Overview

• Purpose: Understand why “just ask AI about our data” doesn’t work, and what foundation you need first.
• You will implement: A simple data dictionary for your key sources and a sample pack for testing AI prompts.
• Prerequisites: Access to your main data sources (ERP exports, spreadsheets, BI views).

The Story: “Can’t You Just Ask AI?”

The moment I knew we had a problem came during a budget review.

Our CFO asked a reasonable question: “Why is Q3 software spend $70K over budget?” She’d heard about AI tools and suggested, “Can’t you just upload the data and ask ChatGPT to explain it?”

I thought I could. I exported our vendor spend data, uploaded it to ChatGPT with Code Interpreter, and asked: “Explain why software spend is over budget.”

ChatGPT responded:

“Looking at your data, I see a ‘Category’ column with values including ‘SW’, ‘Software’, ‘Tech-SW’, and ‘SOFT’. These appear to be software-related. The ‘Amount’ column shows negative and positive values. Could you clarify: 1. Are all these categories supposed to be the same? 2. Are negative amounts credits or expenses? 3. What time period does this cover? 4. What was the budget amount?”

I couldn’t answer most of these questions confidently. The data came from three different systems, each with its own category names. Nobody had documented which categories mapped to “software.” The negative amounts were credits, but that wasn’t noted anywhere. And the budget was in a separate spreadsheet with a different category structure.

The CFO’s “simple” question required three hours of reconciliation work before AI could help at all.

The lesson: AI can’t analyze data you don’t understand yourself. If you can’t quickly explain what each column means, where the data comes from, and how things should add up, AI will just be confused in a more sophisticated way.

“SW,” “Software,” “Tech-SW,” and “SOFT” are all the same thing… right?

Why AI Fails on Messy Data

AI tools are pattern-matching engines. They’re good at: - Finding trends in clean, labeled data - Explaining what’s in front of them - Drafting narratives about numbers

They’re bad at: - Knowing what your data should look like - Understanding your company’s definitions - Reconciling inconsistent sources - Filling in context that isn’t there

When you upload messy data, AI will confidently analyze it–and produce confidently wrong answers.

🏆 Minimum Viable AI Win (10 Minutes)

Try this: Pick one data source you use regularly (a key report, an ERP export, a spreadsheet). Write down:

1. Where it comes from (system, report name)
2. What each column means (in plain English)
3. How often it updates
4. What the numbers should add up to (a control total)

That’s a data dictionary for one source. Took 10 minutes. Now when you share that data with AI, you can paste these definitions too, and you’ll get much better answers.

The Minimum Viable Data Foundation

You don’t need a data warehouse project. You need:

Field: Category
Source: ERP GL Export
Values: SW, HW, PS, MKT, FAC, OTH
Mappings:
- SW = Software (includes SaaS subscriptions)
- HW = Hardware (computers, equipment)
- PS = Professional Services (consulting, audit)
- MKT = Marketing (advertising, events)
- FAC = Facilities (rent, utilities)
- OTH = Other
Note: "Tech-SW" and "SOFT" in legacy data should map to SW
Owner: Controller

sample-pack/
├── gl_sample.csv        # 20 rows of anonymized GL data
├── budget_sample.csv    # 10 rows of budget data
├── mapping.csv          # Category mappings
└── README.md            # What each file is and how to use it

A Real Data Dictionary Example

Here’s what a useful data dictionary looks like for a spend analysis:

The Quality Checklist

Before trusting any data for analysis, check:

When Data Issues Block AI

If you ask AI a question and get a confusing answer, the problem is often data, not AI:

Tools That Can Help

Start Small: The 5-Field Exercise

Don’t try to document everything at once. Start with one source:

Pick your most-used data source. The one you touch every week.

Document 5 key fields:

## Monthly Revenue Report

### Source
- System: CRM export (Revenue Ops)
- Refresh: Daily
- Granularity: Line item by customer by month

### Key Fields

1. **customer_id**: Unique customer identifier
   - Format: CUST-12345
   - Source: CRM master

2. **revenue**: Monthly recognized revenue
   - Currency: USD
   - Negative values: Credits/refunds
   
3. **segment**: Customer segment
   - Values: Enterprise, Mid-Market, SMB
   - Based on: ARR at time of booking
   
4. **product_line**: Product category
   - Values: Platform, Add-ons, Services
   - Note: "Legacy" products map to "Platform"
   
5. **period**: Fiscal month
   - Format: YYYY-MM
   - Timing: By service delivery month

### Control Totals
- Monthly total should match Flash report line 1
- Customer count should match ARR report

That’s it. Now you have something useful you can share with AI when you need help with this data.

Common Mistakes

Key Takeaways

1. AI can’t fix data you don’t understand. Document before automating.

2. Start with one source. Document 5 key fields. Build from there.

3. Definitions matter. “Category = SW” doesn’t help unless you know SW means Software.

4. Control totals are essential. Always have a number to check against.

5. Good enough beats perfect. A one-page Markdown file is infinitely better than nothing.

What You’re Shipping This Week

• ☐ Pick one data source you use weekly
• ☐ Document 5 key fields in a Markdown file
• ☐ Note one control total you can check against
• ☐ Share it with one teammate for feedback

Next Up

Chapter 2 covers safe and responsible AI use–the security, compliance, and data handling practices you need before experimenting with company data.