Lesson 6 · 16 min
Your bank's credit approval doesn't go through a chat assistant, and that's not a delay: it's engineering. Almost everything that runs at scale is still classical AI. Someone who knows how to choose between generative, classical, hybrid, or no AI stops being a vendor and becomes an advisor.
Start the lesson
Goal: choose between generative AI, classical AI, hybrid, or no AI for a case, stating the reason and what stays out.
Four outcomes. None of the questions is "which technology is trending."
Below a certain volume, generative is great. Above it, cost per operation and predictability decide on their own.
Answering on the spot, with the reason, is what changes the way the client sees you.
Fraud detection, price optimization, demand forecasting and inventory optimization.
A mistake costs money directly, the decision has to be reproducible, and nobody wants a hallucinated price in the storefront.
Customer service and chat assistant, with the classic model routing the ticket or detecting the intent before answering.
Sentiment analysis today is almost all generative, with a keyword filter first: a legal-risk term goes straight to a person.
Content personalization (the classic picks the segment, the generative writes), churn prediction (one predicts who leaves, the other writes the message) and document processing.
A recommendation system is hybrid and labor-intensive: don't promise a short deadline.
A big vendor's experimental lab is fine for learning, not for the client's critical path. A tool that breaks completely on a version change is fine for a prototype.
An app that does everything is a doorway to understand what's possible, not the business's infrastructure.
For an overloaded client, a list of twenty tools creates paralysis.
Recommend four things, one per category: a chat assistant, a model for deeper analysis, an automation tool and an assisted-coding environment.
Each one opens weeks of learning, and the client's curiosity turns into a meeting agenda.
Knowing fifty tools on the surface means knowing the name of every one and the behavior of none.
Someone who knows the automation tool in depth knows, for example, that it already has built-in tables and saves you from creating one more spreadsheet and one more account.
Ask item by item: why this tool, what exactly does it give you?
Often the assistant the company already pays for does the same thing. The client cuts a license and sees your value before you build anything.
The board is for drawing with the client; you run the prompts before the scoping meeting.
CASE: <client's request in one sentence> Volume: <operations per day> Cost of an error: <what happens> Data: <structured / loose text / mixed> DECISION: [ ] generative [ ] classic [ ] hybrid [ ] no AI REASON: <the criterion that decided it> WHAT STAYS OUT: <tool or approach we won't use, and why> QUESTION TO CONFIRM: <one question for the client>
You are an AI solutions architect. Classify each case below as GENERATIVE, CLASSIC, HYBRID or NO AI, using these criteria and nothing else: - generative: creates content, converses, unstructured data, creative output, tolerates an error of a few percent; - classic: predicts something, demands high precision, numeric and highly structured data, money at scale involved; - hybrid: multi-step flow, where the classic calculates and the generative interprets and communicates; - no AI: a fixed step solves it, or the problem is a messy process/data. Client cases: """<paste the list of requests here, one per line>""" Estimated volume per case: <operations per day, if known> Error tolerance: <what happens when the system gets it wrong> For each case return: classification · the criterion that decided it · what breaks if I pick another family · one question to confirm with the client. When volume or error tolerance is missing, write "undefined" instead of guessing.
You are a solutions architect. I want an audit with an opinion, not a neutral summary. Tools in use (name, what it's for, monthly cost, number of users): """<paste the list here>""" Context: <industry, size, what the team does day to day> Constraints: <data that can't leave, legacy systems, who approves> Return three blocks: 1. KEEP, and the exact function that justifies the subscription. 2. CUT, what another tool on the list already covers, with the estimated monthly savings and the risk of cutting. 3. REPLACE, with which alternative, the effort and what is lost. Always prefer the solution with FEWER tools. Mark every cost estimate as an assumption when I haven't given the number.
Think of real requests from your niche.
Pick 3 common requests in your niche and fill in the decision board for each one. At least one must end in "no AI" or in hybrid.
Done when: the 3 boards have the decision, the reason and what is left out filled in, with no "it depends".
List a client's paid tools (or the ones in your own routine) and run the stack audit.
Done when: the list has at least one item under CUT with the savings and the risk stated.
4 questions. The explanation appears after your attempt.
Question 1 of 4
An insurance company wants to automatically decide thousands of claims a day, with high accuracy and an audit trail. What do you recommend?
Question 2 of 4
The client wants to set the prices of the store's products automatically. Which family?
Question 3 of 4
An overwhelmed client asks you to recommend tools. What do you deliver?
Question 4 of 4
What is part of the solutions architect's deliverable?
Pick an answer. The explanation shows up after you try.