A $1B robot maker lets AI find the parts and plain code handle the motion. Build your CRM automations the same way

In its machine tending system, a factory robot company that cannot afford a wrong move uses AI for one narrow job and code for the rest.

What happened

Latent Space interviewed Standard Bots, which makes AI driven industrial robot arms for tasks like machine tending, welding and assembly. The company raised $200 million at a $1 billion valuation in June, according to SiliconANGLE, and lists customers including NASA, Amazon and Lockheed Martin.

Its head of AI, Leif Jentoft, described the split in its machine tending system. A zero shot model handles perception, locating and identifying the parts. Conventional programming handles the motion and the cell logic around it.

CEO Evan Beard said the company focuses on short horizon tasks to meet production requirements for cycle time and reliability, and uses conventional programming for parts of a workflow that do not need learned behavior. The company captures failure signals and human corrections from deployments, and Jentoft said in situ interventions can fix an edge case with a few dozen examples. Inference runs on site because, as Jentoft put it, most factories do not have reliable internet.

Latent Space closes with the point that the same approach applies to software agents.

My take

This is the design I push for in CRM and ops automations, and it is good to see it hold up where mistakes break hardware.

Give the model the fuzzy part: read an inbound email, work out intent, pull the name, budget and timeline into fields. Then hand off to plain workflow logic for everything that has a right answer: which pipeline stage, which owner, when the follow up fires, what happens if a field is empty.

Two habits from the article are worth stealing:

1. Keep tasks short. One extraction step is easy to test. A twenty step agent run is not.
2. Log corrections. When someone fixes an AI decision, save the input and the fix. A few dozen of those make a far better prompt or test set than guessing.

That split is the backbone of what I build: AI for reading and judging, code for routing, updates and timing. Lead intake, CRM hygiene, reporting and agent workflows that behave the same on Monday as on Friday. See examples at romielwillautomate.dev.