Cases/AI prototype/North American Stainless
Manufacturing · Steel2026AI prototype

AI that reads steel mill shift reports and finds the hidden losses

Shift leads at North American Stainless write what went wrong in free text. We built a working prototype that reads those reports, finds repeat failures and missed work orders, and hands each finding to a named owner with its source.

North American Stainless
2,428
delay minutes analysed in one month
7
AI capabilities designed
100%
read-only toward plant systems
  • Manufacturing · Steel
  • AI prototype
  • 2026
The challenge

The answers are in the reports. Nobody has time to read them.

Every line at the mill files Short Interval Control (SIC) reports each shift: delays, fixes, leaks, safety notes. Most of the useful signal sits in free text, and forms with yes/no boxes miss it.

Read more

Reading the Mill Bay record for July 2026, we found the same patterns again and again:

  1. 01

    Fixes that did not hold

    A work request was raised for a down-coiler side guide on July 11. The liner slipped again on July 19 and popped out on July 27, costing 76 minutes.

  2. 02

    Warnings without a work order

    Four tours in a row on one day logged the slab caster hoist: "10 TON NOT WORKING" through "fix the ten ton already". No work request number appears next to any of them.

  3. 03

    Leaks handled by topping off

    "Oil was 5+ inches deep creating a hazard walkin thru the cellar." The forms do not total top-offs per asset, so chronic leaks stay hidden.

  4. 04

    Severe events the boxes do not show

    A report described steel sprayed everywhere and catching fire, while the safety incident box said No.

Shift report entries next to the loss they hide
The noise and the hidden loss: real entries from the Mill Bay and slab caster reports.
What we built

Seven capabilities, each with an owner

Release 1 runs on SIC data alone. Release 2 adds read-only MES data. Every capability delivers a list, a draft or a flag to a named owner.

Read more
  1. 01

    Repeat-failure review

    Did the fault come back after a recorded fix? A weekly list for the maintenance planner.

  2. 02

    Fault-response memory

    What was tried last time on this fault? On-demand chat for shift leads and on-call techs.

  3. 03

    Top-off ledger

    Which assets are topped off again and again, a sign of hidden leaks.

  4. 04

    Severe-event flag

    Free text that describes a severe event, whatever the yes/no boxes say, goes straight to safety and the shift manager.

  5. 05

    Suspect-coil queue

    Coils that passed through the line during a fault, mapped for QA review.

  6. 06

    Event dossier

    A cited 30-day history of a major stop, assembled for the root-cause owner.

  7. 07

    Field value audit

    Which form fields carry no signal. One Mill Bay efficiency field was blank on 63 consecutive reports.

A finding delivered as a draft five-why chain with cited steps
Each finding arrives as a draft 5-why chain. Every step cites the report, and the engineer completes or rejects it.
How it works

The LLM reads the language, plain rules do the rest

The reports arrive as PDF files. The model does only what needs language: it resolves assets written many different ways, then classifies, extracts and summarises with a citation for each item. Time windows, thresholds and joins run as plain, auditable code.

Read more
  1. 01

    Read-only

    Nothing is written to the AS/400, MES or Level 2 systems.

  2. 02

    Cited

    Every item points to its report, date, shift and tour.

  3. 03

    Unknown is shown

    Silence in a report is never shown as a fact. No work request in the text means no work request evidenced, not none.

  4. 04

    People decide

    The platform does not decide the root cause. The engineer does.

  5. 05

    No ranking of people

    Findings are grouped by asset, area and question, never by crew or person.

It runs inside the client's own AWS account on Amazon Bedrock and AgentCore, and delivers email digests, chat, Power BI views and an on-prem API.

The outcome

A business case sized on the mill's own data

We sized the value from loss pools in the mill's own reports, as scenarios rather than forecasts: about $0.10M a year in a conservative case, $0.27M in the base case and $0.64M in the upside case.

Read more

The plan starts with a short setup phase that measures the return on real data and ends with a go/no-go decision, so the mill only pays for the next phase if the numbers hold.

Value scenarios from conservative to upside
Value scenarios sized on July 2026 Mill Bay and plate shop data. Scenarios, not forecasts.

AI where your people already write things down

If your plant logs shifts, inspections or tickets in free text, the signal is already there. We can prototype on your own records, read-only, and show what it finds before you commit to a build.

Your story next

Tell us the outcome.
We will write your case study

Discovery in two weeks. ROI proposal after. Public case study optional, by your choice.