Session 1 · Overview
How AI Works in Business
From Prompt to Governed Decision
| Duration | 60 minutes |
| Audience | C-level executives |
| Format | Live demonstration on one screen. Participants watch; laptops are not needed. |
| Dataset | DataCo Supply Chain, sanitized extract |
| Output | Each participant places one AI initiative on the AI Readiness Rubric |
Workshop Concept
One business question. Five layers of enterprise AI.
Session 1 uses the DataCo supply chain dataset to follow one late-delivery problem through five rounds. Each round adds a capability that separates enterprise AI from a chatbot: structured instructions, business context, working tools, continuous operation, and governed authority.
OpeningOur delivery data tells us more orders arrive late than on time. The board wants to know where, why, and what to do about it. Over the next hour we will ask that same question five times. Each time, we give the AI one more thing it needs to be useful: clearer instructions, business context, real tools, the ability to run on its own, and rules about what it may and may not do.
The DataCo Dataset
One slide. The key point: this is line-level data — each row is a single order line, not an order summary.
| Field | Example | Notes |
|---|---|---|
Order Id | 76255 | One order may have multiple lines |
order date | 1/15/2017 | |
Shipping Mode | Standard Class | Standard Class · Second Class · First Class · Same Day |
Customer Segment | Consumer / Corporate / Home Office | |
Category Name | Cleats, Men's Footwear, Fishing | DataCo product categories, not aggregated |
Department Name | Fan Shop, Apparel, Golf, Technology | |
Delivery Status | Late delivery / Advance shipping / Shipping on time / Shipping canceled | Historical outcome |
Days for shipment (scheduled) | 4 | Promised shipping window |
Days for shipping (real) | 7 | Actual days taken |
Order Profit Per Order | 25.50 | Behaves as line-level profit in this extract |
Order Region / Order City | West of USA / Los Angeles | |
Late_delivery_risk | 1 | Historical flag. It equals Delivery Status = Late delivery on every row, so it is an outcome, not a prediction |
Important for analysis. Order-level metrics such as unique order count or order-level late rate require deduplication on Order Id. Delivery Status, actual shipping days and Late_delivery_risk are historical outcomes, not predictions.
What the data actually shows
The facilitator should know these numbers before going live. The AI will surface them during the demonstration. Derived Evidence
| Cut | Result |
|---|---|
| Company-wide late rate | 54.8% of 180,519 order lines, and the same at order grain across 65,752 orders |
| By shipping mode | First Class 95.3% late · Second Class 76.6% · Same Day 45.7% · Standard Class 38.1% |
| Share of order lines by mode | Standard 59.7% · Second 19.5% · First 15.4% · Same Day 5.4% |
| By product category | 50 categories, every one between 47.7% and 60.1% |
| By region | 23 regions, every one between 48.8% and 58.0% |
| By customer segment | Consumer, Corporate and Home Office within 0.4 points of each other |
What this means on stage: ranking categories or regions produces lists whose differences are noise — every category is within four points, every region within ten. The only real lever in this data is shipping mode, and the premium services are the late ones. Layers 1 and 2 are designed so the room discovers this live.
Facilitator note. Load the dataset into your AI tool before the session. Paste the prompts below rather than typing them live.
The Five Layers
Each layer adds one capability that separates enterprise AI from a chatbot. They are built in order, because each depends on the one below it.
- Prompt Engineering — asking better questions.
- Context Engineering — giving it your rules, targets and history.
- Workflows and Tools — letting it query data, build charts and draft documents.
- Agentic AI — running on a schedule instead of on request.
- Guardrails and Human Authority — what it may decide, and who signs.