AI for operations research analysts
Ultramarine Work drafts formulations, fits input distributions, builds capacity models and writes the memo. It works from the extracts, workbooks and prior studies already in your project folder.
Scanner-trade-study.xlsxExample| Vendor | Weighted score | Rank | Lead over next | Sensitivity |
|---|---|---|---|---|
| Norvik | 7.58 | 1 | 0.10 | Kestrel passes it if integration goes to 0.26 and scan rate to 0.24. |
| Kestrel | 7.48 | 2 | 0.02 | Brightway passes it if TCO goes to 0.21 and scan rate to 0.27. |
| Brightway | 7.46 | 3 | 0.36 | No shift of 0.05 between two weights drops it to fourth. |
| Tallis | 7.10 | 4 | It needs a shift of at least 0.12 between two weights to pass Brightway. | |
| AHP weights | CR 0.04 | The consistency ratio is under 0.10, so the pairwise judgments hold. |
NoteKestrel and Brightway are 0.02 apart, and a 0.01 weight shift reverses them, so present them as tied for second.
Ask it for
“Build the overnight replenishment MIP: 46 stores, 2 DCs, shelf caps, a 12,000-unit pick limit per DC, minimizing transport and stockout cost.”
Replenishment-MIP-model.xlsx“Fit arrival and service distributions from 6 months of ticket exports, then recommend warm-up and replication counts for the queue study.”
Input-modeling-plan.xlsx“Score the four in-cab scanner vendors against our six weighted criteria, compute the AHP consistency ratio and show which weight changes flip the ranking.”
Scanner-trade-study.xlsx“Convert the 32-bed med-surg ward’s 91% occupancy and 8.5 HPPD standard into RN and aide FTEs using 1,768 productive hours per FTE and an 80% RN skill mix.”
FY27-staffing-plan.xlsx“Run Erlang C by half hour for the billing queue at 80/20, cap occupancy at 85%, gross up for 30% shrinkage and mark every interval where the ceiling binds.”
Billing-queue-staffing.xlsx“Explain why the 6 to 8 pm missed-pickup rate rose from 4.2% to 7.9% in Q3: rule out the May tracking change, then split mix and rate by market and reconcile.”
Q3-missed-pickup-findings.docx
Inside the Operations Research mode
Optimization model
An allocation, scheduling or routing problem arrives as a stated objective and constraints, and you read the binding constraints and the shadow prices.
Simulation study
It fits the input distributions, checks the warm-up, runs enough replications to defend the confidence interval and says what was simulated and what was not.
Decision analysis and trade study
The weights decide a trade study, so it scores the alternatives, tests the consistency ratio and shows which weight change overturns the recommendation.
Workload capacity and staffing model
Give it the volumes and their standards (HPPD, minutes per case, caseload caps) and it returns FTEs, shifts and overtime, and names the bottleneck.
Contact center staffing model
Erlang C sizes every half hour and channel against the 80/20 target, the occupancy cap and shrinkage, and it flags the intervals where the ceiling binds.
Forecasting and time series
Forecasts come with rolling-origin backtests and prediction intervals shown next to the point forecast, and the outlook brief states its driver assumptions.
Ad hoc analysis and findings
Before it splits the move into mix and rate by segment, it rules out the definition and tracking changes, and every piece has to reconcile back to the total.
- 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
Open the project folder
Drop in the data extracts, the constraint list your ops lead wrote, last year’s model workbook, the memo that set the service standard and the study your director approved.
- 2
Ask for the model or the study
It states the objective and constraints in words and in math, builds the workbook or script and runs the numbers. The file comes back with a source note on every input and a flag on anything it could not reconcile.
- 3
Validate before you brief
Test the cases whose answers you already know, then walk the route or roster that failed on paper. Read the shadow prices and decide, because the solver solved the model you gave it, and only you can judge whether it matches the operation.
Who it’s for
Operations research, optimization, simulation and decision analysts in logistics, airlines, manufacturing, healthcare operations, energy, defense and consulting.
Questions
Can I trust a formulation it drafted, or do I rebuild the model myself?
Check the draft the way you would any first formulation, because it can miss a constraint your dispatchers or floor leads work around informally. You can see the decision variables, objective and data assumptions written out. Its skills tell it to cite the file behind each number and to note what it could not reconcile. Feed it test cases with known answers and confirm the model stays feasible before it goes anywhere.
Will an AI-assisted model pass our model risk validation?
Your independent validation group decides that. The 2026 model risk guidance from the FRB and OCC describes validation as conceptual soundness, effective challenge and outcomes analysis, and it explicitly excludes generative and agentic AI models from its scope. Keep the development documentation, testing and change history in the folder so your validator can see them.
What documentation does it leave for the audit trail?
It leaves the formulation with its data assumptions spelled out, names the files the numbers came from and records what it could not find or reconcile instead of guessing. You still add the sensitivity runs, the V&V results and the credibility assessment your program requires when the study feeds an AoA or a model inventory submission.
Is study data kept confidential?
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.
Can it pick the winning alternative in a trade study?
No, the pick stays with you. It builds the decision tree, tests the value of information and moves the weights around until a single change flips the ranking. You take the ranking, the sensitivity results and the uncertainty estimate to the review board.