A $20,000 AI Agent Replacing a $90,000 SDR: Manny Medina on Pricing in the Agent Era

Hosts Pat Grady and Lauren Reeder sit down in London with Manny Medina, founder of Paid, the billing, pricing, and margin stack he is building for agentic AI companies. Medina previously ran Outreach, the sales-engagement unicorn, and he is now selling infrastructure to the same founders who are trying to turn demos into durable businesses. Asked what is working in AI apps, he reaches for Isaiah Berlin: this is hedgehog land, not fox land. Take a very narrow problem, become the best at that one thing, and you are printing money.

The examples he loves are almost aggressively small. Quandri handles insurance policy renewals. Owl, also in insurance, reviews claims data — a market that looks tiny until you add up leftover claims volume. HappyRobot calls truckers on behalf of freight brokers: if you want to send beer from Milwaukee to Boston, it spins up 2,000 agents, they negotiate with drivers, book the load, and keep calling through delivery. Sequoia-backed XBOW does pentesting: it ranks traditional agencies, then continuously bangs on every app and backend at a quality and cadence that human shops cannot match.

The pattern is not "software eats a job." These are people-heavy problems with no clean software product, and the thing being displaced is the BPO. Medina's line is blunt: wherever BPOs have a big role, that is food for the apex predator in AI right now, which is the agent. The converse is also useful. Harvey is doing well by staying inside legal rather than boiling the ocean; Crosby, in the same space, is trying to fully replace the commercial-contracts lawyer. AI SDRs as a category are already too broad — the phrase means too many things to too many people, which produces swirl. AI executive assistants are not there yet either: Lindy and Fyxer exist, but Medina works across three time zones and Fyxer cannot quite keep up, even though it would be a perfect fit for a single-timezone real-estate broker. Big problem, narrow application, nailed it.

Which markets transform first? Medina's answer is a direct quarrel with Kleiner Perkins' Mamoon Hamid. He remembers the moment Sequoia's paper on the shift to cognition landed around OpenAI's Strawberry / o1 release, and he remembers hearing Hamid's hypothesis: AI will start with the highest-paying jobs because that is where the money is — developers, lawyers, accountants, doctors. Medina fully disagrees. Highest-paid people will buy AI as a side thing and ditch it with the same regularity they ditch other gadgets. The landing he is actually seeing is full takeover of roles nobody else wants.

Nobody wakes up wanting to be an insurance actuary or an insurance adjuster — Sequoia partner Roelof Botha, perhaps the only licensed actuary in venture, is the joke exception. Policy-renewal staff are exiting and not being backfilled. Nobody wants to bang the phones at a BPO for six months on the way to a different job, so turnover is extreme. The companies doing well, Medina says, are addressing labor pools that are disappearing into retirement, trying to leave, or already run by BPOs. In that segment he sees stickiness, expansions, growth, and good economics: the upper bound of price is labor cost, the lower bound is your own margin, and people are going to town in between.

Grady tries to keep both theses. For higher-paying creative work, the copilot approach — Harvey in legal, OpenEvidence in medicine — seems to give people superpowers. For lower-paying, less creative work, full autopilot replacement looks right. Collaborative workflows, he adds, are still "the eighth wonder of the world" in software: once you are deeply embedded, customers do not want to switch. Medina is not so sure the old playbook still holds. Everyone will eventually crowd the rich-job market, including OpenAI if it needs another revenue line, because lawyers and accountants pay a lot and buy everything. That market will be hot, then brutally competitive.

Medina's skepticism about copilot moats has two layers, and the first is about workflows. Asana- and ServiceNow-style collaborative systems were revolutionary when the category did not exist. Today anyone can spin one up: "you can vibe code a ServiceNow." Asked whether he believes in vibe coding, he does not hesitate. Paid itself was stood up in about a month and a half, and half of it was vibe-coded. They have to throw code away, which he treats as the point. Do not hang up on debugging vibe-coded software; throw it away, put it in production, and when it breaks, start again. Starting is cheap.

He grants Grady's point that becoming the definitive workflow for a given job is still sticky. What has changed is the speed of competition. The explainability of copilot value is genuinely hard to land: how do you differentiate one copilot from the next when every vendor can say "I am the same as X, but better or cheaper"? That is a recipe for swirl. Specialization is the escape hatch. If Harvey said it was the best at patent law and owned 78% of that market, that would be a landing. You would have moved into a very narrow application inside a rich market — rich because you own all of it, not because the whole category is rich.

On pricing and packaging, Medina sees four things sticking. First is charging by activity: a credit-consumption model where you can show the work that was done. Second is charging by workflow — string activities together and price a document review as a unit, so a short file and a long, complicated one can carry different prices because they have different consumption patterns. Grady notes that this already moves you from cost-based toward value-based pricing. Medina agrees: leaving the treadmill of charging for pure work lets you charge for something that is actually worth something to someone.

Third is outcomes, and his tactical advice is not to jump straight to pure outcome pricing. Charge an outcome bonus: if a measurable result of a particular quality happens, bill for it. That opens a conversation about value alignment, and once that door is open, contracts become more bespoke — historically hard, now newly practical. The old world wanted every customer in SKU boxes, with HQ holding a discretionary discount. Large ServiceNow and Salesforce deals were already bespoke; the CRO flew out, sat with the counterparty, and came back with terms that would not fit a CPQ tool. In the agentic world Medina sees no reason not to do that with every customer you want to go big with. You can put a chat interface on the contract corpus and ask it to interpret a deal and give you the annualized value.

Fourth is pay-by-agent. Working with AI SDR companies, Medina has been introducing a simple comparison. A fully loaded human SDR costs somewhere between $70,000 and $90,000 a year and produces a bounded set of activities: a book of accounts, contacts, a hundred calls a day. An agent doing about 80% of that work can be sold at $20,000, with a bonus for meetings booked and another for opportunities closed. Draw a boundary around the activity, charge for the agent because the output is the same, and you have something that looks like hiring. The budget trick matters: you dip into the headcount pool rather than the tools pool, so you are not stuck in RevOps' sliver of CRO spend. You are in HR's much larger bucket.

The pitfall of charging as a tool is getting pigeonholed into seats. Stay on activity-based pricing and someone will come along and do the same thing cheaper; then you are one of fifty vendors who look and sound identical, and the only way customers separate wheat from chaff is to try you and churn. Customers will always default to the easiest first-year buy — a fixed price or a consumption price — to see if it works. If it does, it is on the agent builder to go back and align on what actually matters to that customer, then charge for it. Pricing, Medina notes, quoting Seth Godin, is part of your story. If the story is different and the pricing is the same, the story ends up the same. That is why Sierra does not put pricing on its homepage: it sits down with each customer, finds out whether success is time to resolution, a ticket that turns into a purchase, CSAT, or NPS, and builds an outcome function — the same idea as an objective function in machine learning — then iterates toward it.

The vast majority of AI agent companies, Medina says, are doing proofs of concept right now. They look like contracts and they look like money, but every company on the planet has an AI mandate, so somebody went and bought software as a trial. Grady coins the name on the spot: vibe revenue. Medina takes it. Soon the industry enters renewal land, and that is what will separate stickiness from theater. Real monetization schemes will be designed after customers can see who made money, who did not, and what outcomes actually showed up.

Early customers told him bespoke, outcome-based contracts would never happen. Now they call back: they just got their first contract that only pays per outcome, and they do not know who else to call. He thinks this gets pulled by the buyer as a way to mitigate risk, because AI is risky on several dimensions and paying for outcomes lowers that risk. Getting into a BPO budget often starts by matching the BPO price, going cheaper, running 24/7, and accumulating the data the BPO used to own — an intermediate step, not the end state. BPOs will not sit still. They are large enough to buy someone else's technology, deploy their own agents, and train on internal data. Medina has talked to BCG; they would not admit to a lot of BPO work, but he sensed it. He does not know how the balance of trade levels out. He does know that unless you start asserting quality in the price — five qualified meetings, or an agent that replaces a human — you get competed away.

The cost side is another place Medina is contrarian. Everyone tells him token costs will go down and models will commoditize. In a world of reasoning, where inference-time compute is worth more than training-time compute, he does not see token prices falling in the immediate future, because you are going to require deeper thinking. The agent-building habit he dislikes is boxing work in LangGraph and stringing steps together. If you hallucinate at the start of that chain, you are screwed for the rest of it, and no amount of evals will rescue you. Better to let the model do most of the work, wrap it in a good eval framework, and roll it out. In that lower-error, heavier-thinking world, cost per token if anything goes up as models get smarter.

Grady offers the long-arc horse race: price for a fixed workload may fall while workloads grow because you throw more compute at inference or take on harder jobs. Medina agrees over history and says he simply does not know in the very short term. His aha is multimodal. Nail text, then you want a phone call, an inbound line, an avatar — costs that are not LLM bills but third-party APIs. Cloud plus model plus telephony plus data plus whatever else has to sync is what actually drives the cost. For a phone agent the bill is talk time and dials; for an avatar company it is how long the avatar runs. Those compound.

The ugly part is attribution. All agent activity typically goes through an eval proxy before tokens hit the LLM, so you cannot see which customer is profitable, which one is wrecking margins, which agent is doing a good job, and which is not. You see a blob. That is why the margin problem is such a bear. Agent companies are new and hungry, so they price however the buyer wants to buy and only later find out whether they won or lost. Paid looks up the human equivalent of a given activity or workflow — by country, by type of work, by hourly wage and duration — and hands founders a ceiling: you can price up to that. Simon-Kucher will visit once a year and leave you holding the bag; this kind of guidance is needed every time you talk to a customer. The problem Medina keeps seeing is that value is accruing to the customer, not to the agent business. Customers capture all the savings. That needs to change.

Paid exists because of a problem Medina could not shake after rolling out agents at Outreach. He wanted to understand margins and the value those agents added, and the underlying software was not built for a world where pricing has to evolve and the product delivers full outcomes rather than bits and bytes. After he stepped aside as CEO and moved to London, he spent a couple of months calling founder friends with a cross-education trick: they could AMA him on building sales teams, and he would ask what it was like to run an agent business. The answer was spreadsheets. The problems were difficult but not intractable — a lot of work, not a couple lines of code — which is the kind of gap he likes. He also wanted to live in founder-to-founder conversations, which he found raw, full of wonder, and almost not like selling. Cold calls got returned. When he said he was building this, people said they would try it. He claims a 100% hit rate on that line.

Paid, he says, is the business engine for AI companies: billing, invoicing, monetization, pricing, collections, revenue recognition, vendor management, and margin management — the back office required to understand unit economics and actually run the company. The first wedge is the monetization and margin engine, because that is the pain he sees most; adjacent pieces follow because the problem compounds. What transfers from Outreach is the human work: building teams, inspiring people, and selling ahead of capability. What is new is how savvy company-building has become. Lift the hood on an AI agent startup and you find people building a company, only much further along than he was at the same stage of Outreach, because they have seen the whole SaaS arc.

He is not sure whether that is all good. The positive case is better-informed decisions; the negative is blind pattern-matching. What he actually observes is different initial conditions: keep the company small, own more of the destiny, raise differently, refuse growth-at-all-costs, and spend longer honing the ideal customer profile. At Outreach he could sell to everyone — startups, Adobe, the Hillary campaign, a repo company — because they all had communication and workflow problems, each with a completely different path to a product. He wallowed in that sea of business. If he relived it, he would ask what would make Outreach a billion-dollar run-rate company and stack from there. New founders start with the narrowest profile they can sell without friction. The risk is that they restrain the ICP and never go from single product to platform. Grady's counter is the one Medina fully buys: quality scales. Nail it for a small audience rather than being mediocre for a big one. Roadmaps get clearer, tickets look the same, word of mouth compounds.

In the lightning round he puts Jeff Bezos, Pendo's Todd Olson, Sam Altman, and the Collison brothers on his founder Mount Rushmore. The book he wants every AI founder to read is Christopher Manning and Hinrich Schütze's *Foundations of Statistical Natural Language Processing* — Markov chains, look-back and look-forward, the statistical guts still underneath today's models. The product he cannot live without is Perplexity, which even found his wife a UK driving instructor; Claude is the imaginary friend he writes with and "shouldn't be talking to." Are models commoditizing? Not yet, not in a reasoning world: each new model's input tokens cost six, seven, or eight times the generation they replace. AGI, he thinks, is already here in an underused way, behind door number three, unacknowledged. His one piece of advice: stay focused on a very narrow set of customers. Do not worry about TAM. Disregard VC advice about big TAMs. Small TAMs become big TAMs as long as you deliver a superb experience.