Demo report, kept very short on purpose. It's run on a synthetic company with defects planted in it, and shows only the headline. A real report has much more: every check with its threshold and the reason for it, the measured value behind each fix, findings and recommendations for each dimension, what couldn't be measured, and the reviewer's notes.
AI Data Readiness
Halden Freight
Synthetic freight company: warehouse (raw and analytics schemas), dbt project and 1,610 documents
Go0
Go with limits1
Fix first1
No-go1
Use cases
Each verdict comes from the checks agreed for that use case before the run.
| Use case | Verdict | Checks passed | Time to go | Why |
|---|---|---|---|---|
| Support assistant over SOPs and customer contracts Answers support agents' questions from the SOP library and each customer's contract terms, with citations. | Fix first | 9/28 | 15 to 104 working days | Real questions have an answer in the documents: 75%, needs at least 80%; 14 more checks to fix. |
| Shipment ETA prediction Predicts delivery time at booking and updates it from carrier tracking events. | No-go | 14/26 | 10 to 74 working days, if the blockers can be closed | Months of raw.shipments.delivered_at to train on: 8.8 months, needs at least 24 months; 6 more checks to fix. |
| Questions over shipments and invoices Operations managers ask questions in plain English; the assistant writes SQL against the marts. | Go with limits | 28/38 | now | Answer delivery-time and on-time questions only on shipments booked from 2026-Q2 onward, and tell the person so. In analytics.fct_shipments, weight_kg: 0.3% below zero. Treat totals and groupings on those columns as approximate until they are cleaned. And 6 more terms. |
Terms of use
What each go with limits comes on, until its fix is done.
Questions over shipments and invoices
- Answer delivery-time and on-time questions only on shipments booked from 2026-Q2 onward, and tell the person so.
- In
analytics.fct_shipments,weight_kg: 0.3% below zero. Treat totals and groupings on those columns as approximate until they are cleaned. - In
analytics.dim_customers,country: 3 value(s) differ only by case or spacing;created_at: placeholder dates at the start of the calendar. Treat totals and groupings on those columns as approximate until they are cleaned. - And 5 more.
Use case by dimension
The worst result among each use case's checks on that dimension. A dash means the use case has no check there.
| Use case | Fitness | Meaning | Quality | Delivery | Ownership | Access | Privacy and rights | Documents |
|---|---|---|---|---|---|---|---|---|
| Support assistant over SOPs and customer contracts | fix | — | — | — | fix | fix | fix | fix |
| Shipment ETA prediction | blocker | limit | fix | fix | — | — | — | — |
| Questions over shipments and invoices | limit | limit | limit | pass | pass | pass | pass | — |
Fix list
The first 3, ranked by how many use cases each fix unblocks.
| # | Fix | Needed by | Owner | Effort |
|---|---|---|---|---|
| 1 | Months of `raw.shipments.delivered_at` to train on Recover delivery times for 2024-25 from carrier proof-of-delivery files, if the carriers keep them, so delivered_at and is_late cover the full history. If they don't, ETA prediction waits until Q1 2028 for two years of labels, and both column descriptions should state the February 2026 TMS migration. | Shipment ETA prediction (blocker) Questions over shipments and invoices (limit) | Ops data team | 10 to 30 days |
| 2 | Features can be read as they were at prediction time Start snapshotting the feature tables now, and build the training set from the snapshots. | Shipment ETA prediction (blocker) | to agree | 5 to 20 days |
| 3 | Joins from `raw.delivery_exceptions` don't repeat rows Deduplicate on event_id in staging and make the uniqueness test fail the run. That also stops the join from exceptions repeating rows. | Shipment ETA prediction (fix) | Ops data team | 1 to 2 days |
Scored against the Quabyt AI Data Readiness scorecard: 8 dimensions, each 0-5 or not assessed. A 3 on every dimension means the data can support named AI use cases with the fixes listed. Profiling ran as aggregate SQL in the client's environment; no rows left it. Every query that ran is in queries.sql. Use case verdicts come from the checks listed for each, agreed with the client before the assessment.