Your DIY GTM agent is burning money. And it doesn't even know it. 🔥

I read a paper recently that I can't stop thinking about.

Stanford and UMich studied how much agents actually spend running real tasks. And the numbers lined up almost exactly with what we keep seeing in agentic GTM.

Here's the part that got me:

Same task. Same model. Different runs. The cost swings 30X.

Not 30%. Thirty times.

And here's the kicker- the expensive runs aren't the good ones. They're the runs where the agent gets stuck in a loop. Re-reading the same files. Re-enriching the same contacts. Re-running the same searches it already ran two minutes ago.

Sit with that for a second. 🤯

You can't build a GTM engine on costs that swing 30X on the same input. You can't forecast. You can't estimate ROI. You can't tell finance a number you actually believe.

This is one of the quiet flaws in DIY GTM agents that nobody warns you about. ⚠️

You know the setup. GPT + a scraping layer + some glue code. It works beautifully on 10 accounts in a demo. Then you point it at 500 accounts and your token bill comes back 8X what you modeled. You have no idea which workflow is the money pit. No guardrails. Just vibes and a scary invoice.

The paper's core insight is the part worth tattooing somewhere: it's the input tokens, the context piling up- that drive the cost. The agent rediscovering and re-ingesting the same context on every single loop.

This is exactly why we built Tapistro the way we did. 🎯

✅ We unify and pre-organize your signals into structured profiles before the agent ever runs, so the model gets clean, curated context instead of going exploring from scratch every time

✅ We control what goes in, the agent doesn't go fishing through raw data, it reads from a pre-built intelligence layer

✅ And we make the costs deterministic. Behind the scenes, we do the work so one agentic step costs the exact same credits every time, change your question 10 times, same cost. You get a number you can actually plan around.

Control the context, control the cost. That's the whole game.

If you're building agentic GTM on raw API calls and hoping the economics just work out at scale, they won't. Not with 30X variance baked into the foundation. 💀

Curious what you're seeing on your end. Hit reply.

📄 Paper: "How Do AI Agents Spend Your Money?" - Bai et al., 2026