How AI Works in Business

Step 1 — Define your objective

Before writing a single prompt, answer four questions:

Step 2 — Build the prompt

A prompt is assembled from building blocks. Not every use case needs all of them — the specificity of your desired output determines which to include.

Building blocks

Prompt types

Three levels of prompt structure. Each adds building blocks as the task demands more precision.

Demo A — Basic prompt

Prompt type: Basic (task description + context)

Using the attached DataCo supply chain dataset,
what are the top late delivery products?

What happens. The AI has the data, so it pulls real product names — but the output is an unstructured paragraph with no ranking, no percentages, and no clear takeaway. It mentions a few categories without explaining why they matter.

Lesson. The AI has the data but no direction. A minimal prompt produces a minimal answer — real numbers buried in a shapeless paragraph that no executive can act on.

Demo B — Focused prompt

Prompt type: Focused (task description + role + context + specific requirements)

You are a supply chain analyst.

Using the attached DataCo supply chain dataset, identify the top 10
product categories with the highest late delivery rate.

Output a Markdown table with exactly three columns: Category Name,
Total Order Lines, and Late Delivery Rate (%).
Sort by late delivery rate descending.

What happens. The AI returns a ranked table with real numbers from the dataset. Lacrosse leads at 60.1%, followed by Pet Supplies, Cameras, Strength Training and Music; the tenth entry sits near 56%. It looks precise and actionable.

Lesson. A specific prompt produces specific numbers — but numbers are not insight. What the room does not yet see: every one of those ten categories sits within four points of the company average. The ranking is noise.

Demo C — Comprehensive prompt

Prompt type: Comprehensive (all six blocks: task description + role + boundaries + context + specific requirements + reasoning)

You are a supply chain analyst presenting to the CEO.

Using the DataCo dataset, analyse late delivery patterns and produce:

1. Top 10 product categories by late delivery rate (% of order lines late)
2. For each: the total profit associated with late order lines
3. Compare late-delivery rates across shipping modes
4. One actionable recommendation per category to reduce delays

Boundaries: base every finding on the dataset provided. Do not
speculate beyond what the data supports. If a pattern is ambiguous,
say so.

Format as an executive summary with a table. Flag any category where
late delivery rate exceeds 50% as CRITICAL. Keep it under 500 words.

Before generating the table, think step-by-step in a brief paragraph
about how the 50% threshold interacts with the baseline company average.
Then apply the threshold exactly as given.

For each recommendation, explain the reasoning: what evidence supports
it, what trade-off is involved, and what would need to be true for it
to work.

What happens. The AI returns a structured executive brief — prioritised, quantified, with per-category recommendations. Two things surface that no one expected.

• 49 of 50 categories are flagged CRITICAL — because the company-wide late rate is 54.8% and the threshold sits below it. The AI followed the instruction precisely. The result is useless precisely because it did.
• Item 3 uncovers the real story. First Class 95% late, Second Class 77%, Standard Class 38%. The premium shipping modes are the late ones — and this only appeared because the prompt happened to include a shipping-mode comparison.

Lesson. Prompt structure matters — role, constraints, format, and reasoning direction produce a markedly better output. But this prompt applied a bad threshold flawlessly and found the one useful pattern by accident. Better instructions improve direction. They do not replace missing context or missing evidence.

Step 3 — Polish and iterate

No first prompt is final. After reviewing the output, refine with follow-up prompts: ask the AI to add what is missing, go deeper on a specific section, or change the format.

Then apply human expertise — check for accuracy, question the assumptions, verify the ethics, and make final edits before the output leaves your hands. AI drafts; you decide.

Step 2.1 — Add business context

COMPANY CONTEXT:
- We are a B2B distributor serving corporate and home office segments
- Our SLA commitment: delivery within scheduled shipping days
- Board KPI: reduce late delivery rate from 54.8% to below 30% in 6 months
- Current allocation by order lines: Standard Class 59.7%, Second Class
  19.5%, First Class 15.4%, Same Day 5.4%
- Budget constraint: logistics cost cannot increase more than 12%

DATASET: [DataCo supply chain data attached]

Given this context, analyse our late delivery problem and recommend a
phased action plan that stays within budget. Prioritise by profit impact.

Format the plan into three phases: Immediate (0-30 days), Short-Term
(30-90 days), and Medium-Term. For each phase, explicitly state the
expected budget impact against the 12% ceiling.

The 54.8% starting rate and the four allocation percentages come from the data Derived Evidence. The SLA, the board target and the budget ceiling are invented for the exercise Workshop Scenario.

What happens. The recommendations change completely. The AI will not propose shifting volume to premium shipping — it now knows both the budget ceiling and the fact that premium modes are the late ones. Expect it to move volume toward Standard Class, the cheapest and most punctual mode in this data, and to prioritise the segments named in the context.

Lesson. Context engineering is the practice of giving AI your company’s rules, KPIs, constraints, and priorities. This is where enterprise AI separates from consumer AI.

Step 2.2 — Add historical context, or memory

PREVIOUS ANALYSIS (Q1):
- We shifted 15% of Second Class volume to Standard Class in West of USA
- Result: late deliveries dropped 8% in West of USA but increased 3% in
  US Center
- Root cause hypothesis: the Standard Class carrier has limited US Center
  capacity

Given this history, update your recommendations. Do not repeat strategies
that failed. Account for regional carrier capacity differences.

Provide a "Revised Action" table contrasting the Q1 approach with the new
approach for US Center.

What happens. The AI adjusts course. It keeps the direction the data supports, avoids the approach that failed in US Center, and proposes different tactics for different regions.

Lesson. Enterprise AI without memory will repeat the mistakes the organisation already paid to learn from.

Step 3.1 — Multi-step workflow, live demo

WORKFLOW:
1. Query the DataCo dataset: pull all order lines flagged as late delivery.
   Drop any rows with null values in the target columns.
2. Segment by: region, shipping mode, customer segment, product category
3. Generate a visual dashboard with 4 charts. Use a consistent corporate
   colour palette (navy and gold) and label every chart axis clearly:
   - Late delivery rate by region (bar chart)
   - Shipping mode vs delay days (box plot)
   - Profit associated with late deliveries by category (treemap)
   - Trend: late delivery rate over time (line chart)
4. Identify the three strongest combinations associated with late
   delivery (e.g. Standard Class + West of USA + a specific category)
   and state what additional evidence would be required to establish
   causation
5. Draft an email to the VP of Logistics summarising findings with
   the dashboard attached

What happens. The AI produces analysis, four visualisations, a pattern assessment, and a draft email to the VP of Logistics — in one uninterrupted flow.

Lesson. AI with tools becomes a work multiplier. Without them, it can only talk about the work. With them, it does the work.

Step 3.2 — Tool awareness

Point out what the AI just used:

• Code execution — ran Python or SQL over the dataset
• Visualisation — generated the charts
• Document generation — drafted the email
• Data retrieval — queried specific subsets

Key takeaway

An AI model without tools is a consultant with no laptop. It can advise, but it cannot deliver.

Step 4.1 — The autonomous agent scenario

Name: Supply Chain Monitor
Trigger: Runs daily at 6:00 AM

INSTRUCTIONS:
1. Pull yesterday's delivery data
2. Calculate late delivery rate by region and shipping mode
3. Compare to the 30% target KPI
4. IF any region exceeds 50% late delivery rate:
   - Analyse contributing factors (carrier, product mix, volume spike)
   - Generate alert with supporting evidence
   - Draft recommended action
   - Send summary to VP Logistics
5. IF all regions are below 30%:
   - Log "All clear" and update weekly trend tracker
6. Every Friday: generate weekly performance summary with trend analysis

Show the output. Walk through a bad day and a good day. In the historical data, most regions sit between 53% and 58% — so the alert is today’s reality and the all-clear is the target state.

Lesson. Agentic AI operates across time without being prompted. The executive question is not how many agents you deploy, but what each one may access, recommend, and execute.

Step 4.2 — Beyond a single agent

Organisations rarely run one agent. What they run is a small set of specialised agents in a hub-and-spoke arrangement, not a peer-to-peer network negotiating with itself.

The structure is what keeps this manageable. Several agents contribute to one unified picture for leadership, without unchecked interaction between the agents themselves.