any product they have actually released.

Prediction: Through Q2 2027, Netic will continue leading with the 70% AI-first deployment stat rather than a third-party-audited causal revenue figure. The attribution claim cannot be independently verified, and the pattern across applied-AI startups is to keep the softest impressive number front and center until a buyer forces precision. We are wrong if Netic publishes a controlled-holdout measurement isolating incremental revenue the agents actually caused.

Melisa Tokmak, founder and CEO of Netic, went on No-Priors with Elad Gil to make a claim that should interest anyone shipping AI into a real business: her voice-and-text agents now handle the first customer interaction for over 70% of her essential-services customers, and she says the platform has generated $600 million in revenue for those customers. HVAC, plumbing, pet care, hospitality. Not coding demos. Dispatch, scheduling, and triage in real time. The question for a technical team lead: is the moat she describes, orchestration and domain depth above the model, real enough to build a company on, or is it the thing a frontier lab eats in eighteen months?

This is a Type 2 read for most of you. Nobody's forcing a decision this week. But the underlying bet, "vertical depth beats raw model capability," is a Type 1 architecture choice the moment you commit engineering quarters to it.

The Skeptic. $600 million is a great number until you ask what it means. Revenue "generated from AI-handled interactions" is not revenue caused by the AI. If a customer calls to book an emergency furnace repair and the agent takes the booking, did the agent generate that revenue, or did the broken furnace? Attribution here is doing exactly what marketing-mix models do badly: crediting the last touch. The 70% AI-first figure is cleaner and more believable, because it's a deployment fact, not a causal claim. For a PM: "AI-first" means the robot answers the phone first, which is real and measurable. "$600M generated" means someone counted every dollar that flowed through a conversation the agent touched.

The Researcher. Tokmak's actual technical claim is the interesting one, and it lines up with the literature. She says the bottleneck is not model capability, it's the orchestration layer and domain-specific harness. The saved "Checking Problem" paper says the same thing from the regulated-firm side: enterprise AI stalls not because the model is dumb but because nobody can verify its output fast enough to ship it. That's a harness problem, not a weights problem. Her robotics argument is also grounded, not hype: dexterity and building variance push embodied automation out decades, so all the near-term value sits in software. For a PM: the smart model is the easy part now. The hard part is the plumbing that checks it and connects it to your dispatch system.

The Open-Source Advocate. Here's where her thesis actually helps the rest of us. If the moat is orchestration and domain data, not the model, then the model is a commodity input. That means an open-weight Llama or Qwen with a good harness competes with GPT-4o-class calls for a lot of this work, and the economics of voice agents at scale demand exactly that. Real-time voice triage at a million calls a month on frontier API pricing is a bad P&L. The teams that win vertical AI will be the ones who can swap the model underneath without rewriting the harness. Tokmak's own argument, that labs optimize for generalizability while she optimizes for depth, is the case for treating the model as replaceable.

The Compute Pragmatist. Voice is the tell nobody on the podcast priced out. Text agents are cheap. Real-time voice with sub-second latency, multi-turn, plus scheduling and dispatch optimization running alongside, is a latency and cost budget most people underestimate. Every turn is speech-to-text, inference, text-to-speech, and you're paying for all three under a hard latency ceiling because a human is waiting on the line. Model-swappability is survival economics, full stop. If your inference bill scales linearly with call volume on frontier pricing, your gross margin dies exactly when you succeed. The interesting engineering at Netic is almost certainly caching, routing cheap models for easy turns, and reserving the expensive model for the hard ones.

The Builder. What would I actually ship from this? Not the whole platform. The pattern. Agent answers first, triages, escalates to a human on low confidence. That confidence-gated handoff is the entire game, and it's what the "Checking Problem" paper is really about. Build the eval harness before the agent: transcripts scored on did-it-book-correctly, did-it-escalate-when-it-should, did-it-hallucinate-an-appointment. Tokmak's team lives or dies on the false-confidence rate, the calls where the agent thinks it handled it and didn't. That's your 3 AM page: a customer whose emergency call got mishandled by a bot that was sure of itself.

Where they part ways. The Skeptic and the Researcher split on the same $600M. The Skeptic reads it as attribution theater. The Researcher shrugs and says the deployment metric is what matters anyway, so who cares. They're both right, which tells you to trust the 70% and discount the dollar figure. The bigger tension is Open-Source Advocate versus the whole "labs can't compete" thesis. Tokmak needs the model to be a commodity for her moat to hold, but a commodity model is also what lets a well-funded competitor stand up a rival harness fast. The depth is the moat and the low barrier at the same time. That's not a contradiction she resolved on the podcast.

What it hinges on. One belief: does the orchestration-and-harness layer compound faster than a frontier lab can generalize into the vertical? Tokmak bets yes. Elad Gil's aside is the counter-signal worth sitting with, that founders are now "too worried about what labs are doing" and won't enter verticals they'd have fought for in normal times. If he's right, the labs' biggest weapon is the fear that freezes competitors before they start, not any product they've actually shipped. For your own team: if you're building vertical AI, the thing to de-risk is not model quality. It's your eval harness for false confidence and your ability to swap models when the price or latency math turns on you. Build both as if a lab will ship a competing agent next year, because the harness is what survives when the model underneath changes.

Prediction: By the time Netic next discloses metrics or raises its next round (through Q2 2027), it will still lead with the "70% AI-first" deployment stat, not a third-party-audited causal revenue figure, because the attribution claim can't be independently verified.

Confidence: Medium — deployment metrics are checkable; causal revenue claims almost never get audited in vertical AI.

Why: Tokmak's two headline numbers are different species: 70% AI-first is a fact about who answers the phone, while $600M "generated" credits every dollar in an AI-touched conversation to the AI. The first survives scrutiny, the second is the exact last-touch attribution problem that marketing-mix analysis spent a decade discrediting, and no vendor voluntarily replaces a big flattering number with a smaller honest one. The pattern across applied-AI startups is to keep leading with the softest impressive figure until a buyer or auditor forces precision, and nothing in this segment forces it.

Revisit by 2027-06-30: We're right if Netic's next public metrics still headline deployment share or an unaudited "revenue generated" figure. We're wrong if it publishes a third-party or controlled-holdout measurement isolating incremental revenue the agents actually caused.