01 the business
Ready-meal producer
220 staff, two production lines
Goods-in, in-process and food-safety checks happen on paper and in spreadsheets. Failed checks get noted but not always followed up. When a retailer or auditor asks for records, someone spends a day digging.
02 before
The problem
Inspections happen on paper or in a mix of spreadsheets and phone photos. Failed checks get noted but not followed up, and nobody sees that the same supplier or machine keeps causing trouble. When a customer or auditor asks for records, someone spends a day digging.
- Records are scattered across paper and files.
- Failed checks are not tracked to closure.
- Repeat problems by supplier are invisible.
03 the build
What the AI does
- Step 1: Gives inspectors a phone checklist for goods-in, in-process, final or food-safety checks.
- Step 2: Reads photos and measurements and compares them with the specification for that product or area.
- Step 3: Raises a non-conformance with the batch, supplier and photos when a check fails.
- Step 4: Routes the action to the right owner in Microsoft Teams and follows up until it is closed.
- Step 5: Shows repeat failures by supplier, line or site in a weekly report in Power BI.
paused: waiting for a person to approve
How it was put together
- Gave inspectors a phone checklist for each type of check.
- Compared photos and measurements with the product specification.
- Raised a non-conformance with batch, supplier and photos when a check failed.
- Routed actions to owners and reported repeat failures weekly in Power BI.
04 human in the loop
Where a person stays in charge
The quality lead decides whether failed stock is held, reworked or released, and signs off every closed non-conformance.
- The quality lead decides whether stock is held, reworked or released.
- Every closed non-conformance needs the quality lead's sign-off.
- Records cannot be edited after sign-off; changes are added as new entries.
05 rollout
From kickoff to live in 6 weeks
- Weeks 1–2Turn paper checklists into digital ones and agree specifications.
- Weeks 3–4Pilot on one line.
- Weeks 5–6Roll out to both lines and start weekly reporting.
06 after
What changes
Inspection records are complete and easy to find. Recurring problems show up as patterns instead of being fixed one at a time.
Assume four audits or retailer requests a year at 8 hours each to gather records, falling to 1 hour.
4 × 8 hours = 32 hours a year today against 4 hours, before counting the cost of failures that were never followed up.
What to measure
- Time to retrieve records for an audit.
- Non-conformances closed on time.
- Repeat failures by supplier.
- Customer complaints about quality.
07 evidence
What the research says
Gartner research put the average cost of poor data quality at $12.9 million a year per organisation.
Gartner, Data quality: why it matters and how to achieve it (opens in a new tab) · 2020
NIST's framework organises AI risk work into four functions, Govern, Map, Measure and Manage, with a profile for generative AI added in July 2024.
NIST, AI Risk Management Framework 1.0 and Generative AI Profile (opens in a new tab) · 2023
08 lessons
Mistakes to avoid
- Digitising a bad checklist. Fix the checks first.
- Letting the system release stock. That is the quality lead's call.
- Collecting data nobody reviews. Hold the weekly review.
09 questions
Common questions
- Does it work with SafetyCulture?
- Yes. It can use SafetyCulture for checklists and add the follow-up and reporting.
- Can it judge photos?
- It compares photos with reference images and flags differences for the inspector to confirm.