The AI feature is ordinary software

The final lesson of the Ground-Up Path has one job: to show you that the most intimidating technology of the decade is, architecturally, a system you already speak fluently.

This curriculum built toward this moment deliberately. So let’s reconnect, piece by piece.

Calling an AI API

Very few companies run models themselves, since lesson one’s GPU economics price most of them out. The coffee app is calling an AI APIConcept · lights on your mapAI APIA provider’s endpoint that answers prompts. The backend POSTs the prompt to it over HTTPS, key in a header, JSON both ways, under the usual status codes, timeouts, and retries. Module 6 taught every part of the call except who answers it., and here is what the call actually looks like on the wire.

Read what that request contains: a POST to an endpoint, a JSON payload, a key from the vault, and every rule of API etiquette you learned in Module 6. It is Mini-capstone A’s trace, with a model at the far end.

The provider is a third-party service

And the provider on the other side? That is AI as a third-party serviceConcept · lights on your mapAI as a third-party serviceAnthropic, OpenAI, or Google occupying the seat Module 5 built for Stripe. Rented capability under a versioned contract, with deprecation notices, pricing shifts, and outages that become yours. Every relationship rule you learned for vendors applies without edits., Stripe’s seat at the architecture table with a new occupant.

Every dynamic from the API economy transfers unchanged, and build-vs-buy is decided the same way. The contract is versioned, and a model deprecation notice, the warning that a version you depend on is being retired, is Module 6’s migration email, verbatim. And their bad day becomes your status page, exactly as Module 5 warned about every rented dependency.

The two new numbers: latency and cost

The operations story is also yours, and it drops Module 10’s tradeoff table onto new terrain with two loud new line items.

AI latencyConcept · lights on your mapAI latencyA model reply is measured on a human clock, whole seconds of it. Streaming exists to spend that wait visibly, and the craft shifts to how long the wait feels, since how long it is rarely moves.: inference takes seconds, not milliseconds, which is why answers stream in as they generate. Perceived latency becomes a UX lever when real latency won’t budge, and streaming is that lever pulled, since an answer that begins at once reads as fast even when it ends slowly. The slowest AI work, the report nobody is watching the screen for, takes Module 9’s queue like any other background job.

AI costConcept · lights on your mapAI costEvery request runs the token meter twice, once for what you send and once for what comes back. Spend scales with traffic, so trimming menu and history out of the prompt is money saved, not only quality traded.: every token is a line item, in and out, on every request. Prompt size (how much menu, how much history) is now a cost decision, not just a quality one.

Everything else runs the way Phase 6 taught. The feature emits logs, metrics, and eval scores onto Module 13’s dashboards, and its rollout hides behind Module 12’s feature flags. Shadow mode, another Module 12 resident, was practically invented for AI features. Running the assistant silently on real traffic, scoring everything and showing nothing, is exactly the audition probabilistic software needs.

What leaves the building

One flow deserves special attention because it is new in shape: AI privacy considerationsConcept · lights on your mapAI privacy considerationsThe Module 7 and 14 questions aimed at a brand-new flow, since every request now carries customer text outside your walls. Redaction, the provider’s retention and training terms, and the rules that travel with exported records all hang on one audit. Know what the prompt holds.. Prompts full of customer text now leave the building on every request.

The prompt is the vehicle. Whatever the feature packs in arrives on another company’s servers the moment the request goes out. The cargo might be a customer’s message, their account details, or an internal document pasted in as context.

Customer data has left buildings before, but it left through doors somebody opened on purpose, an export, an integration, a report. This flow is ambient instead. Code decides what goes into each prompt, decides again on every request, and no person watches it happen.

Modules 7 and 14 hand you the questions ready-made:

• What PII rides along in the prompt, and should it be redacted or kept out entirely?
• What may the provider retain, or train on? (A contract question, so read it.)
• Which regulations follow the data out the door?

Each question has a working answer. Redaction strips names and card numbers before the prompt ships, the contract settles what the provider may keep or learn from, and residency clauses keep regulated records inside their borders. None of it is exotic, and none of it happens by default.

“What exactly is in the prompt?” is this decade’s “does that export contain PII?”

The reconnection

Now stand back and draw the whole feature honestly. Each row below is one working part of it, and the label beside each part names the modules that taught you that part.

The model is new, but the system is the one you have been learning since “what is technology actually made of?”

That was the design all along, and it is why you’re ready for what’s next.

The mental model to remember

An AI feature is an API call: a POST with a JSON prompt, a vault key, and Module 6’s etiquette, to a provider sitting in Stripe’s old seat (rented capability, versioned contract, their outage on your status page).

Its two loud new numbers are latency (seconds, hence streaming) and cost (per token, where prompt size is a budget). Everything else lands on the dashboards and behind the flags you already know.

And its sharpest new question is a privacy question. What exactly is in the prompt that just left the building?

You should now be able to sit in any AI product meeting and hear it for what it is: a systems conversation you already speak, with one probabilistic newcomer, managed by the toolkit this module gave you.