How to Keep Audit-Ready Records of Factory Decisions, Including AI
DBR77 Team7 min read
On this page
An audit-ready decision record shows what was decided, who decided it, when, on what data and why, and it cannot be changed without leaving a trace. For AI-assisted decisions, the record also keeps what the AI recommended, what the person did with it, and the reason for any override. If you can rebuild any decision from last month in a few minutes, from one system, your records are audit-ready.
Most plants record the result of a decision, not the decision itself. The lot was released or the check skipped, but the who and why live in an email or someone's head.
What "audit-ready" means for a decision
Regulated industries have a well-tested definition. The US FDA's data integrity guidance for drug manufacturing says data should be attributable, legible, contemporaneously recorded, original or a true copy, and accurate, known as ALCOA. The same guidance defines an audit trail as a secure, computer-generated, time-stamped record that allows the course of events around the creation, modification or deletion of an electronic record to be reconstructed.
US rule 21 CFR Part 11 adds the system side. It requires secure, computer-generated, time-stamped audit trails that independently record operator entries and actions that create, modify or delete records, and changes must not obscure earlier entries. It also requires access limited to authorized individuals and authority checks on who may sign or change records.
These rules apply to FDA-regulated records. Outside that scope, they are still a sound checklist for customer and certification audits.
A lot of 4,000 molded housings is on hold after cavity 3 drifted.
What a decision record must contain
Most of these fields can be filled automatically if the decision is made inside the system where the data lives.
| Field | Example | Why an auditor asks |
|---|---|---|
| What was decided | Lot 2291 released for shipment | The decision itself, not only its effect |
| Who decided | Name and role, signed in | Attributable, and the role had the authority |
| When | Date and time, set by the system | Contemporaneous, not typed in later |
| On what data | Inspection results, readings, linked order and lot | The decision can be checked against the facts |
| Why | Reason text, standard or procedure applied | Especially for exceptions and overrides |
| AI involvement | Recommendation, confidence, accepted or changed or rejected | Shows where a person used judgment |
| Result | Verification step and its outcome | Proves the decision was followed up |
Records for AI-assisted decisions
When AI ranks, suggests or drafts, the auditor's question changes from "who decided?" to "who decided, and what did the system tell them?" Four rules keep that answerable:
- Store the recommendation as it was shown, with its inputs, its explanation and its confidence, not a later version.
- Store the human action: approved, changed or rejected, by whom, and when.
- Require a reason for overrides and rejections. A rejection with a reason is useful evidence, and a silent one is a gap.
- Never overwrite. A changed decision is a new entry linked to the old one.
The EU AI Act sets a similar bar for systems it classes as high-risk: they must technically allow for the automatic recording of events (logs) over the lifetime of the system, to support traceability. Most shop-floor recommendations will not fall into that class, but the logging habit is the same.
Illustrative example: a lot of 4,000 molded housings is on quality hold after a dimension drifted on cavity 3. The AI recommends releasing cavities 1, 2 and 4 and scrapping cavity 3, based on 120 re-measured parts. The quality manager changes it: release cavities 1 and 2, re-inspect cavity 4, scrap cavity 3. The record keeps the AI's recommendation and its data, her change, her reason ("cavity 4 within limits but trending"), the re-inspection result two hours later and the final release. Six months later, a customer asks about that lot. The answer takes three minutes.
Who owns which data
Records are only as trustworthy as the lists behind them. Give every type of data one owner by role:
- Master data (assets, materials, inspection plans, roles and rights): one owner each, usually in engineering, quality or IT.
- Operational records (orders, inspections, work orders, holds): owned by the role that closes them.
- Rules and thresholds (alert limits, approval gates, sampling rules): owned by the process owner, changed only with a logged reason.
- AI configuration (which recommendations need approval, by which role): owned jointly by operations and quality.
Check access at least once a year. People change jobs, and old rights to approve or edit are a common audit finding. For who answers for open issues at a shift change, see How to Keep Accountability Clear Across Shifts.
A 15-minute audit test
Pick three decisions from last month: one quality release, one skipped or postponed check, one closed work order. For each, try to answer from records alone:
- Who made it, and did their role allow it?
- What data did they have at that moment?
- Why did they decide that way?
- Was any part suggested by a system, and did they change it?
- Can you see every later change to the record?
If any answer needs a phone call, that decision is not audit-ready. Note which field was missing and fix both the record and the process that left it empty.
How this works in DBR77 IRIS
Decisions in DBR77 IRIS are made in the same system that holds the data. IRIS QMS links each non-conformance to its production order, machine and material lot, and documents each step of 8D and 5-Why work from containment to verification. How to Link Quality Issues to Production and Maintenance explains how those links work. IRIS QMS also supports internal and external audits with tracked findings and corrective actions. IRIS CMMS keeps a full audit trail of work orders with technician notes, time logs and parts consumed. Role-based access sets what each role sees.
For AI-assisted decisions, Data & AI keeps each recommendation with the basis for its score, the action a person took on it and, for applied ones, the outcome. Decision Traces show how a decision was reached, and the Compliance screen keeps a register of AI uses and high-risk decisions. The IRIS assistant only reads data, so every change to a record is made by a named person on the module screen. When AI Should Recommend and When People Should Decide covers which decisions to leave to people in the first place.
FAQ
Do we need to meet 21 CFR Part 11?
Only for records under FDA rules. Outside that scope, Part 11 and ALCOA are still a good checklist for any customer or certification audit.
Should we keep rejected AI suggestions?
Yes. Rejections with a reason show that people reviewed the output, and they help improve the rules and the model.
How long should decision records be kept?
As long as your regulations and customer contracts require. Part 11 asks that audit trail documentation be kept at least as long as the records it covers.
Can AI write the record for us?
It can draft summaries and fill fields from data. A person confirms the decision and the reason.
Conclusion
Run the 15-minute test this week on three decisions from last month: a quality release, a skipped or postponed check and a closed work order. Wherever an answer needed a phone call, fix that field in the process. For AI-assisted decisions, also check that the record keeps the recommendation as shown, the human action and the reason for every override.