Every data project has the same hidden cost. Before a pipeline gets built, someone spends two days emailing a client admin about access. Before a dashboard goes live, someone spends hours figuring out what a confusing report measure actually means. None of this is hard work. It just takes time, and nobody has automated it, because it depends on the account and the situation.
This is the gap that Snowflake Cortex Code (also called CoCo) is built to fill. Built on Snowflake Cortex AI, it offers a set of 21 skills, grouped into six common tasks that clients ask for. Each skill follows the same rule: it writes out the steps for a person to check and run; it never runs anything on its own, and it never makes up a table, schema, or role name that does not exist. And this is not just a nice claim. We ran a real build and measured it against actual usage data.
The main result: On a real dbt project (12 raw tables, 21 models, full tests, a semantic layer, and access setup), CoCo finished the work in about 3 hours instead of an estimated 17 to 28 hours of manual work, roughly 75 to 80 percent faster, at close to 90 percent lower cost (about $95 versus about $1,700). At a bigger scale (5 million rows, 40 to 60 models), the pattern holds up: 1 to 3 days instead of 2 to 3 weeks, still around 75 percent faster, even though the total savings grow larger in dollar terms. The point for the business is simple: the bigger and more complex the project, the more this approach helps, not less.
The Core Sequence: Building the Pipeline
The main use case is data pipeline automation - building a full pipeline, from raw data all the way to analytics, using a Snowflake Cortex AI skill called Snowflake-pipeline that manages each stage one step at a time. Some steps run every time; others only when needed. It works like a checklist: load the data, model it, test it, schedule it, then make it easy to query in plain English.
Time saved: roughly 75 to 80 percent on a small project (3 hours instead of 17 to 28 hours), staying close to 75 percent even on a much bigger one (1 to 3 days instead of 2 to 3 weeks). The biggest time savings come from the parts nobody enjoys writing by hand, like a dozen staging models, the config files every project needs, and the first full round of tests. Together, those alone eat up 8 to 12 hours in a manual build.
Before the Pipeline: Access and Onboarding
Before any build starts, someone has to get access and understand a data source they have never seen before. One skill writes out a least-privilege access request for the client's admin to approve, explains why each permission is needed, states clearly that full admin access is not being requested, and includes a simple way to remove access later. It also catches a common mistake: forgetting to include future tables, which works fine at first and then quietly breaks later when the client adds new data. A second skill handles the first look at a new source, checking things like whether records change over time, what the unique key is, and whether there is sensitive data, because those answers shape everything that gets built afterward.
Time saved: this covers tasks like profiling raw tables (1 to 2 hours by hand) and setting up access (30 minutes to an hour by hand), both cut down to a quick review, in line with the roughly 75 percent time savings seen across the build. In practical terms, that is the difference between losing the first week of a project to access requests and losing an afternoon.
The Bookends: Takeover and Handover
Every project has a beginning and an end, and both are usually poorly documented. Reviewing work left behind by a previous team, and writing up your own work before you leave, are the two moments that shape whether a client trusts what comes next, or ends up disputing the invoice. These skills produce real documents you can point to: one works as paid discovery work, the other keeps the last week of a project from turning into an argument about what was actually delivered.
Time saved: likely in the same 70 to 80 percent range as the rest of the framework. Writing documentation, like schema summaries, handover notes, and access summaries, is exactly the kind of repetitive writing that this data shows CoCo handles quickly.
Getting Logic Out of BI tools
One of the least exciting but most valuable use cases is Tableau to SQL migration - pulling calculation logic out of Tableau and Power BI reports and rewriting it as SQL that can actually be tested. BI tools build up hidden logic over time that nobody can track or check properly. The core problem is that a report measure often means different things depending on which chart it is used in, so a number labeled 'net revenue' might be one total on a summary card and a yearly breakdown on a chart, and both are technically right. Writing it in SQL forces one clear definition, so the number means the same thing everywhere it is used.
Time saved: similar to the roughly 75 percent seen on complex, back-and-forth work at larger scale. Translating BI logic into SQL is exactly this kind of detailed, iterative work, the same category where this data shows the time savings hold up well even as things get harder.
When Something Breaks
The last use case is the one nobody wants to need: the numbers on a dashboard are wrong and there is a meeting in an hour. Here the value is not cleverness, it is having a clear order to follow, a fixed set of steps that replaces guessing with a checklist, right when it is easiest to skip steps under pressure.
Time saved: this data shows that debugging alone takes 2 to 4 hours on a small project and 8 to 16 hours at a larger scale, the single biggest source of manual time, so a fast and reliable troubleshooting process likely saves the most time of any use case here, even before counting the cost of the incident itself.
The Pattern Underneath
Snowflake Cortex AI continues to expand what's possible in automated data engineering. None of these six use cases are unusual. As data pipeline automation tools mature, they are simply the parts of an AI data engineering project that are too specific to fully automate with a script, and too repetitive to keep doing from scratch each time: access requests, source reviews, pipeline builds, handover notes, BI migrations, and troubleshooting. Measured against real usage data, this approach cuts delivery time by around 75 to 80%on typical projects and close to 75%at large scale, while cutting cost by around 90%and 80%respectively. It holds up because the cost of this approach depends on how complex the conversation and the models are, not how much data there is, while manual work scales with both. That is a smaller claim than saying AI replaces data engineers, and a more useful one: it takes over the parts of the job that eat up hours without adding much value, so the team can spend their time on the parts that actually need judgment.
See how Snowflake Cortex Code can cut your team's delivery time by up to 80% and reduce project costs by 90% - talk to Mastek's Snowflake experts about what CoCo can automate for your data projects.
