--- title: "Case study: how Suri used internal spreadsheets as evals for their semantic layer" description: "How Suri, a sustainable oral-care brand, treated the monthly tracker finance already trusted as the eval set for its semantic layer, scoring every rebuilt metric against it before a human ever reviewed one." canonical: "https://rig.so/case-studies/suri-metrics" format: markdown --- [All case studies](https://rig.so/case-studies) UK · Series A · D2CSustainable oral care # How Suri used internal spreadsheets as evals for their semantic layer Metric definition was Suri's biggest concern about moving to a warehouse. They've now certified every core metric at speed, and every dashboard and report pulls from that one set. > “Not long ago the answer to almost any question here involved some form of hand-edited spreadsheet export. In just shy of two months working with Rig we've flipped the script on how traditional data infrastructure projects normally go, centralising 95% of the company's sources and getting a solo data function outputting what would usually take a 10+ person team.” TalAI and Data Strategy Manager @ Suri Hear from Tal, AI and Data Strategy Manager @ Suri Tal walked through this workflow in detail at [Rig's client showcase](https://rig.so/events/rigfest-client-showcase). ## Teams using Claude got different numbers for the same question Ingestion was the easy half. Once 40+ sources were in one place and anyone could ask a question, the answer still depended on whose definition of the metric was reached for. Visibility Siloed divisions, no shared view Fast growth split the company into divisions that could not see each other's work. Commercial had no view of marketing or growth, and each function reported off its own spreadsheet, so the same metric meant different things to different people. Definitions Metric definitions were the bottleneck Getting to agreed definitions for net versus gross revenue, contribution margin, AOV and LTV was the hard part. Data consultancies quoted long, expensive, opaque engagements to do it. Attribution Attribution broke at checkout iOS checkouts severed the source chain. Northbeam and Triple Whale promised a clean cross-channel picture but shipped fixed dashboards that were hard to change and not actionable. The unlock ## Your existing reporting is already the eval set Suri already had a spreadsheet the whole company trusted: a monthly tracker finance kept, estimating where revenue was landing against target. Rather than treat it as legacy to be replaced, Tal treated it as ground truth, and the numbers it held became the test every new definition had to pass. Existing dashboards went through Claude and Rig together to map the business logic behind every definition. Rig then rebuilt each metric from the warehouse and scored it against the tracker, only passing it to a human when the two landed inside an error boundary. Sign-off stopped being a quarter of back-and-forth between data and finance and became one 50-minute meeting. ## How they did it with Rig Rig ingests Suri's commercial stack into one governed context layer, joined on Shopify orders. Certification then ran as a loop rather than a project: every definition is mapped, rebuilt and scored before a human is asked to look at it. Step 1 Map the logic with Claude and Rig Suri's existing dashboards go through Claude and Rig together, so the business context and the warehouse are read side by side, and the logic behind every definition across the company is written down rather than remembered. Step 2 Score it against the tracker Rig rebuilds each metric from the warehouse and scores it against the same figure in the monthly tracker. Agree inside the error boundary and it goes to a human; miss and it never reaches one. Outside the boundary it goes straight back to step 1, and nobody is asked to review it yet Inside the boundary Step 3 A human signs it off Only then is the definition certified. Running the tracker as the eval set also caught errors in the tracker, including SKUs missing their cost of goods and costs being rolled up into revenue that did not belong there. Flowchart: every definition is mapped with Claude and Rig, then rebuilt and scored against the monthly tracker. Scores outside the error boundary return to step 1 without reaching a person; only scores inside it go on to step 3, where a human signs the definition off and it is certified. ### Then, on top of the certified layer Once the definitions were signed off, the same layer carried two more things. Access Rig Connect in Claude - Anyone can query the warehouse in Claude, with no queue behind the data team - Certified definitions travel with every answer - Numbers stay consistent across finance, marketing and growth Attribution Attribution & marketing planning - Own cross-channel attribution model, geo-split MMM to read the incrementality of spend - Tracks spend and volume around peaks like Prime Day, derived from Klaviyo - Marketing, influencer and leadership summary dashboards ## The impact | Area | Before | With Rig | | --- | --- | --- | | Metric definitions | Months of consultancy back-and-forth, and definitions drifting between teams | Auto-evaluated against existing reporting, human-reviewed only on match | | Cross-team access | Everything queued behind the data team | Self-serve via Rig MCP in Claude, on certified metrics | | Attribution | Fixed Northbeam and Triple Whale dashboards, with iOS gaps | Own cross-channel model with geo-split MMM for incrementality | > “Not long ago the answer to almost any question here involved some form of hand-edited spreadsheet export. In just shy of two months working with Rig we've flipped the script on how traditional data infrastructure projects normally go, centralising 95% of the company's sources and getting a solo data function outputting what would usually take a 10+ person team.” TalAI and Data Strategy Manager @ Suri ## Want results like this? [Book a demo](https://rig.so/book-demo)[See how Rig works](https://rig.so/walkthrough) --- - [This page as HTML](https://rig.so/case-studies/suri-metrics) - [Site map for language models](https://rig.so/llms.txt) - [API and agent documentation](https://rig.so/developers)