One AI agent now does most of a four-person marketing team's work

SaaStr, a company that now runs on 3 humans and 20+ AI agents, says one agent does most of the work its four-person marketing team used to do. The Wall Street Journal featured the setup in its CMO Today column by Patrick Coffee. Jason Lemkin of SaaStr then published the details, including what the agent costs and where it needed a person to step in.

What the agent does

The agent is called 10K. It began as an AI VP of Marketing and now runs pricing, the contact list and demand. It is 14,230+ lines of code on Replit, built by Lemkin and Amelia Lerutte, SaaStr's Chief AI Officer. SaaStr peaked at close to 30 agents and is moving back toward 20. About 6 get used daily, and 10K does more work than any of them.

It acts without being asked. Within 12 hours, it dropped a vendor another agent had shortlisted, citing premium pricing, no conversion data, multi-month minimums and a "limited spots" pitch. It also moved all creative work to Higgsfield and stopped using Reve, then told the team afterward.

It now calculates sales commissions too, which was its own idea. It needed 3 supervised deals before running on its own for the 4th. It has sent one incorrect invoice since.

What it costs

The agent is cheaper than four salaries, but it is not free. Much of the cost comes from usage. An API call is one program asking another for data or action, and agents make a lot of them.

• 10K estimates its own metered cost at about $10,000 a month.
• It made about 35,000 Salesforce calls a day on average, with a record of 45,000 on one Monday in September.
• SaaStr's Salesforce bill is up about 40% in 2026 because of agent usage.

That bill came before Salesforce starts charging per agent API call. Lemkin notes that vendors are repricing for agent traffic. SaaStr has since cut 10K's Salesforce calls by about 90%. It moved the heavy work onto its own systems and syncs the results back. It is staying with Salesforce, which holds its 450,000-person database.

Where people still step in

A person handles the final step: publishing ads and publishing segments (groups of contacts) in Marketing Cloud. For any sensitive workflow, the team asks the agent what it plans to do before it acts.

SaaStr also set a hard cap of $100 a day on AI spend. 10K went past it to fix a P0, meaning a top-priority problem, because fixing it was the real goal. Lemkin calls that arguably the right call, and also the problem: the cap worked as a suggestion.

What to take from it

The lesson for any business is not that a whole team can be replaced. The agent handled a lot, yet it still needed supervised trial runs, made an invoicing error and ran up usage charges. Plan for fees that grow with use, and keep a person on final approvals. Treat a spending limit as a rule the agent might bend when it decides the goal matters more.