How Finance Cut Forecasting to Hours With AI

Snowflake's sales finance team cut its monthly forecast and variance analysis from more than a week to a few hours using a CoCo profile that supplies standing context about the team's data, terminology and workflows, according to a Snowflake account of internal usage. The profile reduced steps per answer from roughly 3.0 to 1.3, lowering the cost of each question, based on Snowflake internal CoCo usage data from February to May 2026 across 44 finance users. The team built 36 skills in two months, and 87% of first adopters still use CoCo daily, with Senior Director of Sales Finance Gayatri Kannan saying the tool "has truly become the first choice for all of our team's routine workflows and ad hoc analysis.

On Snowflake's sales finance team, we used to spend more than a week developing our monthly forecast. Today, it takes a few hours. What changed was how much AI knows about our data before anyone types a question. That knowledge also made our AI more efficient. Our analytics team built a profile that gives the AI standing context about our data, terminology and workflows. With it, Snowflake CoCo https://www.snowflake.com/en/product/snowflake-coco/ reaches an answer in less than half the steps, so each question costs less even as the team asks more of them. This post covers the CoCo profile behind those results, why it makes each answer more efficient and how our analysts built 36 skills on top of it. Key takeaways - Snowflake's sales finance team cut its monthly forecast and variance analysis time from more than a week to a few hours. - A CoCo profile cut the steps per answer from roughly 3.0 to 1.3, which lowers the cost of each question Snowflake internal CoCo usage, February to May 2026, n = 44 finance users . - Every answer traces back to a data source certified by the analytics team. - The team built 36 skills in two months, and 87% of first adopters still use CoCo daily. A weeklong forecast process now takes hours The sales finance team runs more than 30 recurring analyses each month for Snowflake's sales leaders, covering the topline forecast, sales pipeline, commissions and business performance reporting. The forecast is the most involved of these, and it comes with a variance analysis that explains what's moved since the previous forecast. In the past, building the forecast meant reviewing pipeline deals, reading account notes in the CRM, comparing against historical trends and checking current field behavior. The work was split across several analysts, and because it took more than a week, it could only run once a month. Using the topline forecast skill in CoCo, we now complete the same analysis in a few hours. That faster turnaround has changed the cadence. We now publish a midmonth update, so sales leaders see where the quarter is heading while there is still time to act. Any analyst can also rerun the skill to test a scenario. "CoCo has truly become the first choice for all of our team's routine workflows and ad hoc analysis. Not only has it reduced turnaround times; it has unlocked the ability to process complex analysis. Any individual can use the skill to dive deeper and run scenario analysis." Gayatri Kannan Senior Director of Sales Finance, Snowflake Capable AI only helps when it knows your data Our data changes daily, spread across dozens of certified sources and a CRM that reflects deals revised overnight. Our first attempts were familiar to anyone who has pointed AI at real finance work. The AI was capable, but it didn't know our world. It would reference a related but incorrect data source or ask for clarification that an experienced analyst wouldn't need. Both failures came from missing context: Every session, the AI had to infer things we knew but had never written down. "When we first launched AI to the team, there were a lot of struggles with inconsistent and inaccurate results. It felt like we had to constantly remind a new hire about relevant terminology and data sources. Having a dedicated GTM Finance profile changed everything. The team is empowered to run complex analyses and rethink workflows themselves with minimal technical support and no governance concerns." Gayatri Kannan Senior Director of Sales Finance, Snowflake What a GTM Finance profile actually contains Our partners on Snowflake's strategic finance analytics team built our CoCo profile for 'GTM Finance'. A profile is standing context the AI reads before it answers anything, so nobody has to restate it in a prompt. Without it, the AI would spend about three steps per session finding the right data source, resolving terminology and inferring what an experienced analyst already knows. With the profile in place, the first step of every conversation is already the work itself. | Layer | What it holds | Why finance needs it | |---|---|---| | Certified data sources | 20+ sources certified by the analytics team | Answers come from sources the analytics team has approved | | Terminology | Definitions for bookings, pipeline, quota and other core metrics | No ambiguity on reported metrics | | Repetitive analyses | Our recurring monthly and quarterly work | The AI recognizes the task | | Output formats | Our standard report and variance layouts | Output arrives ready for review | | Escalation paths | Where to go when a number looks wrong | A person resolves exceptions | Memory scaling: Why a profile makes each answer more efficient Memory scaling is the effect where an AI agent gets cheaper per unit of work as it accumulates context about its environment. It answers in fewer steps, and each step uses compute. Across 44 finance users, those with a profile reached an answer in about 1.3 steps, compared with about 3 steps for those without one. Two things make each answer more efficient. The profile cuts the number of steps, and steps drive cost: Without a profile, each extra step adds about 0.05 credits roughly 10 cents , so a five-step session costs about three and a half times a one-step session. The profile also makes sessions cheaper at the same length. At every step count in the data, a session with the profile cost less than one without it. Finance analysts now build their own AI skills Once the foundation was in place, our analysts could build their own skills on top of it, without relying heavily on the analytics team. A skill is a reusable workflow the AI recognizes and runs on request, and building one requires no technical background. In two months the library grew from three starter skills to 36, and several Excel workbooks became shared skills the broader team could run. Three months after launch, 87% of first adopters were still using CoCo daily. Here's a selection of the skills we built. | Skill | What it replaced | |---|---| | Bookings BvA | Manual month-end variance analysis across data and finance | | Sales travel expense report | Hours of cross-team reporting and manual email drafting | | Topline forecast | Weeklong multiperson process | | Commission reconciliation | Excel-based manual workflow | The reason this works is ownership. The analyst who runs commission reconciliation 15 times a month understands that workflow better than anyone building it from the outside. Our corporate FP&A team made the same argument when they wrote about who gets to build finance applications https://www.snowflake.com/en/blog/cortex-code-fpa-real-time-insight/ , and our skill library is what that shift looks like at scale. Bookings BvA Our corporate FP&A colleagues wrote about a similar build, their automated variance analysis workflow https://www.snowflake.com/en/blog/finance-variance-analysis-workflow/ , including the AI-generated variance commentary and the governance model behind it. Bookings BvA applies the same idea to bookings: An analyst types "bookings bva" into CoCo, answers a few parameter prompts and has month-end variance across deal categories in minutes, including which opportunities slipped and why. It is the clearest example of a skill that replaced a recurring cross-team assembly job. Sales travel expense report Travel is one of the largest controllable cost lines for a field sales team, and accountability requires regular visibility. The sales travel expense report skill determines the reporting window on its own; generates per-leader insights covering budget variance, out-of-policy spend and top spenders; and then drafts the email to each leader. Work that took hours across two teams now takes about 15 minutes, and any team member can run it, with more consistent insights than the manual version produced. Keeping control when analysts build their own skills Opening skill-building up to a team raises an obvious question for anyone accountable for the numbers: What happens to control when analysts write their own logic? It is the same concern behind the trust and auditability requirements finance leaders bring to AI https://www.snowflake.com/en/blog/transforming-modern-cfo-role/ , and it has to be answered before any of the numbers above matter. Three things keep it in bounds: - The analytics team certifies the source of truth: A technical source guide underpins the profile and defines which data sources to use and why. Follow-up questions and ad hoc analysis draw on those same sources, so answers stay consistent and traceable. - The context layer is owned centrally: Analysts build on top of the profile; the analytics team defines the data sources and terminology underneath. - Skills are shared and inspectable: A skill that runs commission reconciliation is visible to the team that depends on it, which is a stronger review position than a spreadsheet on one laptop. How to replicate this on your team Our approach applies beyond finance: Platform experts set the foundation and a functional team creates its own skills on top of it. Three steps get you there. 1. Certify your sources first: Decide which sources are authoritative for each question your team answers, and write down why. If you skip this, then every downstream workflow inherits an unresolved question about which number is right. 1. Encode terminology and output formats second: Definitions your team treats as obvious are exactly what the AI cannot infer. Capture them once, centrally. 1. Then open skill-building to the team: The people closest to a workflow build the best version of it once the layer underneath them is trustworthy. If you reverse that, then you get fast workflows built on sources nobody agreed on. This approach supports work well beyond recurring reporting. Our corporate FP&A team has done the same for long-range planning, asking questions of the plan in natural language. https://www.snowflake.com/en/blog/fpa-ai-modernization-snowplan-v2/ Without standing context, every conversation starts from zero. What separates a team that has an AI tool from a team where AI is central to daily work is usually how much the AI knows about the work. Learn more at Snowflake Expedition Learn more about how finance teams are transforming how they work using Snowflake. Join us at Expedition https://www.snowflake.com/en/expedition/ams/agenda/?fu=leadersfinance , our global virtual event on Nov. 3-5, 2026. Data: Snowflake internal CoCo usage, February to May 2026, comparing finance users with and without a profile n = 44 users . 1 credit = US$2.