Forward-deployed / Learning zone
Technical product sensea standalone module
Lesson 09

Technical sense for AI systems

TL;DR

An AI feature is a normal distributed system — everything in this module still applies. It has one unusual component wired in: a probabilistic model, often a third-party API, that's slow, priced per token, and occasionally confidently wrong. Technical sense for AI means knowing the anatomy of that system (guardrails → retrieval → prompt → model → validate → tools), where its latency and cost live, how it fails, and how you'd ever know it's working. That's why evals and observability aren't optional extras — they are the reliability of the feature. The model is the easy part to add. The system around it is the product.

🎯 For the AI PM

Why it matters — This is where every prior lesson converges. Architecture, APIs, data, latency, reliability, and debt all take on an AI-specific twist at once. Miss the system and you've shipped a demo, not a product.

What it changes in your decisions — Scope the feature to what the system can make reliable and affordable, not just what the model can do in a demo. Fund the unglamorous parts — retrieval quality, evals, observability, guardrails — as the feature itself.

Ask yourself — "What's the full path around the model call, and which part — not the model — is most likely to make this feature fail?"

Risk if ignored — The classic AI-product failure: magical in the demo, untrustworthy, unobservable, and unaffordable in production.

The anatomy of an AI feature

The model call is one box in a pipeline. The product lives in the boxes around it:

The anatomy of an AI feature

The model call is one box · the product lives in the boxes around it

Guardrails → retrieval → prompt → model → validate → tools — every step emits traces.

User input
Guardrails — input checks
Retrieval — fetch context
Prompt assembly
Model call
Validate / parse
↺ invalid output loops back to prompt assembly for repair
Tools / actions — valid output
Response to user
Observability + evals — traces, tokens, cost, and latency from the model call, validation, and tools, graded continuously. Without this, you are flying blind.

The four technical dimensions, AI-flavoured

Measuring "is it working?"

With deterministic software, correct is correct. With a model, quality is a distribution, so you have to measure it. An eval is a graded set of representative and adversarial cases the feature must pass, run continuously so regressions surface before users find them. Pair it with observability — traces of each step's tokens, cost, latency, and errors — so when quality drops you can see which box caused it. Together these are the AI Engineering track's evals and observability lessons. For an AI feature they are its reliability.

Scope to the reliable frontier

Model capability is jagged — brilliant at some tasks, unreliable at adjacent ones. The highest-leverage technical-product decision is scoping: point the model at the jobs it does reliably and inside your cost/latency budget, and use a deterministic path (or a human) for the jobs where a wrong answer is costly. Often the best AI product uses the model for the delightful 20% and boring, correct machinery for the 80% that must not fail.

A worked pass: the napkin for "AI answers support tickets"

Before the roadmap, do the arithmetic. Proposal: draft answers for inbound support tickets. Volume: 60,000 tickets/month. Per ticket, the context is the ticket thread plus retrieved help-center passages — call it 4,000 input tokens at p50. But check p95 too: long threads run 12,000, and cost scales with usage, so price the distribution, not the average. Output is about 500 tokens. At list prices for a mid-tier model, that's roughly a few cents per ticket — say $2–4k/month before caching. The system prompt and help-center boilerplate are identical across tickets, so prompt caching should cut real input cost hard. Verify the hit rate; don't assume it.

Against what baseline? If a drafted answer saves an agent 2 minutes at a loaded $40/hour, that's about $1.33 saved per ticket used — but only for tickets where the draft is accepted. Now the number that rules them all: acceptance rate. At 70% acceptance the feature prints money. At 25% it's a cost line plus an annoyance. That's why the first build artifact isn't the prompt — it's the eval: 100 real tickets, graded drafts, a pass bar agreed with the support lead. The napkin tells you whether to start. The eval tells you whether to ship. Neither requires writing a line of code, and both are the PM's job.

Failure modes

Practitioner checklist