Golden Data

Issue #26 · October 18, 2026

The Planning Team Used To Be Three. It's One Analyst Now, Snowflake Says.

The spreadsheets were the team.

Paul BrownWrites Golden Data from Ogden, Utah5 min read

On Snowflake’s second-quarter earnings call on September 2, CEO Sridhar Ramaswamy said long-range financial planning used to take a three-person team and more than fifty spreadsheets. It now runs on one analyst and a series of models built around the company’s pricing structure and consumption dynamics. He said it in the product section of the call, between the marketing team’s keyword research dropping from ten hours to twenty minutes and 70% of first emails to inbound leads being written before an SDR touches them. Finance was the second item in a list of three.

The tell is the placement. Snowflake did not report the planning change as a finance outcome. It reported it as a product demo. The CFO’s own section, later in the call, carried the number that makes it matter: 334 employees added year to date, 173 of them from the Observe acquisition, against 935 added in the same stretch a year earlier. Non-GAAP operating margin reached 15.3%, up from 11.1% in the same quarter last year. CFO Brian Robins credited strong revenue growth and disciplined headcount management, with slowing headcount expense offsetting growing cloud costs.

A CFO should read that twice. The analyst who owns a forecast should read it three times, because the question is not whether the planning team shrank. It is what the two seats that went away were holding, and where it went.

The teardown: Snowflake on Snowflake

Give Snowflake its due first. This is not a vendor describing a customer. It is a company reporting on itself, on the record, in the same call where it raised full-year product revenue guidance to $6.07 billion. In the Q&A, Robins said adoption inside the finance organization is almost 100%, and named deal desk, tax, accounting, internal audit, FP&A and treasury as places he uses it. The margin expansion is in the press release. The headcount is in the CFO’s prepared remarks. When a company that sells the platform the models run on says its own forecast runs leaner, the burden is on the skeptic.

Eval surface. What is public is a headcount, not an accuracy figure. Nobody on the call said the forecast got better, tighter, or faster to close. No variance to actuals, before or after. No statement of who reviews the models’ output before it reaches the board. The sentence proves the planning function can be staffed at one. It does not prove that one is enough, because the only test disclosed is that the company kept running.

Cost exposure. Fifty spreadsheets is not a mess. Fifty spreadsheets is where three people kept what they knew: the renewal that always slips a quarter, the region that books in a different currency, the customer whose consumption curve is nothing like the cohort’s. Some of that moved into the models. Some of it went wherever the other two seats went, and the call did not say whether that was another team or out the door. The models now hold the institutional memory, and one analyst holds the models. That is a cheaper line on the income statement and a single point of failure in the forecast. The call reported only the first.

Data shape. Snowflake’s planning models forecast consumption of Snowflake, using Snowflake’s own consumption data, on Snowflake. The inputs are governed, metered, and already in one place. On the same call, the CEO introduced Cortex Sense, which he described as capturing the business definitions and institutional knowledge an agent needs. That is a fair description of what the fifty spreadsheets held. The three-to-one compression happened on top of the cleanest planning dataset a model will ever see. Your long-range plan reads a CRM and an ERP that disagree about what a customer is. The compression transfers only as far as the definitions do.

What I’d do this week

User moment: The CFO forwards the transcript and asks what one analyst plus models would look like here. The scope is a planning function that does not run on its own product’s telemetry.

I would not start by consolidating spreadsheets. I would start with a backtest. Pull the last eight quarters of the current team’s long-range forecast, line by line, against actuals. Rank the lines by absolute variance. Most of the error lives in three or four of them. Those lines get a model. The rest stay in whatever spreadsheet already works, because a spreadsheet that has been right for eight quarters is not the problem.

Shape: One model per high-variance line, built on the governed definition of that line, not on the spreadsheet’s local version of it. Each model emits a number, the two inputs that moved it most, and a confidence band. It runs on the close calendar, in the warehouse the finance data already lives in. It is a planning tool, not a dashboard. One person who did not build the model signs the number before it leaves finance.

Time budget: A day for the backtest. One analyst’s quarter for the models it picks, which is the only new spend. If the old versions weren’t kept, that is the first finding, and the backtest starts next quarter.

Artifact: The variance table: eight quarters, every line, ranked by error, with the three or four modeled lines marked. From then on, every forecast a model emits gets logged against the actual, by line, by quarter, with the two drivers it named.

The variance gate. A model replaces a spreadsheet only after two quarters of shadow forecasting in which its error on that line is at or below the human team’s. Not on average. On that line, at the horizon the board uses. If it clears the gate, the spreadsheet retires and the person who kept it signs off on what the model got right and what it still misses. If it doesn’t clear, the spreadsheet stays and the model keeps shadowing.

Success: Two quarters in, a variance table where the modeled lines are tighter than they were under three people and the spreadsheet lines are unchanged.

Credible failure: A planning function that is cheaper and no more accurate. Cheaper is the only part Snowflake’s call proved.

Sources

Crafting runs weekly. The job is to be useful, not to sell. If a line in here is close to a move you're working on, the email reaches a person.

Hiring rather than hunting? Signal is the column for your side of the desk.

Email meRead more Crafting

← All Crafting issues · Drafted with Claude · Edited by Paul Brown