Which companies are ready to buy AI tooling?

Best move: score prospects on what their stack runs, not on what their funding round says.

Why it works: a company that already ships AI has the budget, the team, and the integration surface to buy more. That commitment lives in running code, and running code is hard to fake.

Key takeaways

• The companies ready to buy AI tooling are the ones already shipping it: a live MCP endpoint, an llms.txt, a vector store, or three or more AI signal groups firing together.
• Bassethound classifies every domain as shipping, experimenting, or none, and scores readiness as distinct signal groups hit over six, capped at 1.0.
• The icp_signal field (hot, warm, cold) ranks a target list by buying readiness before you make a single call.
• A detected stack beats funding, headcount, or job postings as a buy signal, because infrastructure is expensive to stand up and pointless to fake.
• A cold verdict means no outside-visible AI, not no AI: server-side or fully proxied stacks stay invisible, and Bassethound flags that gap in the dossier.

What makes a company “ready to buy” versus just curious?

Ready-to-buy maps to the shipping verdict. A company shipping AI has already crossed the hard part: standing up infrastructure, integrating a model provider, putting an agent or a retrieval pipeline in front of users. That company has a budget line, an owner, and a reason to buy adjacent tooling (observability, evals, a vector store, a security layer). A curious company has a chat widget and nothing behind it.

Bassethound draws the line with the verdict field. Shipping means a machine-facing endpoint (MCP, a .well-known AI file, or a technical llms.txt) is live, or that three or more distinct signal groups fire together. Experimenting means at least one signal fires. None means the stack is dark. The score (distinct signal groups detected, out of six) tells you how deep the commitment runs.

A company at 5/6 with a live MCP endpoint is not evaluating whether to build with AI. It already did, and it maintains that decision every day the endpoint stays up. That is who buys. The gap between shipping and experimenting is the gap between a customer with a budget and a lead you have to educate first.

Why is a shipping verdict a better buy signal than funding or headcount?

Funding tells you a company has money. It does not tell you where the money goes. Headcount tells you a company is growing. It does not tell you it grew an AI team. Job postings are aspirational and gameable: a company posts an “LLM engineer” req the day it decides to try, not the day it ships. All three are lagging or noisy proxies for the one thing you want to know, which is whether this company builds with AI right now.

A shipping verdict is the thing itself. A live MCP endpoint is a company telling you, in running code it has to keep alive, that it builds for agents. A Pinecone client in the bundle means someone wired up retrieval and pays for it every month. A LangSmith or Helicone trace means someone cares enough about the pipeline to watch it. These signals cost real engineering time to stand up, and faking them buys you nothing. You cannot post a vector store to look busy.

That is why a detected stack outranks any firmographic score for AI-buying intent. Firmographics rank the company. The stack ranks the commitment. You sell to commitment.

How does the ICP signal turn into a prospect list?

Bassethound emits icp_signal on every domain: hot, warm, or cold. Hot is a shipping company with depth across the layer. Warm is experimenting, one or two groups firing, budget forming. Cold is a dark stack. Run your target accounts through the AI-readiness layer, sort by icp_signal, and you rank the whole list before writing a single email.

The move is to pair the signal with what it detected. A hot account running LangChain and no observability is a lead for an observability product. A hot account with a vector store and no eval tooling is a lead for evals. A shipping company with model providers but no security signal is a lead for a guardrail layer. The verdict tells you readiness. The layer contents tell you the pitch.

You are not spraying a list. You are matching a gap. That changes the first email from “do you use AI” to “you run X and you are missing Y,” which is the difference between a cold open and a warm one. The AI-readiness layer does the segmenting; you do the selling.

What can’t you see from the outside?

Honesty matters here. Detection is a static crawl plus optional JS rendering. It reads what the front end loads and calls. A purely server-side integration (a model invoked from a backend that never touches the browser) can stay invisible. A fully proxied setup (all AI traffic routed through the company’s own domain) hides the vendor behind it. A private internal tool never appears on the public site at all.

So a cold verdict is not proof of no AI. It is proof of no outside-visible AI. Bassethound reports these gaps in the dossier instead of pretending the crawl saw everything. Treat cold as “unconfirmed,” not “empty.”

The signal is strongest when it fires positive. A detected MCP endpoint is near-certain, because it has to be reachable to function. A named vector-store hostname in a network call is a company paying a bill. Absence is weaker evidence than presence, and the read is asymmetric on purpose. Trust a hot signal to prioritize an account. Do not trust a cold one to disqualify a large target without a human look. Read the layer knowing which direction it is reliable in.

How do you qualify one prospect fast?

Point Bassethound at the domain. One keyless call returns the five-layer dossier, AI-readiness included. Read three fields in order: verdict, score, icp_signal. Shipping at 4/6, hot, means you skip the discovery question of whether they build with AI and go straight to the gap. Experimenting, warm, means they are early and worth a nurture, not a hard close. None, cold, means deprioritize or dig by hand, since the crawl may have missed a server-side stack.

Then read the layer contents for the wedge. The model provider tells you their vendor loyalty and where a switch would hurt. The orchestration framework tells you how they build and what they would integrate against. A missing observability, eval, or security signal is your opening line.

The whole read takes one call and under a minute. No key of your own, no waiting on a sales-intel vendor to refresh a quarterly dataset. Sniff the domain, read three fields, name the gap. That is a qualified prospect before the first touch.

Bassethound perspective

Intent-data vendors sell you a readiness score built from funding, headcount, hiring, and web-tracking pixels. We think that model is backwards. Those inputs measure attention, and attention is cheap. A company posts an “LLM engineer” req the week it gets curious, and the intent score lights up months before anyone ships something you can sell into.

A running machine-facing endpoint is different. When a company stands up an MCP endpoint or publishes a technical llms.txt, it is telling you, in code it has to maintain, that it builds for agents. That is not intent. That is a decision already made, paid for, and deployed. It is the single strongest AI-buying signal on the open web, and almost nobody is reading it yet.

So Bassethound sniffs the stack, not the story. We would rather return one honest “shipping, hot” verdict grounded in a detected endpoint than a hundred inflated intent scores. When the crawl cannot see behind a proxy, we say so. Digby does not bark at empty houses. Presence is proof. Absence is a question. Sell to presence.

Sources

• Model Context Protocol: https://modelcontextprotocol.io
• llms.txt specification: https://llmstxt.org
• Pinecone documentation: https://docs.pinecone.io
• LangChain documentation: https://python.langchain.com
• Anthropic documentation: https://docs.anthropic.com