AI for data analysis
Ultramarine Work reads the exports, query results and docs in your project folder, then drafts the analysis, metric definitions and test readouts. Every number in the finished files points to the file it came from.
conversion-mix-rate-decomp.xlsxExample| Channel | Share pre → post | CVR pre → post | Mix (bp) | Rate (bp) |
|---|---|---|---|---|
| Organic search | 42% → 36% | 3.8% → 3.7% | −4.2 | −3.6 |
| Direct | 26% → 24% | 3.5% → 3.4% | −0.8 | −2.4 |
| Paid social | 18% → 28% | 1.5% → 0.8% | −16.0 | −19.6 |
| Referral | 14% → 12% | 2.3% → 2.2% | +1.6 | −1.2 |
| Total | 100% → 100% | 3.10% → 2.64% | −19.4 | −26.8 |
NotePaid social carries 35.6 of the 46.2 bp decline, and its rate fell with the Aug 4 form release, so audit that event tag first.
Ask it for
“Explain why weekly signup conversion fell from 3.1% to 2.6% after Aug 4. Rule out definition and tracking changes first, then split mix and rate effects.”
conversion-mix-rate-decomp.xlsx“Draft the definition sheet for active customer so the board pack and the ops dashboard compute the same number, edge cases included.”
metric-definition-active-customer.docx“Write the checkout test readout: SRM check first, then the primary metric, guardrails, power for the planned MDE, and a ship or no-ship recommendation.”
checkout-test-readout.docx“Forecast weekly orders for the next 13 weeks from three years of history, with rolling backtests, p10 to p90 intervals and driver assumptions.”
orders-forecast-backtest.xlsx“Estimate the effect of the March price change with difference-in-differences, using the two regions that rolled back as controls. Check parallel trends first.”
price-change-impact-memo.docx“Map the census tracts more than 20 minutes from a clinic by transit as rates per 1,000 residents, with the CRS, data vintage and classification stated.”
clinic-transit-access-map.pdf
Inside the Analytics & Data Science mode
Ad hoc analysis and findings
It explains a KPI move only after definition and tracking changes are ruled out, and the mix and rate effects sum exactly to the total.
Operating metrics and reviews
The weekly and quarterly review packs include ARR waterfalls, cohort retention and pipeline coverage, plus a first draft of the commentary on what moved.
Dashboard and metric specification
When two dashboards disagree on the same KPI, the definition sheet pins grain, numerator, denominator, filters and edge cases, so the argument ends.
Data documentation and quality
It documents units, keys, code meanings and nulls in data dictionaries, writes GIS metadata to ISO 19115, and runs checks on the report users distrust.
Ab test design and readout
Every test plan locks the randomization unit, sample size and guardrails in advance, and the SRM check runs before any lift reaches a slide.
Forecasting and time series
Forecasts of demand and volumes come with backtests, prediction intervals and the driver assumptions written down, so planning sees where the model does badly.
Statistical modeling and review
Drafts regression and mixed-effects write-ups with diagnostics and effect sizes, and gives the statistics in someone else’s paper a second read.
Causal and econometric analysis
Where randomization is impossible, it runs difference-in-differences or synthetic control and flags the assumptions most likely to fail.
Model evaluation and documentation
For launch or vendor review, the model card leads with the baseline the model has to beat, then subgroup errors, calibration and leakage checks.
Customer segmentation
Turns customer files into groups the team can name and act on, each with a defensible size and a check that they hold on next month’s data.
Survey design and analysis
Reports weighted toplines and crosstabs with margins of error, significance letters and wave-to-wave comparisons for trackers.
Spatial analysis and mapping
Each map states its CRS, data vintage and classification, and counts are converted to rates per capita when the question is siting or equity.
Data migration and interface mapping
The source-to-target mappings carry transformation rules and reconciliation totals, and the cutover plan names its point of no return.
- Office skills
- Excel, Word, PowerPoint, PDF
- Tools
- Read, edit and write files, Search files, Run code, Search and read the web, View images
- Connectors
- SharePoint, Google Drive, Gmail, and additional custom connectors
- House style
- Your fonts, layouts and rules, on every file
How it works
- 1
Start with the project folder
The folder holds your query exports, CSVs, the tracking schema, the metric dictionary or dbt docs, last quarter’s QBR deck and past test readouts, in whatever shape your team keeps them.
- 2
Ask for the deliverable
It opens the files, runs its own code against them and hands back a workbook or memo. The deliverable spells out the method and the definition it used, and every claim carries a citation to its source file.
- 3
Check it like a peer’s draft
Recompute a number or two by hand, starting from the source file it cites. Check the joins against your metric definitions, and go through its notes on what it could not reconcile before anything reaches leadership.
Who it’s for
Data, BI, product and GIS analysts, analytics engineers, data scientists and the people who lead analytics teams, in any function or industry.
Questions
How do I know its numbers are right?
You check them the way you would check an analyst’s draft. Each number cites the file behind it, the workbooks carry live formulas rather than pasted values, and anything it could not find or reconcile is flagged, with the number left blank. You review every file it touched before the finding goes anywhere.
Does it compute metrics against our definitions, or make up its own?
Once the metric dictionary or dbt docs are in the folder, it uses your definitions and names the definition behind each headline number. Where a metric has no definition, such as what counts as an active customer, it states the assumption it made and shows the alternatives rather than passing a guess off as the number.
Where does our data go when the agent runs?
Ultramarine Work changes files only in the project’s folder. When a request needs it, it can read files elsewhere on your computer, never passwords, keys or app data, and you can turn that off in settings. The model runs on Paper Instruments’ servers, so the content it works on is sent there to be processed. Paper Instruments is SOC 2 Type II, and our security and compliance details are on our compliance page. We make no statement either way about training on customer data, so ask Paper Instruments.
Who is accountable if leadership acts on an AI-assisted analysis?
You are, even though a machine drafted the file. Ultramarine Work drafts, builds, checks and organizes. The sign-off and the professional call are yours. What lands in the deck is what you reviewed and chose to keep, and what you defend when a number is challenged.
Can it run the statistics, or does it just write SQL?
Yes, it runs the inferential work, from regression diagnostics to power calculations, and drafts the write-up with effect sizes, limitations and the checks it could not complete. The judgment stays with you, because a diagnostics table says nothing until a person reads it and decides whether the model is fit to report.