Understand AI and choose an approach

Outcome and preparation

Day 1 · 09:30–10:30 · 60 minutes. Create a use-case decision card. Bring Module 1's task card. Use 10 minutes for concepts, 40 for comparing approaches and 10 for review.

An LLM predicts useful language patterns; it is not a database of guaranteed facts. Transformers are a model architecture, generative AI creates outputs, and an agent uses tools and feedback to pursue a task. None removes the need for source evidence.

Retrieval brings relevant current/domain evidence into context. Fine-tuning changes model behaviour through training; it is not a substitute for fresh weekly source data. More model capacity or reasoning can improve some tasks but costs time and money and must be tested on your workload.

Step by step

1. Step 1

   Read your task card aloud. List which parts are fixed arithmetic, which require interpretation and which require a person's decision. Mark these calculate, assist, or approve.

2. Step 2

   Give Claude the same small source twice: first with a vague question, then with an evidence requirement. Compare what it claims.

   Prompt / exampleCopy

   Northstar revenue was USD 100000 in the week starting 2026-09-28 and USD 110000
   in the week starting 2026-10-05. No customer interview or marketing data is supplied.
   State the observed change, show the formula, and list what you cannot conclude.
   Cite this supplied two-row example. Do not explain the cause unless evidence supports it.
   

3. Step 3

   Ask for a decision matrix. Save the answer as handover/approach-card.md, then replace unsupported assertions with “to verify.”

   Prompt / exampleCopy

   Compare these approaches for my task card: manual spreadsheet, Claude Project with
   uploaded sources, fixed reporting workflow, and agent with read-only tools.
   Columns: benefit, setup effort, recurring cost inputs, evidence quality, failure modes,
   human review needed, and simplest useful pilot.
   Do not invent current model prices or claim that an agent is always better.
   [Paste task card]
   

4. Step 4

   Compare hosted frontier and open-weight deployment choices from a business perspective: licence, data access, hosting/support effort, quality on this task and total cost. Check current vendor plan pages before writing prices, and record currency, date and whether the number is observed or assumed. A subscription and API usage are different cost categories.

5. Step 5

   Select the workshop default: Project context plus a fixed calculator/report workflow, with a narrow evidence-extraction agent. Explain why custom training and a team of agents are unnecessary for the first MVP. If you propose extra agents, state the testable benefit and added cost.

6. Step 6

   Have a partner challenge the choice as a CFO or CTO. Revise the card with the evidence that would trigger a different model, tool or architecture.

Expected output and verification

A one-page approach card names a chosen method, alternatives rejected, cost assumptions and escalation criteria. A correct answer to the two-row example is 10% revenue growth; it cannot establish why revenue grew or whether the company is profitable.

Check whether your matrix compares equivalent task quality. A cheaper output that requires more review may not be cheaper overall. Do not use popularity rankings as proof of suitability.

Troubleshooting

If technical vocabulary dominates, explain each term using the reporting example. If the assistant recommends fine-tuning automatically, ask what specific failure in the current method training would solve. If pricing cannot be verified, keep the amount unknown and list how to obtain it.

Completion checklist

• Approach card saved with a clear default and alternatives.
• Cost and availability claims have dates or uncertainty labels.
• You can explain workflow versus agent in plain English.