What Changes When AI Reads the Model
"The formula is the same calculation. The architecture is the difference between an agent that describes your spreadsheet and an agent that analyses your business."
In Lesson 1, you saw Formula Rot from the inside: coordinate references that break silently, logic that diffuses across dozens of cells, and a reverse-engineering burden that grows every time the model changes hands. You understand the problem. Now you will see what changes when the problem is solved.
This lesson puts two analysts in the same room with the same model and the same question. One analyst's model uses coordinate references. The other analyst's model has been converted to Named Ranges. Both ask Claude the same question: "Explain the Gross Profit formula." The answers they receive are not incrementally different. They are categorically different — in confidence, in business context, and in actionability. That gap is the reason IDFA exists, and it is the gap that every remaining lesson in this chapter works to close.
Coordinate-First modelling defines business logic by location. The formula =B14-(C14*$F$8) means "the value in B14 minus the product of C14 and F8." The meaning of that calculation is not in the formula. It is in the column headers, in the analyst's memory, in the documentation that may or may not exist.
Logic-First modelling defines business logic by name. The formula =Revenue_Y2 - (Revenue_Y2 * COGS_Pct_Y2) means "Year 2 Revenue minus the product of Year 2 Revenue and Year 2 COGS Percentage." The meaning is in the formula. The formula reads as the business rule it encodes.
The shorthand: Coordinate-First asks "where is this number in the grid?" Logic-First asks "what does this number mean?"
The Same Formula, Two Architectures
Here is a Gross Profit formula for Year 1. It calculates the same number in both versions. The math is identical. The architecture is not.
Coordinate version:
=B14-(B14*$C$3)
Named Range version:
=Revenue_Y1 - (Revenue_Y1 * COGS_Pct_Y1)
Read both formulas aloud. The coordinate version sounds like a grid reference: "B14 minus B14 times C3." The Named Range version sounds like a business rule: "Revenue Year 1 minus Revenue Year 1 times COGS Percentage Year 1." The same calculation. The same answer. A fundamentally different level of information available to anyone — or anything — reading it.
What the Agent Sees
When you ask Claude to explain the coordinate formula, here is the kind of response you receive:
Prompt: "Explain the formula in cell D14."
Agent response (coordinate model):
"Cell D14 calculates as B14 minus the product of B14 and the value in C3. B14 appears to contain a revenue figure (11,000,000). C3 appears to contain a percentage (0.59). The result is approximately 4,510,000. This may represent a gross profit calculation, though I would need to verify the column headers to confirm."
Notice the language. "Appears to contain." "May represent." "I would need to verify." The agent is hedging because it is inferring intent from position. It can see the numbers. It cannot see the business rule.
Now the same question on the Named Range model:
Prompt: "Explain the Gross Profit formula for Year 1."
Agent response (Named Range model):
"Year 1 Gross Profit is calculated as Revenue minus the product of Revenue and COGS Percentage. The business rule: Gross Profit equals Revenue after subtracting Cost of Goods Sold, where COGS is expressed as a percentage of Revenue. At current assumptions (Revenue of $11,000,000 and COGS Percentage of 59%), Gross Profit is $4,510,000. A 1-percentage-point reduction in COGS Percentage would increase Gross Profit by $110,000."
No hedging. No "appears to." The agent names the business rule, states the assumptions, calculates a sensitivity, and offers a what-if scenario — because the formula told it everything it needed to know.
Three Dimensions of the Quality Gap
The difference between these two responses is not random. It follows a pattern with three measurable dimensions.
| Dimension | Coordinate Model | Named Range Model |
|---|---|---|
| Confidence | Hedged: "appears to," "may represent," "would need to verify" | Definitive: "is calculated as," "the business rule is" |
| Business context | Mechanical: describes cell references and values | Analytical: names the business rule, identifies assumptions, explains the logic |
| Actionability | Descriptive: tells you what the formula does | Prescriptive: suggests what-if scenarios and quantifies impact of changes |
This is not a matter of prompt engineering. You cannot prompt your way out of a coordinate model. The information the agent needs — the business intent behind the formula — is either in the formula or it is not. No prompt can create intent that was never encoded.
Why This Is a Business Problem, Not a Technical One
Every token an agent spends inferring what a formula means is a token it does not spend analysing what the formula tells you about your business. Coordinate models force agents into inference mode: tracing cells, guessing at column headers, hedging conclusions. IDFA models put agents into analysis mode: naming business rules, identifying assumptions, running sensitivities, recommending actions.
The equation is direct:
Agent capability gap = Business capability gap
A finance team using coordinate models gets descriptions of their spreadsheets. A finance team using IDFA-compliant models gets analysis of their business. The agent is the same. The model architecture is the multiplier.
What Becomes Possible
When every formula reads as a business rule, capabilities emerge that coordinate models cannot support. The full IDFA methodology — which you will build across the remaining lessons in this chapter — unlocks five Finance Domain Agent capabilities. This lesson previews the first three:
Intent Synthesis. Because Named Range formulas encode business rules, the agent can synthesise the intent of an entire model section. Instead of reporting "cells D14 through D18 contain formulas referencing column B and C," the agent reports: "The Gross Profit section calculates margin after COGS for each projection year, with COGS expressed as a percentage of Revenue that decreases from 59% to 54% over the five-year forecast, reflecting an assumption of improving cost efficiency at scale."
Deterministic What-If. Because assumptions are isolated in Named Ranges, the agent can identify every assumption driving a result and propose what-if scenarios. "What if Revenue grows 15% instead of 10%?" is a question the agent can answer precisely — because it knows which Named Range controls revenue growth and which formulas depend on it.
Strategic Goal-Seeking. Because the relationship between assumptions and outputs is explicit, the agent can work backwards: "What COGS Percentage would you need to achieve a 45% Gross Margin?" The agent identifies the assumption to change, the output to target, and the value that achieves the goal — all from the Named Range structure.
These capabilities are not theoretical. You will build and test each one in later lessons. For now, the point is architectural: they require Logic-First design. A coordinate model cannot support Intent Synthesis because the intent was never encoded. A coordinate model cannot support precise What-If because the assumptions are not isolated. Architecture enables capability.
Exercise: Write It Both Ways
Take the Gross Profit formula and write it in both forms yourself.
Step 1. Write the coordinate version:
=B14-(B14*$C$3)
Assume B14 contains Revenue for Year 1 ($11,000,000) and C3 contains the COGS Percentage (59%).
Step 2. Write the Named Range version:
=Revenue_Y1 - (Revenue_Y1 * COGS_Pct_Y1)
Step 3. Give both formulas to Claude (in Claude in Excel, in Cowork, or in claude.ai) with this prompt:
Here is a formula from a financial model:
[paste one version]
Explain what this formula calculates, what business rule it encodes,
and what would happen if the inputs changed by 10%.
Run the prompt twice — once with the coordinate version, once with the Named Range version.
Step 4. Document the difference. Use the three dimensions from this lesson:
- Confidence: Did the agent hedge or state definitively?
- Business context: Did the agent describe cells or name business rules?
- Actionability: Did the agent offer what-if analysis or stop at description?
The pattern you observe is the pattern this entire chapter addresses.
Try With AI
Use these prompts in Claude in Excel or your preferred AI assistant to explore the Coordinate-First versus Logic-First distinction hands-on.
Prompt 1: Formula Readability Comparison
I have two versions of the same Gross Profit formula.
Version A (coordinate): =B14-(B14*$C$3)
Version B (Named Range): =Revenue_Y1 - (Revenue_Y1 * COGS_Pct_Y1)
Both calculate the same number. Rate each version on a scale of 1-10
for: (a) how easy it is to explain to a non-technical executive,
(b) how confidently you can describe the business rule it encodes,
and (c) how useful it is for suggesting what-if scenarios.
Explain your ratings.
What you are learning: This prompt forces the AI to articulate exactly where the information gap lies. The ratings make the quality difference quantitative rather than impressionistic — and the explanations reveal that the gap is architectural (information encoded in the formula), not about AI capability.
Prompt 2: Business Analysis from Named Ranges
Here is a simplified income statement model using Named Ranges:
Revenue_Y1 = 11,000,000
Revenue_Growth_Rate = 0.10
Revenue_Y2 = Revenue_Y1 * (1 + Revenue_Growth_Rate)
COGS_Pct_Y1 = 0.59
COGS_Pct_Y2 = 0.57
Gross_Profit_Y1 = Revenue_Y1 - (Revenue_Y1 * COGS_Pct_Y1)
Gross_Profit_Y2 = Revenue_Y2 - (Revenue_Y2 * COGS_Pct_Y2)
Provide a business analysis of this model. Include:
1. What business assumptions are encoded
2. What trends the assumptions imply
3. Two risks that a CFO should consider
4. One question a board member might ask about these projections
What you are learning: With Named Range formulas, the agent moves beyond mechanical description into genuine business analysis. It can identify assumptions, infer trends, flag risks, and anticipate stakeholder questions — because the business logic is readable, not hidden in cell coordinates. This is what "analysis mode" looks like.
Prompt 3: What-If on Both Formats
I have two versions of a projection model. Both contain the same data.
VERSION A (coordinate references):
Cell B14 = 11000000
Cell B15 = B14 * 1.10
Cell C3 = 0.59
Cell D14 = B14-(B14*C3)
VERSION B (Named Ranges):
Revenue_Y1 = 11,000,000
Revenue_Y2 = Revenue_Y1 * (1 + Revenue_Growth_Rate)
Revenue_Growth_Rate = 0.10
COGS_Pct_Y1 = 0.59
Gross_Profit_Y1 = Revenue_Y1 - (Revenue_Y1 * COGS_Pct_Y1)
For each version, answer: "What happens to Gross Profit if revenue
grows 15% instead of 10%?"
After answering both, explain why one version let you give a more
complete and confident answer than the other.
What you are learning: What-If analysis is a core Finance Domain Agent capability — and it depends entirely on model architecture. This prompt demonstrates that the agent can answer the same question on both models, but the Named Range version produces a precise, assumption-aware answer while the coordinate version forces the agent to guess which cells control growth. Architecture determines capability.
Flashcards Study Aid
Continue to Lesson 3: The Three Layers -->