Nestack Agent Care
Industries / Biotechnology / GMP batch records

Biotechnology AI agent · Manufacturing quality

GMP Batch-Record & Release Agent (Record to Release Package)

Review the executed batch record before release — entries, in-process results, yield, line clearance, deviations and the audit trail — and hand QA and your Qualified Person a release package with what is open.

4–6 weeksTypical delivery
Your stackDeployment
A named QPBatch release
Agent CareAfter launch

What this agent does

Checks the executed record against its master

In
01

Take the executed batch record as it stands, the master it was built from and the order its entries were made in.

02

Read the results, deviations, equipment and material status the other systems hold against that batch.

Reason
03

Check the entries, verifications and signatures the master requires, and mark what is blank, late or out of role.

04

Set in-process results against the limits in force, and yield and reconciliation against the expectation for the stage.

05

Read the audit trail behind the electronic entries: what changed after first entry, by whom, when and why.

Decide
06

Hold the package where a deviation, an investigation or a change control is open, and name who it waits on.

07

Raise what the record cannot support: a missing line clearance, an unstated equipment state, material without status.

Out
08

Present the release package: what was checked against what, and the open items with the person each sits with.

09

Keep the checks run, the trail entries read, the confidence and the reviewer and QA corrections.

Product statement

The agent checks, compares and assembles inside the approval boundaries agreed during implementation, and writes carry an attributable, reviewed trail. It signs nothing, completes no entry for an operator and closes no deviation — certification is a named Qualified Person's own act.

Example workflow

One executed record, end to end

AgentHuman
1Executed record receivedThe record for one batch, the master it was built from and the stages it covers, paper inserts included
2Entries and signatures checkedThe entries, verifications and signatures the master requires, inside the configured exception set, and the role behind each
3Results and yield comparedIn-process results against the limits in force, yield and material reconciliation against the expectation
4Status and trail readEquipment, line clearance and material status, deviations and their state, and the trail behind each entry
No human action required

Stages 1 to 4 run without a person in the loop — nobody is asked to read a batch record until the entries, the results, the status and the trail behind them have been set against the master. A deviation open on the batch leaves for the person who owns it the moment it is seen.

5DecisionSplits on what is open and the confidence gate
Record stands, nothing open

Goes to the reviewer as a release package.

Something is open or unsupported

Held with the open item and its owner named.

Batch-record reviewer, QA and the QP

Read the findings, the open items and the trail, then decide. A deviation is assessed and closed by a named person, and certification is the Qualified Person's own act — this list is not that assessment and does not stand in for it.

Accept · Correct · Escalate
Resolved — handed back
6Package assembledWritten only where access and policy allow, with an attributable trail; nothing is signed and no batch released
7Outcome evaluatedFindings QA dropped, discrepancies found first at review, open items missed and corrections by cohort
Corrections

Findings QA dropped or rewrote are counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required8 items
Releasing or certifying a batch.
Signing or approving a batch record.
Approving or closing a deviation or a CAPA.
Approving a change control or a specification.
Automation boundaryAgent acts unaided
Check entries, verifications and signatures against the master.
Compare results, yield and reconciliation with their limits.
Read the audit trail behind the electronic entries.
Assemble the release package and list what is still open on it.
Write actions run only inside the approval boundaries agreed during implementation. Certification is not one of them.
Deciding a batch conforms.
Deciding a deviation has no impact on the batch.
Editing or completing a GxP entry for someone.
Narrowing what the reviewer is shown.

Example output

One executed record, annotated

Everything the agent raises is attached to the batch and the entry it came from.

Release package · single batchIllustrative example
Batch
Record
Deviations
Proposed
Confidence
Batch status
Commercial batch, one line
Electronic, two paper inserts
One open
Hold for the open deviation
93%
Left for QA and the QP
As receivedThe executed record, the master it was built from and what the other systems hold against this batch.
Evidence used Master entry list Trail for that shift Deviation still open
Why it is heldA deviation open on this batch is unassessed, so nothing that follows from it is settled yet.
ActionAcceptCorrectEscalate
What the score decidesConfidence decides how much a reviewer re-reads, not whether a batch is released.

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value landsValue 01 – 04
Every batch record at reviewFrom execution to the release decision
03Review & checks

Read the record against its master

Use the master document, the limits in force, the equipment and material status and the trail behind each entry.

01Approved path

Start review at the findings

Entries, results, yield, status and open items arrive checked and assembled, so a reviewer works a list rather than a stack of pages.

02Human review

Show what is open before release

Open deviations, unsigned entries, unexplained discrepancies and changes made after first entry reach QA and the QP with owners.

04Build an evidence trail

Retain the checks run, the master version read, the trail entries examined, the confidence and every reviewer and QA correction — on both paths.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

MES & batch recordsWerum PAS-X · Rockwell PharmaSuite
Emerson Syncade · scanned paper records
ERP & materialsSAP S/4HANA · Oracle ERP
Material status · dispensing records
Quality systemsTrackWise · Veeva Quality
Deviations and CAPAs · change control

Agent

GMP batch record & release

Checks the record
Reads the trail
Lists what is open

Equipment & laboratoryCalibration logs · Maintenance records
Line clearance · released test results
Observability & evaluationOpenTelemetry · Langfuse
Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the batch record

Each control wraps the one inside it, and certification sits outside all six. Review by exception holds only where the system raising them is validated for your use and the trail is read with the record.

L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Rollback / safe modeRestrict automation if evaluations or production signals degrade; a record found wrong after a batch shipped goes to recall, field alert and notification.Roll back
L5Attributable trailReads, findings, edits and model, prompt and rule changes are recorded for a review that is defined, not optional.Record
L4Reviewer and QA gateDefine what may be written as a draft; certification, record approval and deviation closure wait for a named person.Gate
L3Exception scopeWhat the exception set covers is configured, versioned and shown with the package, so what it never raises reads as a gap, not as a clean record.Disclose
L2Limit and yield testsResults, yield and reconciliation are tested against the limits in force; a short reconciliation under a passing yield is measured here.Test
L1Master bindingEach check names the master step it ran against; unbound checks are held, and steps the master does not describe are listed as unchecked.Bind
Model coreFinding proposed — the entry, the result against its limit, the open item, the trail change and confidence
L1 – L2Decide whether a finding may stand
L3Keeps the reviewer's field of view honest
L4 – L5Leave certification to the QP, keep the trail reviewable
L6Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate what was checked, not only what was raised.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the release package the reviewer opens
Depth of coverage ▼
E1Final-output evaluationWas each finding real, and the master step it names the one in force?
E2Coverage evaluationWhat did the exception set never put in front of anyone?
E3Trail evaluationWere changes after first entry read, with who, when and why?
E4Routing evaluationDid an open deviation reach its named owner, and how fast?
E5Slice evaluationHow does performance change across specific batch cohorts?
E6Business outcomeHow much was re-read, and what was still open at review?
Floor — the batch decision QA and the QP stand behind

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle. The one this page is built around is a narrowed field of view — a record that reads complete because nothing raised an exception.

Agent lifecycleDirection of processing →
01 · Intake / record2 modes
GB-01

Blank entry reads as complete

A field never filled raises nothing, so a step has no record it was done.

GB-02

Paper insert never reaches the package

An attachment outside the electronic record is not reviewed at all.

Stage gathersThe executed record, its master and what other systems hold
02 · Checking2 modes
GB-03

Reconciliation short, yield in range

The percentage passes while the material count does not add up.

GB-04

Late entry passes as contemporaneous

A value written after the step carries its time, with no who, when or why.

Stage comparesEntries, results, yield and status against the master
03 · Exception1 mode
GB-05

Exception set narrows the review

Entries outside the configured set are never shown, so the record reads clean.

Stage raisesGaps, open items and changes made after first entry
04 · Package1 mode
GB-06

Deviation shows closed after reopening

A state read before it changed puts a live deviation in a clean package.

Stage returnsThe release package a reviewer, QA and the QP read
05 · Change / Version1 mode
GB-07

Rule change moves the exception line

A change to what is raised alters what reviewers see, unremarked.

Stage tracksMaster, limit, model, prompt and rule changes
Sev-1 · the reviewer is not shown what matters Sev-2 · the package misstates the batch Sev-3 · checks degrade, more is re-read by hand

Affected slices

Half-paper records are where findings go wrong

A finding QA drops is one the agent should not have raised, and the cohorts below produce them at very different rates. Paper inserts and contract batches are read into the check rather than read by it.

Slice performance — reported separately, not only in aggregateIllustrative example
SliceFailure rateLift Lift vs. thresholdStatus
Records finished on paper inserts5.3%3.5× Review
Contract-manufactured batches3.3%2.2× Review
Late and corrected entries2.9%1.9× Watch
Routine repeat commercial batches1.2%0.8× Normal
Bar: discarded-finding lift vs. routine-batch baseline · scale 0–4.0× · tick at the 2.0× threshold 2 of 4 slices over threshold

Evidence-linked improvement

The batch closes; the check that missed does not

Every dropped finding is a fault in the check, not in the batch. The correction is made where the check lives, and it clears your change control first.

Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect

Dropped findings, or discrepancies QA finds first, rise in one cohort.

02Diagnose

What QA saw and the check did not is the whole of the diagnosis.

03Improve

The change is used only once a named approver has signed it.

04Verify

Records that failed are checked again, those failures among them.

05Learn

The record that got past the check stays in the evaluation for good.

Learn → DetectThe return edge. The exception set is the part worth watching hardest: widening or narrowing what a reviewer is shown is a change to the review itself, and it is approved as one.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, master and limits, checks and trail, evaluation, integration, then production validation and handover.

Workstream Week 1Week 2Week 3Week 4Week 5Week 6
01Workflow discovery and boundary definition.
02Master batch record, limit and specification mapping.
03MES, ERP and quality-system access review.
04Entry, verification and signature checks.
05In-process limit and yield comparison.
06Reconciliation and material-status checks.
07Equipment, clearance and status reads.
08Audit-trail review behind the entries.
09Deviation, CAPA and change-control state checks.
10Exception-set definition and disclosure rules.
11Evaluation suite and regression records.
12Integration, deployment and Agent Care handover.
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tier PilotOne product, one line ProductionProduction MES integration AdvancedSeveral sites / systems
Introduced at Pilot
Entry, verification and signature checks
Limit, yield and reconciliation checks
Equipment, clearance and material status
Audit-trail and deviation-state review
Exception set disclosed with the package
Release left to QA and a named QP
Baseline evaluation
Introduced at Production
Production MES and ERP integration
Observability and evaluation
Introduced at Advanced
Several sites, products and lines
Multi-site and enterprise controls
Build price From $5,000 From $8,000 Custom quote
Final build priceConfirmed after discovery based on the sites and lines in scope, how much of the record is electronic, quality-system coverage, review controls and deployment requirements.
Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

You bringWe build with it
01Your master batch record and the limits in force Master batch record, limit and specification mappingWeek 1
02How an executed record reaches review today, paper included Workflow discovery and boundary definitionWeek 1
03Read access to the MES, ERP and quality systems MES, ERP and quality-system access reviewWeek 2
04Your rules for entries, verifications and second-person signatures Entry, verification and signature checksWeek 2
05Your deviation and change-control procedure as written Deviation, CAPA and change-control state checksWeek 4
06Batches that went wrong — a blank field, a late correction Evaluation suite, regression records and failure-mode testingWeek 4
07Named reviewers, a QA approver and your Qualified Person Review workflow, then pilot records and production validationWeeks 5–6
Nothing else is required Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases are drawn over the weeks they actually occupy. Week 3 is where the checks meet the trail behind them, and nothing later stands without both.

Phase W1W2W3W4W5W6
Discovery W1
Build W2 – W3
Evaluate W4 – W5
Pilot & Launch W5 – W6
Week focus W1Master records, limits and the automation boundary W2MES, ERP and quality-system access, and entry checks W3Yield, reconciliation, status and the audit trail behind them W4Deviation state, the exception set and batch cohort slices W5Integration, and first packages read by your reviewers W6Production validation with QA and your QP, then Agent Care handover
Reading the bandThe bars mark named work, not effort spread evenly. Week 4 is where the agent's blind spots are written down and agreed, and they are printed with the package from then on.
At the end of W6Nothing about who certifies has changed, and your reviewers can say whether a package saved them a pass. Agent Care then holds the limits, the exception set and the version map.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Biotechnology AI agent

Build a batch-record agent around the release decision you already make.

Show us one product, an executed record as it really is with its paper inserts, and how a batch reaches your Qualified Person today. What an agent does not raise is the part a reviewer cannot see, so we agree that list in week 4, version it, and print it on the package the agent hands your reviewers.

Nestack Agents · GMP batch record & releaseAGT-BT-09 · Agent Care available after launch