How We Turned a 10-Year Financial Forecast into a Conversational Bug Tracker

From Excel Frankenstein to a Real Planning Platform

Long-term planning is the kind of work that quietly consumes a finance team. At Snowflake, the 10-year forecast covered over 40 entities, each with more than 100 cost centers and hundreds of expense categories. That granularity wasn't a nice-to-have—it was a dependency for tax, treasury, and HR teams. The model had to answer completely different questions for each group.

So we did what most finance teams do when the tooling breaks: we built a massive Excel workbook. It grew tabs like a fungus. We patched formulas on top of formulas. New logic got layered over old logic to keep up with a changing business. It worked, but it was a Frankenstein. Maintenance was a nightmare, governance was a joke, and scaling meant more duct tape.

Why We Moved the Model to Snowflake

About a year ago, we rebuilt the whole thing on Snowflake, with Streamlit as the UI layer. We called it Snowplan. The idea wasn't to build a dashboard—it was to build a planning platform. Analysts could edit assumptions directly in the interface, and the changes would write back to Snowflake, run the model, and return fresh outputs instantly. No more broken formulas, no more version confusion, no more hunting for the source of truth.

That architectural shift changed the entire planning process. The model now lives where the data lives. Actuals flow in automatically, assumptions get versioned, and different roles see exactly the level of detail they need. Individual contributors get granular input pages. Managers see logic changes for review. Executives see the consolidated P&L and free cash flow. It's not just a model anymore—it's a governed system that supports multiple planning workflows.

CoCo Makes Scenario Planning a Conversation

Streamlit made Snowplan usable. CoCo made it conversational. Before CoCo, you still had to know where to click, which assumption to tweak, and how to interpret downstream effects. Now I can just ask: compare two forecast versions and summarize the key drivers. What changed between the plan we showed the board last year and the latest version? What's the net impact on margins? Which assumptions should we worry about?

That's huge for board prep. The real question isn't “Can you pull the latest numbers?” It's “What changed, why, and what does that mean for our story?” CoCo compresses what used to be manual analysis into a back-and-forth conversation. And because it's built on Snowflake tables, CoCo can actually trace which sales are affected, compute financial impact, and show the metrics behind the numbers.

A Real Example: Planning Around a Tax Change

Here's a concrete case. We were exploring a potential tax change. In the old world, this would start with a meeting: define affected sales, pull data, build assumptions, update the model, review outputs, do sensitivity analysis, then figure out who else needs to be involved. With CoCo, it became fluid. I could ask CoCo to summarize the tax change, then create a new forecast version assuming it passes.

That immediately triggered the kind of back-and-forth finance teams do in meetings: Is the tax passed to customers or absorbed as margin loss? What percentage can realistically be passed through? Which sales are affected? What's the impact on revenue, gross margin, operating margin, and free cash flow? CoCo can answer all of that, generate sensitivity tables showing how operating margin dilutes under different pass-through rates, and even flag risks—like the indirect costs of compliance that a first-order model might miss.

It doesn't stop there. CoCo can help draft an email to the tax team summarizing findings, assumptions, open questions, and decision points. So it's not just producing a number—it's helping move the whole process forward. That's AI-assisted planning, not just AI-assisted modeling.

Why Trust Is the Real Foundation

For conversational planning to work, the numbers have to be trustworthy. That's why architecture matters. CoCo doesn't hallucinate a forecast out of thin air. It interacts with the same governed data, assumptions, and logic that power Snowplan. Every scenario is versioned, every change is reviewable, and access control follows the app's role model. Analysts and executives can compare before and after, understand what changed, and roll back if needed.

This is the key difference: we're not asking leaders to trust a black box. We're using AI to operate a well-governed planning platform where data, business logic, permissions, and outputs are all visible, explainable, and auditable. That's what makes AI acceptable in enterprise finance.

From a One-Off Tool to a Strategic Platform

The exciting part is that Snowplan has outgrown its original use case. Once the model moved to Snowflake, the architecture became reusable. The same foundation now supports headcount planning, equity modeling, treasury cash forecasting, hedging, legal entity forecasting, COGS planning, and M&A scenario analysis. Each new workflow reuses the same governance base, connects to relevant data sources, and exposes a user-friendly Streamlit interface. With CoCo, each one can be queried, adjusted, and explained in natural language.

I think more finance teams will follow this path: first, move the model to where the data lives; second, build an intuitive app layer; third, make planning conversational with AI. The real ROI isn't making finance more technical—it's giving time back to judgment.

The Bottom Line

Long-term planning was never just a modeling exercise. It's an organizational alignment process. The less time finance spends maintaining models, the more time it can spend challenging assumptions, aligning executives, and shaping strategy. Snowflake and Streamlit gave us a platform that scales, governs, and connects to real-time data. CoCo made it faster and more interactive. For FP&A teams, that's the real value of AI: it removes the manual labor that drags finance down and lets them focus on what actually matters.