Session 9 — Building AI Products

📋 Canvas produced for AI functionality

Complete all 6 sections to turn a vague idea (“let’s have AI!”) into one defensible functionality. The summary is generated at the bottom — copy it into your exercise rendering.
Trick If you resist a section, it is often a sign that the idea is not mature.

0 / 6 sections filled

1. Problem — What specific problem? For whom? What measurable pain today? 2. User — Main persona: role, technical level, context and frequency of use. 3. Role of AI — What exactly does the AI do? Where does it stop? 4. Role of humans — Position on the “human in the loop” spectrum and justification. 5. Metrics — 3 measurable metrics: completion / time saved / confidence / escalation rate. 6. Risks — 3 risks (including ≥ 1 of responsible AI: bias, confidentiality, consent) + mitigation.

📄 Canvas summary

💰 ROI Calculator (Return On Investment)

Compare the cost of 100% manual treatment to the cost of the same AI-assisted treatment, taking into account the precision : cases missed by AI are escalated to humans (full manual processing) and successful cases still require quick proofreading.
⚠ API prices change quickly — check your provider's current prices.

Manual processing

Manual time per unit (minutes) Loaded hourly cost (€/h) Monthly volume (units: tickets, emails, documents, etc.)

AI solution

AI cost per unit (€) — API call ⚠ Monthly fixed cost (€) — infrastructure, supervision AI accuracy (%) — the rest is escalated towards humans Human proofreading per successful unit (minutes) — 0 if no proofreading Initial integration cost (€) — only once

🔬 To test yourself

Two experiences to understand what really drives profitability :

Triple the API cost (€0.08 → €0.24): the economy is barely moving.
Drop the accuracy (85% → 60%): The economy collapses — each escalation reverts to full manual processing.

Lesson : pilot it escalation rate , not price negotiation.

🌳 Build, Buy or Fine-tune?

Answer the questions to get a recommendation. Buy = use an off-the-shelf API (Application Programming Interface) · Fine tune = refine an existing model on your data · Build = build a model from scratch.

🧭 Reminder: reading order

We go down the tree floor by floor — we don't skip any steps:

Buy first — this is the reasonable default for ≈ 80% of cases: fast, low maintenance, predictable cost.
Fine-tune on proof — only if accuracy plateaus despite good prompting, AND you have the data, AND the volume justifies the cost.
Build almost never — total sovereignty required, or the model IS your product, with the budget to match ⚠.