“Why did Site 2’s margin drop 9 points last month?”
That question takes a financial analyst hours in a spreadsheet: tracing relationships across production, cost, and revenue data, checking which costs are fixed versus variable, working out whether the problem is price, volume, or mix. An AI agent with access to your formula structure can do it in seconds, because it follows the declared mathematical relationships your team already built into their KPIs.
That’s AI formula chain analysis. A caveat upfront: it’s only as good as the model it connects to. Stale formulas or broken data connections produce wrong answers. But when the answer is wrong, you can see exactly which formula or data source failed, which is more than any correlation-based approach will give you.
What Changes When AI Can Read Your Formulas
Most AI integrations connect to databases and return numbers. The AI can see that Revenue dropped and Margin fell, but it can’t trace why. Did the cost go up? Did volume decrease? Was it a product mix shift? Without the formulas, it’s left to infer. And when those inferences are wrong, there’s no trail to follow.
There’s a difference between your BI tools showing relationships on a dashboard and exposing them programmatically so an AI can reason with them. Your dashboards drill down, your calculated fields derive metrics, and your analysts navigate those paths daily. That navigation lives in people’s heads. The AI doesn’t get it.
MCP (Model Context Protocol), the vendor-neutral standard now governed by the Linux Foundation, changes this. When an operational platform runs as an MCP server, it exposes not just values but the structure behind them.
A beverage manufacturer’s Capstone model, for instance, contains 114 metrics across 6 disciplines: 37 raw inputs from plant systems and 77 calculations with explicit formulas. The organisational hierarchy spans 121 nodes, from enterprise level down through 3 sites, 4 production areas per site, individual lines, and individual equipment.
Through MCP, the AI gets two things that flat data connectors don’t provide:
The formula graph. Every calculated metric declares its formula. The AI can traverse the entire dependency chain:
| Metric | Formula | Type |
|---|---|---|
| Gross Margin % | Profit ÷ Total Revenue | Financial outcome |
| Profit | Total Revenue − Total Cost | Financial outcome |
| Cost Per Unit | Total Cost ÷ Filler Throughput | Unit economics |
| Total Cost | Fixed Cost + Variable Cost | Cost rollup |
| Variable Cost | Bottle + Packaging + Syrup + Energy + Labour | Cost rollup |
| Energy Cost | Energy Cost/Line/Day × Number of Production Lines | Cost driver |
| Labour Cost | Labour Cost/Shift × Number of Lines × Shifts per Day | Cost driver |
Each formula references other metrics, which reference others in turn. The AI doesn’t stop at “cost went up.” It traces which cost, why, and whether it scales with volume or capacity.
The organisational hierarchy. Metrics roll up from equipment to line to area to site to enterprise. The AI knows that improving Filling Line 02 at Site 2 affects Site 2’s aggregate, which affects the enterprise total. It can compare across sites, trace impacts upward, and quantify them at every level.
The Diagnosis: Site 2’s Margin Collapse
Walk through a real example.
The question: “Why did Site 2’s gross margin drop from 56% to 47% between January and February?”
With formula chain access, the agent traces the declared relationships:
| Metric | Jan 2026 | Feb 2026 | Change |
|---|---|---|---|
| Gross Margin % | 56.1% | 46.9% | −9.2 pts |
| Revenue Per Unit | $0.71 | $0.71 | Flat |
| Cost Per Unit | $0.31 | $0.38 | +22.6% |
| Filler Throughput | 7,885,197 | 4,681,880 | −40.6% |
| Labour Cost | $694,400 | $627,200 | −9.7% |
| Energy Cost | $86,800 | $78,400 | −9.7% |
The agent traces the chain step by step:
| Step | Formula | What the AI Finds |
|---|---|---|
| 1 | Gross Margin % = Profit ÷ Total Revenue | Margin fell 9.2 pts |
| 2 | Profit = Revenue − Total Cost | Revenue Per Unit flat → cost is the driver |
| 3 | Cost Per Unit = Total Cost ÷ Filler Throughput | Throughput dropped 40.6%, but cost didn’t drop proportionally |
| 4 | Labour Cost = Labour/Shift × Lines × Shifts/Day | Formula doesn’t reference throughput → labour is semi-fixed |
| 5 | Energy Cost = Energy/Line/Day × Production Lines | Same pattern → energy is semi-fixed |
Step 4 is where the diagnosis lands. The labour formula doesn’t reference throughput. It references shifts and lines. Labour scales with capacity, not output. Fewer bottles across the same cost base means higher cost per unit.
The root cause isn’t a cost blowout. Volume dropped and fixed costs didn’t follow it down. The correct response is to recover throughput, not cut costs. A correlation-based AI, seeing “costs are up relative to revenue,” would have recommended cost reduction. Exactly backwards.
The AI didn’t infer that labour costs are semi-fixed. It read the formula. And when an analysis like this is wrong, you can trace exactly where it went wrong, because every step references a declared relationship.
What This Requires
This needs a connected model where calculations are explicit and the hierarchy is defined. The example above is beverage manufacturing, but the same logic applies anywhere operational metrics are built from declared formulas: energy, logistics, healthcare, financial services. The AI applies the model your team already built. It doesn’t build its own.
Cost allocation rules must live in formulas, not in analysts’ heads. Data connections between production and finance must be maintained and timely. The org structure must reflect how metrics actually roll up. And the model needs depth. If Energy Cost is a single input rather than a formula breaking down consumption by equipment, the AI’s trace stops there.
This gets more interesting across sites. When the AI can traverse the org hierarchy, it can compare the real financial cost of efficiency gaps across locations, not just the percentages. We explored how OEE percentages hide dollar impact in a previous article. Formula chain analysis is what lets an AI do that comparison autonomously, without an analyst manually building the bridge.
Companies that have already done this integration work, connecting ERP to MES, building KPI formulas, structuring hierarchies, have already created the intelligence. MCP is what lets AI agents use it.
For more worked examples, see What Did Yesterday’s Production Actually Cost and Why Spreadsheets Lie About Cost Per Unit.
Frequently Asked Questions
What is AI formula chain analysis?
A method where an AI agent traces declared mathematical relationships in an operational model to answer business questions. Instead of inferring correlations from data patterns, the AI follows explicit formulas from input metrics through calculations to financial outcomes. Each step is auditable.
How does MCP enable formula chain analysis?
MCP (Model Context Protocol) is a vendor-neutral standard that lets AI models interact with external data sources. When a platform like Capstone runs as an MCP server, it exposes the formulas and hierarchy that connect metrics, not just the values. Any AI model that speaks MCP can traverse this structure to trace root causes.
Can AI do root cause analysis on operational costs without human guidance?
Yes, provided the model encodes the relationships. The AI traces from high-level KPIs (like Gross Margin) down through each formula to find where the numbers diverge. How deep it goes depends on how deep the model goes. If a cost element is a single input rather than a formula, the trace stops there.
If your finance team spends more than four hours a week tracing levers through spreadsheets, your model might be ready. Book a 30-minute audit of your Capstone instance →