Causal AI for operations
Use causal discovery to find what really drives delays, defects, downtime or cost in operations: a worked delivery example, a table template, and a step-by-step routine from data to a tested change.
The question operations teams ask
A key number moves (late deliveries, defect rate, downtime, cost per order) and a dashboard shows ten things that moved with it. Which of them would you change? Causal discovery separates the levers from the bystanders.
A worked example: late deliveries
In this map, complaints follow late deliveries, and late deliveries follow load per driver and weather. Weather cannot be changed; order volume is the business itself. The lever is staff on shift: more staff on peak days lowers the load per driver, which lowers late deliveries and then complaints. A correlation table would also have shown complaints moving with order volume and overtime, inviting the wrong fixes.
Typical questions
- Delivery and logistics: what drives late deliveries: staffing, routing, warehouse backlog, weather?
- Manufacturing quality: which machine settings, materials or shifts drive defects?
- Downtime and maintenance: which conditions lead to breakdowns, and which only appear alongside them?
- Service operations: what drives handling time, backlog and repeat contacts?
- Cost: which inputs actually move cost per unit?
How to lay out your table
One row per unit of time or work (a day per site, a shift, a batch, an order), one column per measure:
| date | site | orders | staff_on_shift | load_per_driver | rain_mm | late_pct | complaints |
|---|---|---|---|---|---|---|---|
| 2026-03-02 | 1 | 1840 | 42 | 43.8 | 0 | 4.1 | 12 |
| 2026-03-03 | 1 | 2210 | 38 | 58.2 | 6 | 9.7 | 31 |
- Keep the columns numeric; code categories (site, shift) as numbers or leave them out of the first run.
- Leave identifiers and dates out of the discovery itself (UCMT Labs' data check flags ID-like columns).
- Include the likely common causes (volume, season, weather): left out, they create false links.
- Aim for hundreds of rows or more; UCMT Labs warns below 500.
From data to a tested change
- Write the question: which number, and which levers you could change.
- Build the table and upload it; read the data check.
- Run the suggested method and one from another family; keep the links both find.
- Trace the paths into your key number on the Reasoning tab.
- Ask what-if questions on the graph: what if staffing rose 10 % on peak days?
- Pilot the change in one site or for a few weeks, and compare with the rest.
Limits
The graph is built from the past. A change outside what the data has seen (a new site, a new process) may not behave as the past did. Treat what-if results as estimates, and confirm the important ones with a pilot.