Your $5,000 AI Agent Needs 20% of an Engineer to Keep Working

Vercel’s lead-qualification agent costs the company about $5,000 a year to run. It also costs 20% of an engineer to keep alive. The second number is the one that never makes it into the business case.

Vercel COO Jeanne DeWitt Grosser laid out the math at SaaStr’s AI event this year. Her team replaced a 10-person lead-qualification function with one person in the US plus a fifth of another covering Europe and Asia. A separate support agent now handles 93% of the case load. SDR quotas climbed 30% in a quarter. She put the return at 32x. Then she said the part that gets left off the keynote slide.

“A demo that works is not a system that works.”

That gap is where most marketing AI is stuck right now. The model is fine. The demo dazzles. Then real inputs show up and the thing falls over.

Sadie St. Lawrence, a consultant at HMCI, handed a batch of agents real business jobs and wrote down what broke.

“The moment something unexpected happens—an email formatted slightly differently, a field missing in the CRM, an edge case the agent wasn’t trained on—it breaks.”

And it breaks without telling you.

“The agent doesn’t tell you it messed up. It just quietly produces a wrong output, and you don’t notice until a client emails you.”

Kayode Faturoti, co-founder of Liners, ran the same experiment across a company’s worth of tasks and landed in the same place. Agents are strong on anything close to code and structured work, he found, and weak anywhere judgment is required. Getting them to hold up meant writing out every exception by hand.

“Every task turns into a long list of dos and don’ts, with edge cases spelled out and exceptions named one by one.”

Here’s what’s actually going on. An agent is software you give a goal, and it works out its own steps to get there instead of following a fixed script. That flexibility is the whole pitch, and it’s also the failure mode. A fixed script breaks loudly when the input changes: you get an error, you fix it. An agent improvises, so when the input changes it doesn’t stop. It guesses. Sometimes the guess is fine. Sometimes it invents 400 venture firms that don’t exist. That’s what SaaStr’s own marketing agent did, until someone asked it to name them and it admitted it made them up. No alert fires for that. You find out when a customer does.

Which is why the maintenance line is the actual job. Running one of these agents means correcting it, over and over. Jason Lemkin, who runs SaaStr, watched his own build quietly stop pulling the YouTube transcripts it depended on. Nothing broke loudly, so nobody noticed. His rule now:

“A five-minute agent that you correct daily for a month outperforms a sophisticated build you set and forget.”

Add it up and a marketing agent’s real cost sits well past the subscription. There’s the token meter, where the AI bills by the volume of text it reads and writes. And there’s the person watching it. Faturoti puts the first half plainly.

“Agents run on tokens, and tokens are cash.”

The second half is the salary you hand to whoever minds the thing. Grosser’s 20% of an engineer. St. Lawrence’s client-caught errors. The reviewer Faturoti says every deployment now needs, because the work didn’t vanish, it moved.

“They do not remove the work so much as change its shape: less doing and more directing, reviewing and correcting.”

None of this means the agents don’t earn their keep. On narrow, well-fenced, low-judgment jobs they clear a bar people won’t. Lemkin rates his AI outbound emails a 3 to a 6 out of 10 on customization and ships them anyway. Decent emails sent with zero errors at scale beat a human’s occasional brilliance. On one batch of event leads his own reps refused to touch, the agents pulled in 15% of the London ticket revenue. The lesson isn’t “don’t automate.” It’s that the wins came from picking a job small enough to fence, and staffing someone to mind the fence.

We’ve watched this from the vendor side too: an agent built to run your marketing end to end also runs a meter, and the companies posting the cleanest AI wins are the ones that scoped the job down until it was boring. The firsthand builders are now saying the same thing from inside their own stacks.

So before the next marketing agent goes into the martech budget, price the whole thing:

1. Scope it narrow. The 93% agent works because someone drew the 93% line. Agents fail at the edges, so give them a job with fewer edges.
2. Name an owner. Not a committee. One person whose job includes catching the silent wrong answer before the client does.
3. Budget the minder. If a $5,000 agent needs a fifth of an engineer, that engineer is the real price. Put it on the line item.

The teams getting value out of AI right now share one habit. They admitted the demo was the easy part, and they staffed for everything that comes after it.