Putting Sensitive Employee Data to Work, Without Lowering the Security Bar

A people analytics team needed to answer questions leadership cares about, like whether one group is leaving faster than others or getting promoted more slowly. The demographic data behind those answers sat in a separate Snowflake account under the company's security rules, so every answer meant working across two accounts. Qbiz designed cell-level encryption that lets that data move into the main account without weakening its protection, and worked the design through every team that owned a piece of the security model.
AT A GLANCE | |
Client | A large enterprise whose small people analytics team handles some of the company's most sensitive employee data. |
Business challenge
| People analytics couldn't reliably answer leadership's questions about retention, promotion and workforce demographics without manually combining data from separate systems. The company's security requirements kept its most sensitive employee data apart from the workforce data needed to make those decisions. |
What Qbiz did | Designed cell-level encryption that runs inside Snowflake, built a working prototype and production-ready pipelines alongside Snowflake's resident solutions architect, and brought the security, reporting and platform teams into the design until all of them approved it. |
Business impact | Questions like whether one group is leaving faster than others, or getting promoted more slowly, used to mean combining data from two separate Snowflake accounts by hand. The design gives people analytics a path to identify retention and promotion gaps across employee groups and deliver consistent EEO reporting from one governed source, without compromising the protection of sensitive employee data. |
Technical and Operational impact | The client can consolidate people data into its main Snowflake account, retire the separate account and apply the main platform's governance, lineage and features to sensitive data. |
Technology |
The data behind workforce decisions was the hardest to use
The client is a large enterprise, but its people analytics team is small. That team handles work the business can't afford to get wrong: EEO reporting, workforce population studies, and other analysis that depends on employee records from Workday, including demographics like ethnicity and gender.
Part of that data was rated above what the company's main Snowflake account was approved to hold, so it sat in a second account set up for it alone. That met the security requirement, but it split people analytics in two. Combining demographics with the rest of the workforce data meant working across two Snowflake accounts.
Workday data landed in the separate account, and only selected fields were shared to the main account for broader use. Any analysis that needed demographics alongside headcount, levels, internal mobility or attrition had to work around that split. A question like whether attrition in one org differs by gender meant pulling from both accounts and lining the results up by hand. The second account also got new features months or even a year after the main one, and with changes to single sign-on and multi-factor authentication on the way, keeping it running was about to take even more of a small team's time.
The goal was to let people analytics use that data like any other workforce data, so questions about retention and promotion could be answered from one place, without weakening how it was protected. Getting there meant working through real trade-offs between security and reporting, and bringing in every team that owned a piece of the answer.
Their rule: tier 4 data can't sit in a tier 3 platform. Encrypted, it can.
The client classifies data by sensitivity. A work email sits low on the scale and a personal email sits higher. Demographics like ethnicity and gender sit at tier 4. The main Snowflake account was approved for tier 3, and under the client's policy, encrypted tier 4 data counts as tier 3.
That made the real question how fine-grained the encryption had to be. The answer was cell level. Every value gets its own data key, so a single exposed key reveals one value, not a column.
Our job was to turn that rule into a design the client's security, governance and platform teams could approve and support, not just one that technically met the policy.
Easy to encrypt, deliberately harder to decrypt
Five decisions shaped the architecture. The core idea is simple: the data stays put, and only keys travel.

Encrypt where the data already lives. A small web server running in Snowpark Container Services holds the public keys used for encryption, so encryption happens inside Snowflake, locally and fast.
Check identity once per request. Before anything is decrypted, an identity check confirms the user is allowed to use the decryption key.
Send keys out, keep data in. Data keys go to AWS KMS in batches to cut down on network calls, and KMS unlocks each one individually. The container then decrypts each value locally. Only keys make the trip, so raw data stays inside Snowflake by design.
Built from standard parts. Docker, AWS KMS and a generic call to an identity and access management (IAM) service, built alongside the client's Snowflake resident solutions architect.
Separate systems for separate jobs. Identity checks, key decryption and value decryption run in separate systems, so compromising any one of them on its own doesn't expose sensitive data.
The harder problem was getting every team to yes
Every "no" was protecting something
At least six teams owned a piece of the security model, and each had to approve the design: information security, governance, security, the internal Snowflake platform team, engineering, and people analytics.
Their requirements competed. Security wanted as few records decrypted at a time as possible. People analytics needed every row for its reporting. The platform team needed something it could support alongside everything else in the main account. Each requirement protected something real, and the client was right to be careful with its employees' most personal data. The work was folding those requirements into one design, then taking it back to every team, round after round, until all of them signed off. Getting that approval was as much of the project as the encryption itself.
One thing helped more than any diagram. When the platform team kept turning down proposals, we asked what they would do instead. Once they put their own approach on paper, we had something to build from, and it became their design too. By the final rounds, they were the ones defending it in the room.
The hardest trade-off: every row, under the same controls
The client's security guidance says to decrypt a few records at a time. EEO reporting and workforce population studies need every row in a column. Both requirements were legitimate, so the design had to hold both.
Full-column reports run through the same controls as any other request: an identity check up front, and every value unlocked with its own key. That makes them slower than they would be with one key per column. An engineer on the project wanted to speed that up. We made the case to keep the per-value design and accept the longer run time, so reporting gets every row it needs without a shortcut that unlocks a whole column at once. For the business, that's what makes the numbers defensible. An EEO report built from every row went through the same controls as a lookup of a single record.
The lesson we took from it: decide up front which uses legitimately need everything, and agree on what serving them will cost. Design only for the worst case, and the reporting gap shows up later as a surprise.
Shipped, and ready to put data to work
Delivered. An approved design, a working prototype, and production-ready pipelines. The client's team has already customized the solution for their environment, swapping in their own key management wrappers and identity service, so it runs on tools and infrastructure they already operate.
Enabled. With encryption in place, the client can retire the separate account and run all people data on one platform. That gives it:
For people analytics: demographics sit alongside headcount, levels, mobility and attrition data in one governed source, so retention, promotion and EEO questions get consistent answers
For the platform: one Snowflake account instead of two, with the main account's new features, governance tools and lineage applied to the most sensitive data
It also leaves the client better positioned for what comes next. Sensitive people data that lives in the governed main account, under controls its teams already approved, is a stronger starting point for future AI applications than data kept off on its own.
Revisit the architecture without relaxing the requirements
Many of the companies we work with keep their most sensitive data in a separate platform, often for good reason, as this client did. But the reasons get set once, and the platform keeps moving. Meanwhile the questions leadership asks about retention and promotion get harder to answer from data that sits apart. Others have the opposite problem: their HR data already sits in the main account, without the protection it should have. Either way, the questions are the same: Does this data need to move, and what's the secure path for it? Both are cheaper to answer before a deadline forces them. If you're weighing whether your sensitive data can live alongside everything else, we can help you work out whether it should move and design a path your security teams can approve.



