二十个 Agent、1.2 个人：SaaStr 用 AI 跑销售团队之后发生了什么

Jason Lemkin, founder of SaaStr, traces the decision to the community's 10,000-person annual event. The company went in with a seven-figure budget and an eight-figure top line, and two well-paid salespeople quit on site. It was, he says, the eighth sales team he had built, and he turned to his chief AI officer, Amelia, and declared that SaaStr was done hiring humans in sales.

He had reason to think agents could do it. Before the event, a general-purpose digital clone of himself called Deli had closed a $70,000 sponsorship on its own. If a horizontal agent that was not trained for sales could close one deal, he reasoned, a few purpose-built agents deployed with time to spare could do much more, and he no longer wanted to pay a junior SDR $150,000 a year only to watch them leave.

Today the office has ten desks that used to belong to go-to-market staff, each now labeled with the name of an agent. Lemkin calls it the quietest office around. He adds that leaders at top AI companies would also prefer the smallest sales teams they can run, partly for cultural reasons.

SaaStr's sales work breaks into two very different motions. High-end sponsorships average roughly $70,000 to $80,000, while tickets, ranging from a couple of hundred dollars to about $2,000 for last-minute buyers, are a self-serve, high-volume business worth four or five million dollars a year. Because the motions differ, SaaStr uses separate agents and, for now, separate vendors for each, plus more agents to win back lapsed buyers.

The old team was two or three SDRs and up to five account executives. Now there are 20 agents, one full-time AE, and Amelia, who spends about 20% of her time orchestrating the agents; that is the 1.2 humans. Lemkin says the business performs about as it did with ten people, and that if two more great humans wanted to join, he would hire them tomorrow.

What the agents replace, he argues, is the mid-pack and below: people who do not really understand the product after three months. He puts it bluntly: AI is replacing the jobs people do not want to do and displacing "the midpack and the mediocre," while the best humans gain superpowers.

Lemkin separates plays from playbooks. Outbound, webinars, podcasts and in-person events all still work, and that is why companies like ElevenLabs keep running road shows. What no longer works is the 2021 playbook, because growth at older SaaS companies has decelerated so much that these plays lack the ROI to justify their budgets.

The market has split in two. Hyper-growth AI companies are swamped with demand and are choosing which leads to answer; Lemkin recounts an old colleague at Bolt saying half his job was picking whom to respond to, and that Bolt closed a seven-figure deal Lovable had left unanswered. At the other end, slower companies are under cost pressure. Both ends therefore have a reason to push AI hard in go-to-market in 2026.

Lemkin also describes an unusual buying moment. Traditionally only 3 to 5% of prospects in a category were in market in a given year; for high-ROI AI products, he says, it is now north of 50%, because everyone is trying to adopt AI at once. He expects this to be a bubble that eventually cools, since an enterprise cannot buy everything every year.

Lemkin predicts that the classic junior SDR who sends cadence emails will be about 90% displaced by AI next year. People who qualify inbound contact requests should, in his view, be extinct, since a visitor no longer needs to wait a day for a 21-year-old who does not know the product to ask what they do. SaaStr's own agent qualifies everyone on the website without them noticing and simply books the meeting with a salesperson.

For account executives he is less certain, because the tools are not there yet. He guesses about 70% of an AE's job is safe through the end of next year, trending toward 40 to 50% after that, and sees no reason a great agent could not close a deal when there is little to negotiate. The harder question is how anyone learns sales if entry-level SDR and AE jobs vanish; Lemkin notes companies are already cutting junior hires in favor of experienced people who use tools like Cursor.

Leadership roles, he says, are safe because nobody has produced an autonomous CEO, and managing a million leads takes real leaders. But he thinks only about 20% of current go-to-market people are the right ones for the future, and that a big-name résumé does not guarantee the skills a startup needs in 2026 or 2027.

Lemkin's advice for anyone worried about their job sounds simple and almost nobody does it: pick one painful problem, such as support, SDR work or inbound qualification, choose a leading vendor that treats you well, and do the work yourself. He recounts a call with a public B2B company worth over $10 billion where nobody on a 20-person call had trained an agent personally; they hoped an untrained agent handed to junior SDRs would sell by itself.

He demystifies the jargon. Ingestion means giving the agent your website, wiki, training documents and other material; training means answering the questions it generates; and QA means correcting the mistakes it makes in practice emails. Spend an hour or two a day for about 30 days, with the vendor's forward-deployed engineer helping, and the agent becomes pretty good.

Lemkin estimates this takes 50 to 60 hours plus vendor evaluation, and that completing one deployment makes you hyper-employable, perhaps even a chief agentic GTM officer. Doing it once also compounds: after months of training three agents, SaaStr distilled the result into a single prompt, gave it to Salesforce's Agentforce, and had it working pretty well in a day.

SaaStr lists its agents at saastr.ai/agents, and Lemkin's first caveat is not to build go-to-market agents yourself unless you are a company like Vercel with engineers eager to do it. He has built plenty in Replit, including a valuation calculator used 800,000 times in 90 days and a pitch-deck reviewer that has reviewed almost 3,000 decks, but none of the GTM stack. Products evolve so fast that an internal build can become obsolete within months.

The journey started with Deli, a digital clone trained on his 10,000-plus pieces of content. Attendees began using it to ask about refunds, discounts and venue details, and it effectively took over event support, which had previously been answered weeks late. Then it sold a sponsorship on its own. Lemkin suggests support is a natural first place to start for anyone, since many companies, including some AI leaders, offer little support at all.

Next came outbound with Artisan, which sent about 60,000 emails, and inbound with Qualified, which booked a sponsor at 11 p.m. on a Saturday. Later, Agentforce was pointed at leads that human reps had judged not worth their time; with a very similar prompt it drew a 70% response rate. The lesson was that agents are best at high-ROI work nobody was doing.

Lemkin says he picked Artisan and Qualified largely because both were SaaStr sponsors that offered to help the most. Another vendor wanted $100,000 up front, and a third feared bad press if the project failed. Since most of these products run on similar underlying models and APIs, he argues the usual feature matrix needs an extra column: who will actually help train my agent.

His practical test is to get the forward-deployed engineer on the phone before signing anything. If the best software has no help to offer and a rival vendor will deploy alongside you, choose the latter. Some of the best vendors even turn away customers they cannot help.

He also rejects the common excuse that a company is too small. A business with 300 customers may have 30,000 website visitors and 31,000 contacts in its CRM, yet nobody is talking to any of them. Even a couple of million dollars in revenue gives enough traffic and data to benefit, and a few months of good content, not twelve years, is enough training material.

Lemkin spent about an hour a day training Deli at the start, after it gave customers wrong dates, and no longer needs to. Now Amelia spends 10 to 15 hours a week reviewing the output of 20 agents, which he calls exhausting because agents work nights, weekends and Christmas. Being the chief AI person is not a job for lazy people, he says.

That workload is also why he is unsure SaaStr can add a 21st agent. He argues the same limit will hit buyers: after a year of onboarding agents, enterprise teams may be too tired to add another, so vendors may have about 12 months to close deals before agent fatigue sets in. Business process change, not software, is the hard part, he says.

To avoid noisy outreach, Lemkin says to take the best email from your best salesperson or marketer, train the agent on it, and let it generate and A/B test variants, with data from a source like Salesforce for light personalization. He blames the failures of 2024 on two things: models that were not yet good enough, and vendors who claimed you could just turn the product on. Terrible emails, he says, come from a poorly trained product from a bad vendor.

SaaStr has sent hundreds of thousands of messages both with and without disclosing that they came from an AI, and concluded that it does not matter if the message adds value and gets an instant reply. Some founders even write back that they can tell it is AI but would like a meeting. He sees worry about disclosure as an excuse not to do the work.

He also points out that most human outbound is poor. A top rep might spend two hours researching one email, but a 21-year-old SDR using automation tools does almost none; a former colleague at Brex, Sam Blonde, once called inherited reps' emails the worst he had read. So the bar for good enough AI outreach is a facsimile of your best person, which beats the mid-pack.

Humans still win in one case: a handful of very large targets. Lemkin describes a whiteboard of the 50 best prospective sponsors, divided among skilled sellers, where no AI is needed. At 5,000 targets that approach breaks down, and he adds that 99% of companies cannot attract a team of elite sellers anyway.

Lemkin sums up SaaStr's structure: agents cover the whole top of the funnel, and one human AE plus 20% of Amelia's time handles what five or six AEs used to do. He is skeptical of new GTM engineer job postings and thinks that for now you must promote someone internally, a data-loving nerd from product, marketing or RevOps rather than classic sales. Running several agents also requires segmenting the customer base so they do not conflict, and master agents that manage other agents are not here yet.

Lenny asks for function-specific advice. For leaders worried about becoming obsolete, Lemkin says buy an agent and deploy it yourself rather than delegating. He notes that today's trainable products typically cost $50,000 and up, plus around $25,000 for forward-deployed help, with Clay starting around $100,000 a year, so juniors are unlikely to control that budget, though cheaper products may follow.

Juniors, in his view, should become the best at working with whatever agent their company uses. He cites Owner.com, whose CEO is aiming for $3 to $5 million per rep, rather than $300,000 to $500,000 a few years ago, with the same 100 reps. He also describes a rep who quit the day SaaStr rolled out an AI RevOps tool that tracks every action in the CRM, having done nothing for 30 days.

Lemkin stresses that AI does not mean less work. He is working harder than ever because the agents are so productive that people have to keep up. Fighting transparency and extra meetings, he says, means being left behind by peers who embrace them.

Lemkin tells startups to staff forward-deployed engineers: people who make sure an agent is trained and working on day one. He notes the idea resonates with advice from an earlier guest, Jen, to start with services and grow into software, but with a key difference: an agent without onboarding will simply fail. Salesforce is now doing something similar with a large force of people, and Lemkin says working with it feels as hands-on as his first purchase 20 years ago.

He distinguishes FDEs from traditional sales engineers, who were often one resource shared among eight reps. The FDE's top job is customer success; he describes an AI leader's $3 million deal where the FDE handled the on-site deployment and sales only managed procurement. His favorite kind, he says, is a former engineer who loves the product.

Looking ahead over the next year, support is already permanently changed, with 50 to 80% handled by AI. He thinks being a "people person" is no longer enough for a salesperson, though field sales and door knocking still work and phone automation faces regulatory limits. The biggest shift, he says, is customers expecting an agent to work before they pay, something even Salesforce's Marc Benioff wishes his customers could get.

Lemkin argues AI will increase demand for go-to-market professionals overall, because the winners are growing so fast and eventually all go enterprise. Replit, ElevenLabs and Vercel are among those building sales teams. He imagines SDRs paid $250,000 who manage ten agents instead of ten people, compared with $80,000 to $90,000 today, but warns the profession is not evenly distributed and old ways put you at risk.

He is exuberant about building: he says he has built 12 apps on Replit in about 150 days despite not coding, and is impressed that agents now debate each other to solve problems. His advice is to pick one of two paths for next year, working harder at a fast-growing company or joining a slower one that still needs fewer people, because he sees no middle path.

For skeptics he proposes an incognito test: over the holidays, use your own product, support and sales funnel with a fresh email address, and fix whatever makes you wince first with an agent. On job fears, he says to be honest that the best people get more productive, noting SaaStr never fired anyone; the old reps quit and agents backfilled. His closing advice for founders with happy customers is not to quit, because it is still early.