AI Agents for Water Testing Labs: Sample Custody and Results
An agent can run sample intake, chain-of-custody logs and result notifications for a water lab while analysts keep sign-off on every reported value.

An AI agent for a water testing lab can receive sample requests, keep the chain-of-custody log complete at each handoff, and send status updates to clients and staff. It works inside permissions the lab defines, so it can move records and draft messages but cannot release a reported analytical value. Analysts stay responsible for sign-off, and the agent's job is to make custody and status visible so that no result reaches a report without a named person checking it.
Key takeaways
- Sample intake, custody logging and result notifications are strong first tasks because they are record-heavy, repetitive and easy to verify.
- The agent should write to the custody log only through defined events such as receipt, storage, transfer and disposal, each tied to a named person or device.
- Reported values stay behind an analyst approval step. The agent may prepare a notification, but it never releases a signed result on its own.
- Every agent action should leave a trace that a quality reviewer can read, including what the agent saw, what it changed and who approved the next step.
- Start with one workflow, confirm the exceptions are handled, and expand scope only after a permission review.
Where an agent fits in a water lab's sample flow
Water testing work follows a sequence. A client or field technician requests analysis, a sample arrives, it is logged and labeled, it goes into storage, it moves to bench analysis, the results are reviewed, and a report is issued. Most delays and most disputes happen at the handoffs between these steps, not at the instrument. An agent is useful at those handoffs because it can watch the queue, ask for missing fields, and keep records current while people are busy with the bench.
The useful split is straightforward. The agent handles coordination and documentation: intake forms, label checks against the request, custody events, reminders about hold times, status questions and notifications. Analysts handle chemistry, calibration status, quality control decisions, dilutions, and any judgment about whether a sample or a value is credible. If a task requires interpreting water quality, it belongs to a person.
How do you design a chain-of-custody log an agent can maintain?
A chain-of-custody log is only as good as its events. Define the events before connecting any tool, and keep the list short enough that staff can follow it on a busy morning. A practical starting set looks like this:
- Sample collected, with the collector's name and site identifier recorded on the request
- Received at the lab, with container condition, seal status and any temperature indicator reading noted
- Labeled, with the label matched to the request before the sample moves
- Placed in storage, with the storage location named
- Removed for analysis, with the analyst's name
- Transferred between staff, subcontracted or returned to the client
- Disposed of or retained, with the retention decision recorded
Each event needs a timestamp, an identity for the person or device that performed it, and a reason when the event is unusual. The agent can prompt for missing fields and refuse to close a record with an empty custody step. Exceptions, however, belong to people. A broken seal, a sample received outside its hold window, or a label that does not match the request should be surfaced by the agent, sent to the responsible analyst, and kept open until someone resolves it. Labs that already track custody for other physical goods will recognize the pattern. The same reasoning appears in AI for Jewelry Repair Shops: Custody Logs and Order Status, where each handoff needs an owner.
Result notifications without releasing values
Clients ask two kinds of questions. The first is about status: when will the results be ready, and has the sample arrived? The agent can answer that from the custody and workflow record. The second is about values: what did the sample show? That question should only be answered from an approved report. Keep the agent away from the number until an analyst has signed the result, and write the notification rules so that this boundary is explicit.
A simple decision guide keeps the behavior consistent:
- If the question asks for status (received, in analysis, awaiting review), the agent may answer from the custody and workflow record.
- If the question asks for a value and the report is signed, the agent may send the approved report through the lab's standard channel.
- If the value is not yet signed, the agent says that review is pending and gives the expected next step, without interpreting or hinting at the result.
- If the client raises a concern, such as an unexpected result or a request for re-testing, the agent routes the message to the responsible analyst and records the request.
- If a message mentions a regulatory deadline, a legal dispute or a health concern, the agent does not reply on substance and escalates at once.
Notification templates should be reviewed by a person until the lab has enough history to trust them. Even then, any message that mentions a result should keep an explicit approval step. Status messages can be automated more freely, because they carry less risk if a timing estimate is off by a day.
Illustrative scenario: a sample with a missing collection time
The following scenario is illustrative and does not describe a real client. A field technician uploads a sample request for a private well and leaves the collection time blank. The agent checks the request against the intake fields, finds the gap, and asks the technician to confirm the time before the sample is logged as received. The sample is labeled and placed in storage with its custody events recorded in order.
Later, the client emails to ask whether the well is safe to use. The agent cannot answer that question. It replies that the report is in review, adds the request to the record, and forwards the message to the analyst responsible for the sample with the custody history attached. The analyst reviews the data, signs the report, and the agent sends the approved copy through the usual channel. The agent made the path faster to follow, but the judgment and the signature stayed with the lab.
Permissions, audit trails and failure handling
Treat the agent like a staff member with a narrow, written role. It should have read access to requests and custody events, write access only to the event types it owns, and no ability to edit analytical data or an approved result. These limits must be enforced by the system, not left to an instruction in a prompt. The OWASP authorization guidance makes this point clearly: access decisions belong in controls outside the model's text, checked on every request.
Client emails and field notes are untrusted text, and a message can contain instructions aimed at the agent. A line asking the agent to ignore the hold rule and send the value should have no effect, because the agent has no permission to send signed values outside the approved path. The OWASP prompt injection prevention cheat sheet is a useful reference for designing those boundaries and for testing them before go-live.
Log the agent's actions with the same care you would apply to an analyst's work. Each action should record the input it saw, the record it changed and the time it acted. The OpenTelemetry observability primer offers useful vocabulary here: traces show the sequence of steps, and logs describe what happened at each step. The NIST AI Risk Management Framework gives a broader structure for identifying and measuring these risks over time. A quality review is much easier when the lab can reconstruct a path without asking the agent to explain itself. Labs that already run nonconformance tracking will find this familiar, as described in AI for Quality Management: Nonconformances and Audit Prep.
Define what happens when things fail. If the agent cannot reach the lab's system, it should stop writing and queue its actions for a person, not guess at missing data. If it is unsure about a field, it should ask. Review its permissions on a regular schedule, and again whenever the lab changes methods, staff assignments or report templates.
How to start with one workflow
The safest path is to start small. Choose one workflow with clear events and a visible backlog, usually sample intake or status questions. Then follow these steps:
- Write the custody events and the exception list before connecting any system.
- Assign an analyst owner to each event type and to each category of exception.
- Set the agent's permissions to the events it owns, and exclude anything that releases or edits a value.
- Run the workflow on a limited set of incoming samples while staff compare the agent's log entries with their own records.
- Review open exceptions regularly, correct templates where the same question keeps coming back, and decide whether the scope should grow.
Be Human designs and operates this kind of setup as a managed service. A first engagement defines the one role the agent plays in the lab, the memory and tools it needs, the approval boundaries that keep analysts in charge of reported values, and an observable success criterion, such as fewer custody exceptions left open at the end of each working day. Managed operations then cover monitoring, recurring evaluations, permission review, integration maintenance, incident response and controlled changes to the workflow. The work is remote and available in English and Portuguese.
If you want to find the first task in your lab worth handing to an agent, the free Be Human AI diagnostic is a practical place to begin.
Frequently asked questions
Can an AI agent sign off on water test results? No, and it should not be configured to. Sign-off belongs to a qualified analyst who checks calibration, quality control and the method conditions for each sample. The agent can prepare drafts, track which results are awaiting review and send a signed report through the lab's channel once a person has approved it.
What does an AI agent need access to in a water testing lab? It needs read access to sample requests, custody events and workflow status, plus write access only to the event types it owns, such as logging receipt or recording a client message. It should not be able to edit analytical data or signed reports. Access should be limited by role and reviewed on a schedule, not granted once and forgotten.
How is chain of custody different from sample tracking? Sample tracking tells you where a sample is. Chain of custody shows who handled it, when, under what conditions, and what happened at each transfer. An agent can support both, but custody requires a stricter record in which each event is named, timestamped and linked to a person or device, and each exception is resolved by someone accountable.
Where should a lab start if its records are spread across several tools? Start with the workflow that creates the most repeated questions, usually status requests or intake corrections, and map the events it depends on. Clean up the event definitions before connecting systems, because unclear events make both the agent and the audit trail unreliable. A narrow first role is easier to permission, audit and roll back than a broad one that touches everything at once.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


