AI Explained
The AI stack and typical integrations
What sits under every AI product, and how companies actually plug it in.
The stack
Seven layers, from chips to apps
Every AI product sits on the same layers. Open each one for what it is, who lives there and the question to ask about it. Most teams build only the top four and rent the rest.
In the real world
Six ways companies wire AI in
Most AI projects are one of these, or two of them together.
Assistant over your documents
Staff ask questions and get answers with the source page.
How it is wiredDocuments are indexed; a question retrieves passages; the model answers from them.
WatchCarry each document's access rules into the index.
Assistant inside the tools people use
AI in email, chat, the CRM or the help desk.
How it is wiredThe vendor's own AI feature, or a connector that acts as the signed-in user.
WatchCheck what each connector may read, and what the vendor keeps.
Coding agent in the repository
An agent that reads the code, makes changes and opens pull requests.
How it is wiredA coding agent on the developer's machine or in CI, with project rules in a file it reads.
WatchReview its changes like anyone's, and keep secrets out of its reach.
Agent that handles requests
Refunds, password resets, ticket triage.
How it is wiredA model in a loop with a few narrow tools, and approval for anything that cannot be undone.
WatchLimits in code, an identity of its own, a way to stop it.
AI step in a data pipeline
Classify, extract or summarize thousands of records overnight.
How it is wiredA scheduled job calls the model once per record, with a fixed prompt and a check on the output.
WatchSample the results, and watch the cost.
Voice and meetings
Transcribe a call, summarize it, draft the follow-up.
How it is wiredSpeech to text, then a model on the transcript.
WatchConsent to record, and where the recording is kept.