The system has three parts: a retrieval layer that indexes and queries your data, an agent runtime that plans a workflow and calls tools, and a model endpoint that does the language work. Ideally, the first two always run inside your environment. The third is the only component where a deployment decision has to be made.
Where each component runs
Component
Runs where
Holds your data
Retrieval index
Local, your cloud tenant or on-premise
Ideally only pass-through
Agent runtime
Local or container in your subscription or account
In memory during a session only
Tool connectors
Your network, behind your firewall
No, they pass through
Session and audit log
Your storage, your retention policy
Yes
Model endpoint
Your choice of the three patterns below
No, see below
Three deployment patterns
Fully in-tenant. The model endpoint is one you already own: Azure OpenAI in your subscription, Amazon Bedrock in your account, or Google Vertex in your project. No data crosses your network boundary at any point in a session.
In-tenant runtime, vendor model API. Retrieval and orchestration stay in your environment; prompt content goes to a model provider. Faster to stand up and tends to be user-focused. Requires your security team to accept a named third-party processor, so it needs a real decision rather than a default.
Self-hosted open-weight model. For environments where no third-party inference is acceptable at all. Runs entirely on your infrastructure. Capability is lower than a frontier model.
Who is on my side of the engagement
One person. I have no subcontractors, no offshore development team and no bench. There is no second party for you to vet, no staff rotation, and no one else who ever holds your credentials or your data.
02
Data handling
The agent is read-only by default. It has no write path into any validated system. Where write capability is ever discussed, it is a separate, explicitly scoped decision with its own risk assessment.
System
Read
Write
How
LIMS
Yes
No
Read-only or existing reporting views
MES
Yes
No
Read-only or historian query
QMS
Yes
No
Read-only or export
ERP
Yes
No
Read-only view on the relevant tables
Document stores
Yes
No
Read-only against SharePoint, Veeva or file shares
Agent output store
Yes
Yes
Drafts and analysis packets only, outside every validated system
What the output actually is
The agent produces drafts and analysis: an investigation packet, a release readiness summary, a stability projection, a draft Module 3 section. These are working documents. They are not GxP records, they are not the system of record, and they carry no signature. They become records only when a qualified human reviews the content and enters it into the validated system under that system's existing controls.
Retention and destruction
Session logs and the retrieval index live in your storage under your retention policy. At the end of an engagement, any working copy on my side is destroyed. During an engagement I prefer to work against a validation or UAT instance rather than production wherever the data supports it.
03
21 CFR Part 11 and ALCOA+
The Part 11 position rests on one distinction: the agent does not create, modify, sign or delete electronic records in a validated system. Your LIMS, MES, QMS and ERP remain the record of authority and remain governed by the Part 11 controls they already have. The agent reads from them and produces analysis for a human. Part 11 applies to the systems holding the records; the controls below describe how the agent layer preserves record integrity rather than undermining it.
Attribute
How it is preserved
Attributable
The session log records who asked, what they asked, and what came back. Output entered into a validated system carries the reviewing human's identity and signature from that system.
Legible
Output is plain text and structured documents readable without proprietary tooling. Session logs are readable JSON.
Contemporaneous
Sessions are timestamped at execution.
Original
The agent never mutates a source record. Every retrieved value is carried with its source system, record identifier and retrieval timestamp, so any figure in an output can be traced to the exact record it came from at the moment it was read.
Accurate
Numerical values are returned by deterministic queries and calculations, not generated by the language model. Narrative claims are grounded in retrieved records and cited.
Complete
The full session is logged: the request, every record identifier retrieved, every tool called, the model and version used, and the output returned. Nothing is summarized away.
Consistent
Logs are written in sequence with timestamps in a single timezone reference, retained in the order of execution.
Enduring
Logs are written to your storage under your retention schedule.
Available
Logs are retrievable by your team without going through me, in a format an inspector can read.
Where the human approval step sits
Between the agent's output and any GxP consequence. Always. The agent assembles evidence and proposes; a named, qualified human reviews the evidence, accepts or rejects the proposal, and applies their own electronic signature inside the validated system under that system's existing Part 11 controls. The agent holds no credentials in a validated system, has no signature manifestation, and cannot execute an approval. If the human does nothing, nothing happens.
04
Validation approach
Validation is risk-based and driven by intended use, following GAMP 5 second edition and ICH Q9. The intended use statement does most of the work here: the system is decision support, its output is reviewed by a qualified human before it has any GxP effect, and it has no write path into a validated system. That places it materially lower on the risk scale than a system making autonomous GxP decisions.
GAMP category positioning
Infrastructure (cloud tenant, container platform, storage) is Category 1 and leans on your existing qualified infrastructure. I do not re-qualify your platform.
Retrieval and orchestration components are Category 4, configured products. They are commercially available or open-source components used as intended and configured for your environment.
Workflow-specific configuration (the tool definitions, the queries, the prompts, the output templates for your particular workflow) is Category 5, bespoke. This is where the testing effort concentrates, and it is the honest place to put it.
Testing something that is not deterministic
A language model does not return byte-identical output on every run, and you cannot test it the way you test a calculation in a spreadsheet. The approach splits the system in two:
Everything deterministic is tested deterministically. Retrieval, queries, calculations and data transformations return the same result every time and are tested with fixed inputs and expected outputs, exactly like any other computerized system. This covers every number that appears in an output.
The generative layer is tested against acceptance criteria over a fixed challenge set. A defined set of representative cases with known correct answers is executed against the system, and each output is assessed against criteria that are objectively checkable: is every factual claim traceable to a retrieved record, is any claim present that the retrieved data does not support, is the correct conclusion reached, is the required content present. The pass threshold is agreed up front and the challenge set is version-controlled as a validation artifact.
05
What I don't do
These are scope boundaries, not current limitations waiting to be lifted in a later phase. If an engagement would require crossing one of them, that is a conversation with your quality organization before it is a conversation with me.
No autonomous changes to validated systems. Ever, in any deployment pattern.
No writes to the QMS. No deviations raised, no CAPAs closed, no change controls modified.
No writes to LIMS, MES or ERP. No results entered, no batch records altered, no master data changed.
No electronic signatures. The agent has no signature manifestation and no credentials to apply one.
No batch disposition. Release remains a human decision made by a qualified person.
No decisions without a named human approver. If no one is accountable for the output, the workflow is not in scope.
No training on your data. Not for my models, not for a provider's, not for a future client's.
No shadow system of record. Your validated systems stay authoritative and the agent's output store never competes with them.
06
Access required
This is the minimum footprint for a Data Diagnostic, which is where nearly every engagement starts. It is deliberately small, because a diagnostic is an assessment rather than a build, and because a smaller ask is easier for you to approve.
What I need
Why
Read-only access or extracts
For the systems named in the scope, limited to the data domains relevant to the named initiative. A validation or UAT instance is preferred over production.
Schema and data dictionary
To assess structure, completeness and joinability without guessing at your field semantics.
Representative records
A real batch, a real deviation, a real stability study. Volume is not the point; representativeness is.
One hour with a system owner
Typically two thirty-minute sessions per system. What the fields actually mean in practice is never in the documentation.
A named IT and QA contact
So that questions about entitlements and controls get answered by the right person rather than inferred.
CDA, and an MSA before any build
Signed before any access is granted.
What I do not need
Write access of any kind, to anything.
Administrator or superuser credentials.
Patient data. CMC and manufacturing data is the scope. If PHI or PII is present in a system in scope, it is excluded and I will say so in writing before access is granted.
Access to systems outside the agreed scope list, which is fixed in the statement of work.
A VPN account that outlives the engagement. Access is time-boxed and revoked at close.
Take this into the meeting
Prints to a clean letter-size document with contact details and no site navigation. Choose "Save as PDF" as the destination.
Questions this page doesn't answer are worth sending directly. kyle@kylelangham.com
Kyle Langham Consulting
Kyle Langham · kyle@kylelangham.com · (240) 848-2562 · kylelangham.com · linkedin.com/in/kylelangham
Governance, security and validation. Current as of the date printed.