How do you tell if a company is shipping AI in production?

Best move: judge the stack a company runs, not the AI it advertises.

Why it works: production AI leaves infrastructure fingerprints that a marketing page never does. A company running retrieval and agents has to load the libraries, call the providers, and stand up the endpoints to do the work, and those are visible from the outside.

Key takeaways

• “Shipping AI” is an infrastructure question, not a messaging one.
• The strongest single tell is a live MCP endpoint or a technical llms.txt, because those exist to serve machines, not visitors.
• Bassethound returns a three-way verdict: shipping when real production signals fire, experimenting when a few do, none when the stack is empty.
• The score counts distinct signal groups (model providers, SDKs, orchestration, vector stores, observability, llms.txt and MCP), so one keyword cannot fake it.
• A shipping verdict is a strong ICP signal for anyone who sells AI tooling, infrastructure, or services.

What separates shipping from experimenting?

Depth and production intent. Experimenting looks like one shallow signal: a mention, a standard llms.txt, a single chat widget bolted onto a marketing site. Shipping looks like several signal groups firing together, and at least one of them built to serve machines. A company with a wired-in model provider, a vector store, and LLM observability watching both is not testing the water. It is running AI it cannot afford to leave unmonitored.

The verdict draws the line at that difference. One signal is a hint. A cluster of them, with a machine-facing endpoint in the mix, is production.

Which signals prove production AI?

Six groups carry the weight. A named model provider (Anthropic, OpenAI) wired into the app. An AI SDK loaded in the client. An orchestration framework such as LangChain or LlamaIndex. A vector store holding embeddings. LLM observability such as LangSmith or Helicone, which is the strongest of all because you only monitor what you run in production. And the machine-facing endpoints: a technical llms.txt, a .well-known AI file, or a live MCP endpoint.

Observability and a live MCP endpoint are the signals that most reliably mean “in production.” Nobody wires up LangSmith or exposes an MCP endpoint for a demo.

Why isn’t llms.txt alone enough?

Because most llms.txt files are marketing. A standard llms.txt is a friendly summary for models, useful but shallow, and easy to add in an afternoon. It says a company has heard of AI. It does not say a company runs it. A technical llms.txt, one that points to real endpoints and specs, is a different animal, and Bassethound classifies the two so a shallow file does not inflate a score. Treating any llms.txt as proof of AI adoption is exactly the mistake that makes shallow readiness scores disagree with reality.

How do you score AI adoption from the outside?

Count distinct signal groups and weight the ones that imply production. Bassethound tallies how many of the six groups fire, normalizes that into an ai_readiness score, and sets the verdict: shipping if a machine-facing endpoint is present or several groups fire, experimenting if at least one does, none if the stack is empty. The scoring is transparent on purpose. A buyer should be able to see why a domain scored the way it did, not trust a black box.

Can a company fake shipping AI?

Not cheaply. A hero headline about AI costs a sentence. A model provider plus a vector store plus observability plus a live MCP endpoint costs an engineering team and a reason to keep it running. That asymmetry is the whole point. Marketing is cheap to fake and worthless as a signal. Infrastructure is expensive to fake and therefore worth reading. Reading the second thing is how you avoid mistaking a landing page for a product.

Bassethound perspective

Every vendor now says it is AI-powered. The claim is noise. The stack is signal. A company that has wired in a model provider, stood up a vector store, and is watching both with LLM observability is shipping AI whether or not it says a word about it, and a company with an AI headline and nothing behind it is not, no matter how loud the headline.

Bassethound reads the second thing. Shipping, experimenting, or ignoring AI is a verdict you can act on, and it is the one signal that predicts whether a company is a real buyer for AI tooling and services. That is why the icp_signal rides on top of the AI-readiness layer. The readiness verdict is not a vanity metric. It is the ICP filter.

Sources

• Model Context Protocol specification: https://modelcontextprotocol.io
• The llms.txt proposal: https://llmstxt.org
• LangSmith documentation (LLM observability): https://docs.smith.langchain.com
• Helicone documentation (LLM observability): https://docs.helicone.ai
• Anthropic API documentation: https://docs.anthropic.com