Turbo AI PM
Kits / workflow

RAG / knowledge assistant

Answers grounded in your own documents. Work top to bottom, or jump to the decision in front of you — each one links to the metrics, tool, and playbook that answer it.

The riskConfident answers that aren't in the source — or are in the source but don't answer the question.
Used in product scenariosRecommendations & Personalization Search & Discovery Customer Support
Take it to your PRD, Notion, or launch review.

The chain

What has to be true at each layer for this product to create value. The hypotheses between layers are yours to hold — and to check after launch.

Model / SystemIs the AI behaving correctly?
R-01 Recall@k↔ R-02R-05 Groundedness / Faithfulness↔ R-07R-07 Answer Relevance↔ R-05
If the right passages are retrieved and answers stay grounded, users trust answers enough to act on them.
Product / UserIs the user accomplishing the task?
B-02 Task Completion Rate↔ B-06B-06 Human Override / Edit Rate↔ B-07B-03 Deflection / Containment Rate↔ B-05
If users resolve questions on their own, tickets and time-to-answer drop.
BusinessIs the product creating meaningful value?
Support cost per resolved question · time-to-answer

↔ marks each metric's shadow metric — the one to watch alongside it so the number can't improve while the product gets worse.

01

Before you build

Should this be AI at all — and which capability fits?

Before scoping a model, check that the problem has a pattern to learn, data you can reach at decision time, and a cost of being wrong you can live with. If a rule would get most of the value, the rule is the better product.

For RAGThe test for RAG: does a corpus exist that actually contains the answers, and can you keep it current? If the answers live in people's heads, retrieval has nothing to retrieve.
Check whether this needs AI →
Evidence, business implication, deeper reading
Evidence you need: Examples of the decision being made today, where the data lives and whether it is available at decision time, and an honest estimate of what a wrong answer costs.
Business implication: Choosing AI for a problem rules could solve buys you cost, latency, and a maintenance burden with no extra value.

Check whether the unit economics can work

Estimate cost per completed task against the value of that task before building — margins are set by architecture, not tuning.

MetricsC-02 Cost per Resolved Task↔ B-02C-01 Cost per Inference↔ P-01C-05 Context Window Utilization↔ R-06
Estimate ROI & unit economics →
Evidence, business implication, deeper reading
Evidence you need: Volume estimate, token estimate per task, model pricing, and the human cost (or revenue) per task.
Business implication: A feature that works but destroys gross margin is a failed product.

Make the case to build it

Ask for one specific decision by a specific date: the problem in their units, why AI and not rules, the expected value with its assumptions visible, the risk said out loud, and the checkpoint you will report back at.

MetricsB-02 Task Completion Rate↔ B-06
Build the business case →Make the case & defend the work →
Evidence, business implication, deeper reading
Evidence you need: The baseline (what today costs and achieves), a value range with the two assumptions that move it most, cost per completed task vs. the human cost, the rules-vs-model comparison, and the top failure modes with their mitigations.
Business implication: A named checkpoint turns a bet into a staged decision — much easier to approve than an open-ended one. For this product: Support cost per resolved question · time-to-answer.

Define what the model may do, must never do, and needs to know

Write the operating envelope before anyone writes a prompt: context the model sees, unforgivables, and what happens when it fails.

For RAGSpecify the corpus, freshness rules, and who may see which documents.
Write the AI PRD →
Evidence, business implication, deeper reading
Evidence you need: Agreed context spec, a written list of unforgivables, and a fallback for each failure mode.
Business implication: Scope decided here drives both cost and risk for the life of the feature.
02

Evaluate

Choose what to measure

Pick one metric per layer and a shadow metric for each, so no number can improve while the product gets worse.

MetricsR-01 Recall@k↔ R-02R-02 Precision@k↔ R-01R-05 Groundedness / Faithfulness↔ R-07R-07 Answer Relevance↔ R-05S-01 Hallucination Rate↔ S-03C-05 Context Window Utilization↔ R-06
Calculate any of 38 metrics →
Evidence, business implication, deeper reading
Evidence you need: The starter metrics below, each paired with the metric that catches it being gamed.
Business implication: The layer-3 metric is the one leadership will ask about — name it now. For this product: Support cost per resolved question · time-to-answer.

Establish a baseline worth beating

A number means nothing without a comparison. The simple heuristic and the incumbent are the two baselines that get skipped.

For RAGKeyword search (BM25) over the same corpus — the heuristic a RAG system has to beat.
MetricsR-01 Recall@k↔ R-02B-02 Task Completion Rate↔ B-06
Evidence, business implication, deeper reading
Evidence you need: Zero-rule, heuristic, and incumbent scores on the same eval set, measured the same way.
Business implication: If a heuristic gets 80% of the value, the model has to justify its extra cost.

Design an evaluation you can trust

Representative data, a scoring method that matches the task, and a holdout nobody tunes against.

For RAGScore retrieval and generation separately so a wrong answer can be traced to the stage that caused it.
MetricsR-01 Recall@k↔ R-02R-05 Groundedness / Faithfulness↔ R-07R-07 Answer Relevance↔ R-05
Evidence, business implication, deeper reading
Evidence you need: A golden set with an owner and a size target, slices defined up front, and a judge calibrated against humans.
Business implication: A weak eval means you ship blind and cannot explain failures afterward.
03

Ship

Decide whether it is ready to ship

The bar depends on blast radius and reversibility, not on the accuracy number alone.

For RAGInternal knowledge tools usually land in the amber quadrant: wide reach, but reversible when answers cite sources.
MetricsR-01 Recall@k↔ R-02R-05 Groundedness / Faithfulness↔ R-07
Decide go / no-go →Make the case & defend the work →
Evidence, business implication, deeper reading
Evidence you need: Your quadrant, the threshold for it, the eval result against the baseline, a working fallback, and the numbers you promised at the funding checkpoint.
Business implication: A documented bar turns post-launch blame into post-launch learning.

Design what users see when it is wrong

Every AI flow needs a designed failure state: specific copy, a way to recover, and a human path.

For RAGWhen retrieval finds nothing relevant, say so and offer a next step — never generate from empty context.
Design for non-determinism →
Evidence, business implication, deeper reading
Evidence you need: Fallback copy for timeouts, low confidence, out-of-scope requests, and refusals.
Business implication: Users forgive visible, recoverable failures; they abandon silent ones.
04

In production

Set up monitoring that separates "the model changed" from "the world changed"

Alert on breakage, review drift daily, review quality weekly, review business impact monthly.

MetricsR-01 Recall@k↔ R-02R-05 Groundedness / Faithfulness↔ R-07R-07 Answer Relevance↔ R-05
Evidence, business implication, deeper reading
Evidence you need: Error and latency alerts, input-mix and confidence drift, slice quality, and the layer-3 metric on a monthly cadence.
Business implication: Degradation you catch in a week costs far less than one you find in a quarterly review.

Turn production signal into a better next version

Design what each interaction captures and where it goes — implicit, explicit, and outcome signal feeding the eval set.

For RAGLog unanswered and corrected questions — they are your next golden-set entries.
MetricsB-02 Task Completion Rate↔ B-06B-06 Human Override / Edit Rate↔ B-07
Evidence, business implication, deeper reading
Evidence you need: A named owner for the signal pipeline and a path from user action to golden set.
Business implication: The flywheel is the only part of an AI feature that compounds over time.
!

Something looks wrong

Start from what you're seeing. Each symptom points to its most likely cause and the metrics that confirm it.

Answers sound right but cite the wrong policy
Retrieval returns near-miss documents and the generator faithfully summarizes the wrong source.
CheckR-02 Precision@k↔ R-01R-06 Context Relevance↔ C-05

Determine whether retrieval is causing poor answers: high faithfulness (R-05) with low precision@k (R-02) points at retrieval, not generation.

Grounded answers, users still unhappy
Answers are faithful to the context but do not address what was asked.
CheckR-07 Answer Relevance↔ R-05B-06 Human Override / Edit Rate↔ B-07
Costs creeping up with flat traffic
Context stuffing — more and longer chunks per query than the answer uses.
CheckC-05 Context Window Utilization↔ R-06C-02 Cost per Resolved Task↔ B-02

Other kits

Agents & multi-step pipelines Copilot / drafting tool Customer-facing chatbot / support Classifier / triage & moderation Coding assistant / IDE copilot