www.iqsource.ai Lemkin Swapped 10 Reps for 20 Agents and 1.2 Humans
Ricardo Argüello, September 7, 2026
CEO & Founder
General summary
Jason Lemkin, founder of SaaStr, replaced a 10-person sales team with more than 20 AI agents run by the equivalent of 1.2 people. The part that rarely travels with the headline is the real operating cost: roughly 30 days of intensive training per agent up front, weekly maintenance after, and a named Chief AI Officer spending about an hour a day on the stack.
- SaaStr went from 10 SDRs and account executives to more than 20 agents run by 1.2 people
- Amelia Ibarra, titled Chief AI Officer, spends roughly an hour a day operating the agent stack
- Each agent required about 30 days of intensive training before it was useful, plus weekly upkeep
- Of the 20-plus agents, about 12 represent what the operation can genuinely sustain
- Lemkin's stated lesson is that training matters more than picking the perfect vendor
Instead of hiring ten salespeople, you hire twenty interns who work for free and never sleep. Each one needs a month of supervision before it is useful, and somebody has to review all of them every day, forever. The cost stopped being payroll and became one person's attention. That is the operation SaaStr built, described with its own numbers.
AI-generated summary
Twelve.
Not twenty. SaaStr runs more than twenty AI agents in its go-to-market stack, and admits that about twelve is what the operation can actually sustain with the maintenance it has.
That admission is why I trust the rest of the account.
The headline and the invoice
Jason Lemkin had ten people between SDRs and account executives. Today the work runs on 20-plus agents managed by the equivalent of 1.2 humans.
The headline writes itself and teaches almost nothing.
What teaches something is what he published about keeping it running. Each agent needed roughly 30 days of intensive training before it was useful. Not configuration, training: correcting output, reviewing answers, explaining what was wrong and why. Then weekly maintenance, permanently.
His summary: training matters more than picking the perfect vendor. Thirty days up front. Commit the time. Every day.
Multiply it out. Twenty agents at thirty days of supervision each is not a technology project. It is onboarding an entire graduating class, except these hires do not learn by watching the person next to them.
The job title that makes it work
Amelia Ibarra holds the Chief AI Officer title at SaaStr and spends about an hour a day across the stack: which meetings got booked, what outbound is producing, whether support answers hold up, what the deal summaries say.
One hour, daily, from someone with judgment.
That is the cost almost no business case includes. Teams compute payroll savings and set them against a software license. The human attention required to keep the system from decaying appears in no cell of the spreadsheet.
And without that owner, agents do not fail loudly. They decay slowly, keep answering, and nobody finds out until a customer complains about something that started three months ago.
We argued why this is a role rather than a title in the agent maestro is a role, not a job title.
One detail that serves no argument
Lemkin mentions that a rep quit the day they deployed the deal intelligence agent, because the system made visible that he had no activity.
That is not a productivity lesson. It is what happens when measurement suddenly arrives somewhere it never existed, and it is worth knowing before you deploy anything that reports on people.
Building the honest business case
If you are evaluating agents for a commercial operation, add three lines your model probably lacks.
Thirty days of training per agent, priced at the cost of the person doing it. That time is not free and it does not compress.
One hour daily of a person with judgment, permanently. Name that person before signing anything. If you cannot name them, do not start.
And a realistic ceiling: how many agents that hour can carry. If the answer is six, launch six, not twenty, no matter how many the platform allows.
On the gap between having agents and running an operation with agents, we wrote the agent directory versus the org chart.
Lemkin pulled it off and it worked. He also published the costs, which is the rare part. If you copy the twenty agents, copy the hour a day too.
Let’s put an owner and an attention budget on your agentsFrequently Asked Questions
SaaStr runs more than 20 AI agents across its go-to-market stack, managed by the equivalent of 1.2 people, after previously operating with 10 SDRs and account executives. Of those 20-plus, the company acknowledges roughly 12 are what the operation can genuinely sustain with available maintenance.
Based on the account SaaStr published, each agent needs around 30 days of intensive training before it contributes, plus ongoing weekly maintenance. Jason Lemkin's stated lesson is that training matters more than picking the perfect vendor, because the same vendor performs very differently depending on that upfront work.
A named person with allocated time. At SaaStr the role exists formally as Chief AI Officer and takes about an hour daily: checking booked meetings, monitoring outbound, quality-checking support answers, and reading deal summaries. Without an explicit owner, agents degrade quietly and nobody notices until a customer complains.
It reduces payroll and substitutes a different cost: the sustained attention of one person with judgment, plus the training time each agent requires. The math can land strongly in your favor, but a company that runs it expecting savings without work ends up with badly performing agents and nobody watching.
Related Articles
Thomson Reuters Trained Its Own Model for $450,000
Thomson Reuters built a legal model that beats GPT 5.4 and Claude Sonnet 5 on its own content. The final training run cost $450,000. The corpus is the moat.
Anthropic's $30 Trillion Market Is Your Payroll
The addressable market Anthropic is pitching investors is not measured in software licenses. It is measured in the human labor its models could replace.
OpenAI Benchmarks Jalapeño Against Nvidia's Racks
OpenAI published first results for its Broadcom-built inference chip: more throughput and lower latency at once, with 40 to 60% less power. Vendor numbers.
Google Entered Legal and Harvey Became a Connector
Gemini Enterprise for Legal shipped August 25 with Cleary, Freshfields and Weil. Harvey appears inside it as an MCP integration rather than a rival.
Porsche Sold Its Consulting Arm, Then Rented It Back
Porsche sold MHP to TCS for 320 million euros and signed a five-year partnership worth 1.25 billion. The accounting works. The learning walks out.
OpenAI Cuts Cursor Off Over a Change-of-Control Clause
OpenAI is ending Cursor's model access on November 12 because SpaceX bought it. Cursor survives because OpenAI was only 5% of its traffic. What is your number?
