GxP audit trail review of every row, person by person — each period decided and signed on schedule
Every audit-trail period reviewed on schedule and signed, rows cited
See one case, screen by screen ↓Software that collects the audit trails of your GxP systems, screens them for data integrity risks and records who reviewed each period and what they decided. Audit Trail Review adds seven agents that collect every system's export on a risk-based schedule, map each event to a person, flag what needs a look with the exact log rows, and draft the deviation and the review record. A data integrity reviewer decides every finding and signs.
How to choose AI solutions for life sciences →
Updated
Inspectors expect the audit trails of critical data to be reviewed, and they ask for the evidence. But a chromatography data system, a LIMS, an MES, environmental monitoring and an EDC each export their own trail, in their own columns, and together they run to hundreds of thousands of rows a week. In the time allowed, QA samples a fraction by hand. A shared shift account or an emergency grant that was never removed hides who actually did what, so even the rows that are read are hard to attribute.
The risk sits in the rows nobody reached. An assay reintegrated three times late at night until a failing result reads as a pass, results approved by the person who entered them, a workstation clock set back before an alarm limit is raised — each can sit in an export until the batch is released. Reviews slip behind, periods with nothing found often have no signed record at all, and a system that had a bad week is reviewed no more often than before.
Estimated split of reviewer hours for one week across the sample company's nine systems, sampling by hand and with the solution.
Five agents collect, map and screen every audit trail into one lane per person; two more draft deviations and the review record. A data integrity reviewer decides every flag and signs.
Northwind Bio runs nine GxP systems across Leiden, Waterford, Raleigh and clinical study VLT-302. It is Wednesday, October 7, and week 40 (Sep 28 – Oct 4) is due for review. Here is what happens next, screen by screen, in the working solution.
Priya opens User swimlanes: 12 people and accounts across seven days, rostered shifts as bands, 20:00–06:00 shaded as out of hours. Every screened event is a tick; 2,219 events of interest, and 7 of 8 flags still open — “Reprocessed ×3”, “3 sessions at once”, “Self-approved ×14”, “Clock set back 2 h 10”, “Back-dated ×37”. CDS Waterford was collected at 06:10 and is waiting to be screened.
“Click a flag in a lane to see its hour, the rows behind it and the rule that raised it.”
Joris is a QC analyst at Leiden, rostered 07:00–15:30. His lane is flagged on Thursday night: a test injection, the sequence, processing at 94.1%, then three manual integrations, the third reported at 95.3%. Four checks raised it — repeated reprocessing, manual integration without a reason, trial injection before a sequence, and out-of-hours edits to critical data.
“Reprocessed 3 times in 21 min, then reported.”
F-112: the automatic integration gave an assay of 94.1%, below the 95.0–105.0% specification for Corventa 50 mg lot CV50-2609-14. Over the next 21 minutes the main-peak baseline was moved three times; the third version reported 95.3% and was signed three minutes later. Two of the three manual integrations have no reason, the “sst-check” test injection was excluded the next morning, and no OOS investigation is linked to the lot. Suggested: open a deviation, confidence 92.
“Rows are shown exactly as exported — the agents never change them.”
Priya chooses “Open a deviation” and notes a possible OOS not investigated. The deviation drafter writes DEV-2026-0612 from 9 cited rows: lot CV50-2609-14 on hold and not released, three other lots the analyst tested after 20:00 in September, an OOS investigation on the 94.1% result, manual integration restricted pending review, and a proposed class of Major. She sends it to Omar Haddad, who approves it, and it is raised in the QMS.
The deviation drafter never raises a deviation in the QMS on its own — QA approves first.
Environmental monitoring at Leiden had a High finding this week: the workstation clock was set back 2 h 10 min, then the particle alert limit for Grade B room 2.14 was raised from 2,000 to 3,500 per m³ (F-115). The schedule proposes moving it up one tier, citing the SOP section that says so.
“QA-DI-007 §5.5 moves a system up one tier for 8 weeks after a High finding: before each batch release, plus weekly, until Nov 30.”
Priya applies it and sends it to QA; Omar Haddad approves. Env. monitoring · Leiden is Critical until Nov 30, and its reviews over the next ten weeks rise from 10 to 26. The plan now holds 205 reviews across 5 critical systems.
Nine records are open. Environmental monitoring is due tomorrow; CDS Leiden (148,220 rows, 3 findings) and LIMS (186,410 rows, 2 findings) in two days; CDS Waterford is still at Screening with 121,906 rows collected. Since go-live in May, 136 records have been signed.
121,906 rows exported, checksum verified, no gaps in sequence numbers. 38 accounts resolved to people and server and workstation clocks compared. 12 rules raise 290 events of interest; the pattern finder compares people and their instruments with their 90-day baselines. Two findings, citing 5 log rows — six impurity peaks integrated by hand with the same copied reason, and an aborted run with a reason recorded.
The record preview fills as the agents work: scope, rule set v4, every finding and its decision.
On the CDS Leiden week 40 record, F-112 is a deviation QA approved. Priya accepts the suggestions for the medium and low findings — ask Anouk Smit why method v7 was used before it took effect, and justify the aborted run. She signs with the meaning “Reviewed — all findings decided”. Every rule states its result, and “no anomalies” is written down for the 8 rules and pattern checks that found none.
“The record is locked. Any change needs a new version with its own signature.”
The same week of audit trails, seen by the five people who carry it — what their review looked like, and what it looks like now.
Five agents collect, map and screen every audit trail into one lane per person; two more draft deviations and the review record. A data integrity reviewer decides every flag and signs.
Exports each system's audit trail for the period on its schedule, with every column, the user and role list and the shift roster; checks that row counts match the system's own count with no gaps in sequence numbers, and stores the export unchanged with its checksum.
Maps each system's columns to one event — who, what, object, old and new value, reason, workstation, server and workstation time — and resolves account ids to named people, roles and shifts, marking generic or unowned accounts.
Applies every active rule in the approved rule set — deletions, repeated reprocessing, manual integration without a reason, trial injections, method changes after a run, clock changes, shared accounts, self-approval, elevated rights — and counts what each rule checked and raised.
Builds a 90-day baseline for each person, instrument and site — late-night clusters, bulk back-dating, unusual rerun rates — and says which baseline, how far off, and what normal looks like.
Groups events on the same record, person and hour into one finding, tells what happened in time order, states the result before and after against its specification, and suggests Justified, Ask the person or Open a deviation, with the reason.
Turns a finding the reviewer sends to deviation into a QMS-ready draft: title, event date, system, description with citations, potentially affected lots and results, immediate actions and a proposed classification.
Assembles the record for each system and period from the review data: scope, export checksum, rule set version, counts per rule, every finding with its decision and decider, links to deviations, and “no anomalies found” where that is true.
Decides and signs; QA approves deviations. The agents propose; a named person decides.
Ask in plain words about a system, a finding, a person or a rule — or say what to change. Every answer points to its log rows.
What needs me before Friday?
Start with F-112: a Corventa assay moved from 94.1% to 95.3% by three manual reintegrations late on Oct 1. Environmental monitoring is due first, tomorrow — one finding, a clock change.
Any shared accounts?
One: qc_shift_b in the LIMS signed in from three workstations within 7 minutes on Sep 30 and entered 41 results. Three people were on shift B, so none of the 41 can be attributed. It was a label-printer account given result-entry rights in 2024. Suggested: open a deviation and disable the account.
Who changed a system clock?
On Oct 2 at 03:15 the environmental-monitoring workstation was set back 2 h 10 min, then the particle alert limit for Grade B room 2.14 was raised from 2,000 to 3,500. Separately, EDC site 104 entered 37 visit dates 3–6 weeks after the visits, against a usual lag of 3 days.
Add a rule: flag sequences rerun more than twice in 24 hours
New rule for both chromatography data systems. Over the last 90 days it finds 3 sequences: SEQ-24-1187 at Leiden (already F-112), and two at Waterford that were reruns after documented instrument faults. It sits in rule set v5 for Omar Haddad's approval, and is in the audit trail.
13 screens from the working solution, on its sample data. Pick one to see it large.
How often each tier is reviewed, what counts as out of hours and who approves what are settings — versioned, and on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Critical systems (CDS, LIMS, MES) | Weekly, plus release-linked trails before each release | Daily · Weekly · Every 2 weeks |
| High-risk systems (environmental monitoring, EDC) | Weekly | Daily · Weekly · Every 2 weeks · Monthly |
| Medium-risk systems (QMS, ELN) | Monthly | Weekly · Monthly · Quarterly |
| Out-of-hours window | 20:00 – 06:00, unless the person was rostered | 19:00 – 07:00 · 20:00 – 06:00 · 21:00 – 05:00 · 22:00 – 05:00 |
| Reviewer must be independent | On | On · Off |
| Second approval for High-severity findings | Omar Haddad | Omar Haddad · Priya Raman |
| Deviations raised in the QMS after approval by | Omar Haddad | Omar Haddad · Lena Ortiz |
| Sign records that have no findings | On — the record states “no anomalies found” | On · Off |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · Northwind Bio and the people named on this page are characters in the working solution. Context: 21 CFR 11.10(e) · FDA Data Integrity and Compliance With Drug CGMP, Q&A (2018) · MHRA GXP Data Integrity Guidance (2018) · PIC/S PI 041-1 (2021) · EU GMP Annex 11 draft revision (2025).
Software that collects the audit trails of your GxP systems, screens them for data integrity risks and records who reviewed each period and what they decided. Audit Trail Review adds seven agents that collect every system's export on a risk-based schedule, map each event to a person, flag what needs a look with the exact log rows, and draft the deviation and the review record. A data integrity reviewer decides every finding and signs.
Every row. In the sample company the agents screened 490,574 rows across 8 systems for one week. Events that are explained — a manual integration with a valid reason, for example — stay out of the findings but are counted in the record.
Repeated reprocessing, manual integration without a reason or with a copied one, trial injections before a sequence, aborted runs, processing methods changed after the run, and edits outside the rostered shift — alongside rules for every system such as deletions, clock changes, shared accounts, self-approval, elevated rights and audit-trail settings changed.
Each system is scored on impact, configurability and findings over 90 days, and the score sets its tier and how often its trail is reviewed, with the SOP section shown beside every date. After a High finding the schedule proposes moving the system up one tier for 8 weeks; it changes only when QA approves.
It resolves every account to a named person from the identity and roster feeds and marks accounts that have no single owner. In the sample, a LIMS shift account used on three workstations at once could not be attributed to any of the three people on shift, so it was raised as a finding with a deviation suggested.
No. The agents never change, filter or shorten an export, never decide a finding, never raise a deviation in the QMS and never change the schedule on their own. They suggest, with every statement cited to its row.
Yes. A data integrity reviewer decides every finding and signs every record with its meaning, name and time; reviewers cannot review periods in which they changed records; High findings, deviations, rule sets and schedule changes need QA approval. A signed record is locked.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your review SOP, risk assessment, system list and sample audit-trail exports — and it goes live once every quality gate has passed. We will walk you through it on your own exports first.
Testing sized to each function's risk, every line linked to its source.
1-min demoAI Use Case Risk ClassificationEvery new AI idea classified in minutes, each conclusion tied to its clause.
1-min demoVendor Security AssessmentA cited vendor risk rating with gaps and controls, ready to sign
1-min demoWe'll run Audit Trail Review on a sample of your own audit-trail exports.