Delegated Calculation — Guardrail 4
In Lesson 6, you learned to attach Intent Notes that document the business purpose of every AI-generated formula. Now you reach the guardrail that governs how the agent interacts with the model itself. The CFO walks into your office: "What if we lose the Johnson account — that is $2M off Year 1 Revenue. What happens to Year 3 margins?" In the old world, you open Excel, change a cell, trace the impacts manually, re-check the formulas, and report back an hour later. With IDFA and the IDFA plugin, you give the agent the question. The agent writes the new assumption. The spreadsheet engine calculates. The agent reads back the answer. Every number comes from the model. Nothing is estimated.
This distinction — between an agent that reports what the model calculated and an agent that reports what it calculated internally — is the most consequential guardrail in IDFA. In finance, these are categorically different. One is audit-valid. The other is an opinion.
The Rule
Guardrail 4 — Delegated Calculation: An AI agent operating on an IDFA-compliant model is prohibited from performing calculations internally. It must delegate all arithmetic to the spreadsheet engine — writing assumptions, letting Excel recalculate, and reading back the results.
The agent provides the reasoning. The spreadsheet engine provides the mathematics. The result is not the agent's estimate — it is the model's deterministic output.
Why This Guardrail Exists
Consider what happens without it. You ask the agent: "What is Year 2 Gross Profit if COGS improves to 58%?" The agent knows the formula. It knows Revenue_Y2 is $11M. It calculates internally: $11,000,000 minus ($11,000,000 times 0.58) equals $4,620,000. It reports this number confidently.
The problem: the model might have a different Revenue_Y2. Perhaps another assumption changed since the agent last read the model. Perhaps the revenue growth formula includes a rounding convention. Perhaps there is a conditional override the agent does not know about. The agent's internal arithmetic is based on its understanding of the model — which may be incomplete, outdated, or wrong in ways neither the agent nor the user can detect.
When the agent delegates calculation to the spreadsheet engine, none of this matters. It writes the COGS assumption. Excel recalculates every dependent cell using the actual formulas in the actual model. The agent reads back the result. The number is mathematically certain because it came from the deterministic calculation engine, not from the agent's reasoning.
When the IDFA plugin is active, Claude has two skills working together:
| What Claude Does | How It Works |
|---|---|
| Write an assumption | Changes a Named Range value in the model |
| Recalculate | Triggers the spreadsheet engine to recalculate all formulas |
| Read a result | Reads the current value of any Named Range |
| Inspect the model | Lists all Named Ranges, their values, and dependencies |
| Read a formula | Returns the formula assigned to a Named Range |
The agent writes to Named Ranges, never to cell coordinates. The spreadsheet engine recalculates. The agent reads back Named Range results. This is the complete interaction pattern.
The Write-Calculate-Read Workflow
Every interaction follows the same three-step pattern:
Step 1 — The agent reasons about the correct input:
The agent determines what assumption to change based on the user's question. This is where the agent's intelligence matters — understanding the business context, selecting the right Named Range, choosing the right value.
Step 2 — The agent writes the input and the spreadsheet engine recalculates:
You ask Claude to change an assumption. Claude writes the new value to the Named Range in the model and triggers a recalculation. The agent does not calculate anything — it writes a value to a Named Range and lets the spreadsheet engine do the rest.
Step 3 — The agent reads back the result:
Claude reads the affected outputs from the model and reports them to you. The numbers it reports are the model's deterministic outputs — not estimates, not approximations, not internal arithmetic.
This separation is absolute. The agent never performs arithmetic on financial data. It reasons about what inputs to change and what outputs to read. The spreadsheet engine handles every calculation between those two points.
Deterministic What-If Analysis
The what-if workflow demonstrates Guardrail 4 at its most powerful.
The question: "What if Year 1 Revenue is $12M instead of $10M? What happens to Gross Profit across all three years?"
Here is what this looks like in Cowork:
You: What if Year 1 Revenue is $12M instead of $10M? Show me all three years of Gross Profit.
Claude: I'll update the assumption and recalculate the model.
Changing Inp_Rev_Y1 from $10,000,000 to $12,000,000... Model recalculated. Results from the model:
Output Value Gross_Profit_Y1 $4,800,000 Gross_Profit_Y2 $5,412,000 Gross_Profit_Y3 $6,098,400 With Year 1 Revenue at $12M, the model produces Gross Profit of $4.8M in Year 1, $5.41M in Year 2, and $6.10M in Year 3.
Every number came from the spreadsheet engine. The agent added zero arithmetic. If a CFO asks "where did these numbers come from?", the answer is: the model calculated them. Not the agent. Not an estimate. The model.
The most dangerous violation of Guardrail 4 is calculating a result internally and then writing it to the model as if Excel produced it.
Wrong: Agent computes Gross_Profit_Y2 = $11M - ($11M * 0.59) = $4,510,000, then reports it.
Right: Agent writes Inp_COGS_Pct_Y2 = 0.59 to the model, the spreadsheet engine recalculates, and the agent reads back Gross_Profit_Y2 from the model.
Internal calculation and deterministic model calculation can produce different results. Only the model result is audit-valid. If the agent ever reports a number it computed rather than read, the entire audit trail breaks.
Strategic Goal-Seeking
What-if answers the question "what happens if I change this input?" Goal-seeking answers the reverse: "what input do I need to achieve this output?"
The question: "What Year 1 Revenue do I need for Year 3 Gross Profit to equal exactly $5,500,000?"
The agent cannot solve this algebraically and report the answer — that would be internal calculation. Instead, it iterates through the model:
You: What Year 1 Revenue do I need for Year 3 Gross Profit to equal exactly $5,500,000? Show me each iteration.
Claude: I'll iterate through the model to find the answer. Let me start with $11M and adjust from there.
Iteration 1: Setting Inp_Rev_Y1 to $11,000,000... → Reading Gross_Profit_Y3 from the model: $5,589,100 — too high.
Iteration 2: Setting Inp_Rev_Y1 to $10,800,000... → Reading Gross_Profit_Y3 from the model: $5,473,728 — too low.
Iteration 3: Setting Inp_Rev_Y1 to $10,850,000... → Reading Gross_Profit_Y3 from the model: $5,502,610 — close.
Iteration 4: Setting Inp_Rev_Y1 to $10,845,000... → Reading Gross_Profit_Y3 from the model: $5,499,726 — closer.
Iteration 5: Setting Inp_Rev_Y1 to $10,846,000... → Reading Gross_Profit_Y3 from the model: $5,500,303 — within tolerance.
Year 1 Revenue of approximately $10,846,000 produces Year 3 Gross Profit of $5,500,303 — within $303 of the $5,500,000 target.
Every intermediate result came from the model. The agent used binary search logic to converge on the answer, but every data point was a deterministic model output. The final answer is not the agent's estimate — it is the input that the model confirms produces the target output.
The Business Bottom Line
The CFO asks: "What if COGS improves 2% faster than planned?" Without IDFA and the IDFA plugin, an analyst opens the model, changes a cell, traces the downstream impacts, double-checks the formulas, and reports back. That takes time, and the analyst's interpretation of the model may introduce errors.
With IDFA and the IDFA plugin: the agent writes the new COGS efficiency assumption to the model. The spreadsheet engine recalculates every dependent cell. The agent reads back Gross Profit for all three years. The CFO has a verified answer in seconds — and every number traces back to the model's deterministic calculation engine.
Agent reasoning plus spreadsheet mathematics equals mathematical certainty. That is Guardrail 4.
Capability Preview
The workflows you practiced in this lesson — Deterministic What-If and Strategic Goal-Seeking — are Capabilities 2 and 4 of the five Finance Domain Agent capabilities you will validate in the capstone (Lesson 11). In the capstone, you will run formal capability tests that verify the agent delegates every calculation to the model, never reporting a number before reading it from the model.
Exercise: Full What-If Workflow
Using the Gross Profit Waterfall model built in earlier lessons (or the IDFA SKILL.md reference values), complete the following in Cowork:
Part 1 — What-If:
- Ask Claude: "Change Year 1 Revenue to $12M and show me all three years of Gross Profit."
- Confirm the results: GP Y1 = $4,800,000, GP Y2 = $5,412,000, GP Y3 = $6,098,400
- Verify: every number Claude reported came from the model — none from its own arithmetic
Part 2 — Goal-Seeking:
- Ask Claude: "Find the Year 1 Revenue needed for Year 3 Gross Profit to equal $5,500,000. Show me each iteration."
- Watch Claude iterate — writing assumptions, reading back results, adjusting, and repeating
- Verify: the final answer was found through iteration, not algebraic solving
- Document how many iterations it took to converge
Success criteria: Every number Claude reported came from the model. Claude performed zero arithmetic. The goal-seeking answer was found through iteration, not calculation.
Try With AI
Setup: Open Cowork with the IDFA plugin installed and your Gross Profit Waterfall model connected.
Prompt 1 — What-If Scenario:
What if Year 1 Revenue is $12M instead of $10M? Show me the impact
on all three years of Gross Profit. Do not calculate any number
internally — every result must come from the model.
What you are learning: This prompt tests whether the agent follows Guardrail 4. Watch for the write-calculate-read sequence. Every Gross Profit figure should come from reading the model, not from the agent's arithmetic. If the agent reports a number without reading it from the model first, it violated the guardrail.
Prompt 2 — Goal-Seeking:
Find the Year 1 Revenue needed for Year 3 Gross Profit to equal
exactly $5,500,000. Iterate through the model — write the assumption,
read back the result, adjust, repeat — until you converge on the
answer. Show me each iteration.
What you are learning: Goal-seeking forces the agent to use the model as a calculator rather than solving the problem analytically. The agent must iterate — write an input, read the output, adjust, repeat. Each iteration produces a deterministic model result. The final answer is the input that the model confirms produces the target.
Prompt 3 — Multi-Variable Sensitivity:
What happens to the three-year Gross Profit total if Revenue grows
at 15% instead of 10% AND COGS efficiency improves to 2% per year
instead of 1%? Change both assumptions and read back all results
from the model. Compare to the base case.
What you are learning: Multi-variable scenarios are where internal calculation becomes most dangerous — the compounding interactions between variables make mental arithmetic unreliable. By delegating everything to the spreadsheet engine, the agent handles complexity without introducing estimation error. The model calculates the compound effects; the agent reports them.
Flashcards Study Aid
Next: Lesson 8: Retrofitting Existing Models — where you learn to convert legacy coordinate-based models to IDFA compliance using the five-phase retrofitting process.