Skip to content
be-human.ai
← Back to the journal

AI by industry

AI for Geotechnical Drilling Firms: Boring Logs and Sample Tracking

A practical guide for drilling firms on turning boring logs and lab sample status into a searchable record an AI agent can cite, with clear approval limits.

Be Human typographic cover with the title: AI for Geotechnical Drilling Firms: Boring Logs and Sample Tracking
Conceptual illustration created with AI · Be Human

An AI agent for a geotechnical drilling company is most useful when it can answer a simple question with a source: which borehole produced this sample, at what depth, who signed the log, and where the laboratory result stands today. That only works if boring logs, field notes and lab status live in one record with stable identifiers, not in separate PDFs, spreadsheets and inboxes. This article explains how a drilling firm can build that record first, so an agent can search it and cite it without guessing.

Key takeaways

  • A boring log is the primary record of what the ground was, so every lab sample should point back to a borehole, a depth interval and a specific log revision.
  • Lab status belongs next to the sample it describes, with a single owner and a timestamp for each step: submitted, received, tested, reported.
  • An agent can only answer reliably from records it can search and that carry identifiers, revisions and authorship.
  • Start with one role, such as sample status answers for project managers, and keep the agent out of log edits and interpretations that engineers own.
  • Success should be observable: every answer cites its records, and a fixed sample of answers is reviewed against the source documents.

Why boring logs and lab status drift apart

On a typical site investigation, the field crew records logs as the boreholes are advanced, a technician or engineer transcribes them into the project files, samples leave in coolers with a chain of custody form, and the laboratory reports results days or weeks later against sample numbers. Each handoff is competent on its own. The weak point is the links between them, which are usually implicit: a label written on a jar, a line in a spreadsheet, a PDF named with the project code and a revision letter that nobody remembers how to read.

When a project manager asks whether the samples from the northern boreholes have been tested, the honest answer is often 'probably, let me check', which means someone opens several folders and compares names by eye. Errors rarely come from carelessness. They come from the fact that two correct records can describe the same interval with different labels.

An agent does not fix this by being clever. It fixes it only when there is a record it can read. The useful test for a drilling firm is not whether AI can read its PDFs, but whether a question arriving early in the morning can be answered with the borehole, the depth interval, the sample, the lab status and the log revision it depends on, with each item traceable to its source.

What the record needs to contain

A usable record links four things with stable identifiers: the borehole, the logged intervals, the sample with its custody history, and the laboratory status with the report it produced. Each entry also needs an author and a time, and every change needs a revision rather than an overwrite.

  • Project and borehole identifiers, with location and drilling method, fixed once and never reused for another hole.
  • Logged intervals with top and base depth, the soil or rock description in the classification language your engineers use, groundwater observations and field test results recorded at that depth.
  • Sample records that carry the identifier written on the container, the borehole, the top and base depth, the sample type, the recovery date and the person who sealed it.
  • Custody events for each handover between field, courier and laboratory, each with a signature or equivalent confirmation.
  • Laboratory status with a timestamp for each state and the report reference and revision that the laboratory issued.
  • Log revisions. When an engineer changes a description after review, the earlier version stays on file and the change names its author.

The same discipline of stable identifiers and revision history appears in other technical record sets, such as the RFI and submittal logs for civil engineers that our team has worked through in a separate article.

A step-by-step plan to build the sample chain

Start with one active project rather than the whole archive. Choose a project with at least one completed borehole and samples still waiting on laboratory results, because that gives you a live test without historical clean-up. Then work through the following steps in order.

  • Write down how the field crew currently names boreholes, samples and log revisions, so the mapping to permanent identifiers is explicit instead of being rewritten by accident.
  • Assign each borehole, sample and log revision a permanent identifier, and keep the original label visible as an alias.
  • Link the original logs, custody forms and laboratory reports to the record rather than replacing them, so the source documents remain the evidence.
  • Pull laboratory status from the status messages or report index the laboratory already provides, and match each entry to a sample identifier. Flag any entry that does not match, instead of guessing.
  • Define the answer format before the agent exists. Each answer should name the borehole, the depth interval, the sample identifier, the lab status with its timestamp, and the record it came from.
  • Test with about twenty real questions gathered from the project manager and the engineer. Record which answers were correct, which cited an outdated revision, and which should have gone to a person.

This order matters. The firm learns whether its identifiers are consistent before any agent depends on them, and the test questions become the first evaluation set.

Scenario: a result that arrives after the log changed (illustrative)

This scenario is illustrative and does not describe a specific client. A project manager asks whether the test on the sample from borehole BH-07 at a particular depth interval has been reported. The laboratory reported last week against the sample identifier, but the engineer revised the description for that interval after review, and a design memo still quotes the earlier description.

Without a shared record, the answer depends on who remembers the revision. With the structure described above, the agent returns the lab status with its report reference and timestamp, then notes that the log has a newer revision than the one cited in the memo. It does not decide which description is correct. It routes the discrepancy to the engineer who owns the log, with both revisions attached. The project manager gets a clear answer on the lab status, and the engineering question stays with the engineer.

Where the agent may answer and where it must stop

The boundary between answering and deciding should be written down before launch. A simple decision guide keeps the agent inside its role.

  • Answer directly from the record: sample status, which borehole an identifier belongs to, whether a custody event is signed, and which log revision is currently on file.
  • Answer with a caveat: any reading of a log description, a question about a result the laboratory has not reported, and gaps in custody records.
  • Escalate to a person: changes to a logged description, relabeling of a sample, release of a report to a client, and any statement about ground conditions that will feed a design.
  • Never do: edit a logged interval, delete a revision, or mark a sample as received without the laboratory's confirmation.

Permissions should follow the same discipline. The agent reads the record broadly, writes only short notes attached to a sample, and has no write path into the logs themselves. The OWASP authorization guidance frames this as checking each action against what the requester is allowed to do, which is the right lens for an agent that touches field data. The same logic applies to quality nonconformance and audit preparation, where the agent organizes evidence but a responsible person decides the disposition.

Keeping answers checkable over time

An answer is only useful if someone can check it later. Record each agent answer with the question, the records it read, the revisions it used and whether a person corrected it. The OpenTelemetry observability primer describes this kind of traceable recording of what a system did and when, and the same idea applies to an agent that answers from documents.

Retrieval-based designs, described in Google Cloud's architecture guidance for retrieval-augmented generation, make the check practical, because the answer is assembled from retrieved passages that can be shown beside it. Review a sample of answers on a fixed schedule, compare them with the source documents, and rerun the original test questions whenever a laboratory format, a naming convention or the agent's instructions change. A regression in a single answer is often the first sign that a record has drifted.

For material documentation in a fabrication or testing setting, the same pattern of linked certificates and job records is covered in AI for metal fabrication shops.

Choosing the first task for an agent

Be Human designs, implements and operates this kind of personalized AI infrastructure as a managed service. A first engagement defines one useful role, such as sample status for project managers, along with the memory and tools that role needs, which here means the log store, the laboratory status feed and the custody records. It also sets the approval boundaries and one observable success criterion. Managed operations then cover monitoring of answers, recurring evaluations against real questions, maintenance of the laboratory and field integrations, permission reviews, and controlled improvements when naming conventions or report formats change. The agent keeps its identity, memory and permissions under the firm's control.

The right first role is usually the one where a wrong answer is cheap to catch and a correct answer saves a phone call. Take the free Be Human AI diagnostic to find the first task worth handing to an agent in your drilling operation.

Frequently asked questions

Do we need to replace our laboratory or field software before starting? Usually not. The record can start as a structured store that links the original logs, custody forms and laboratory reports, with identifiers mapped to what the crew already uses. Replacing systems is a separate decision that should follow once the identifiers prove consistent.

Can an agent work with scanned boring logs? It can, but scanned pages are only useful once their key fields are transcribed and linked to the borehole, depth and revision. Keep the scan as the evidence and treat the transcribed fields as the searchable layer, with a person checking a sample of transcriptions during setup.

Who should own the record inside the firm? Ownership should sit with the person accountable for the project files, often a project engineer or document controller, while the agent is given read access and narrow note-writing rights. Log content stays with the engineer who signs it, and the owner decides how identifiers and revisions are maintained.

How do we know the agent is working? Define one observable success criterion before launch, such as every answer citing the record it used and the lab status matching the laboratory's report. Then review a fixed sample of answers on a schedule, rerun the original test questions after any change, and log every correction so the failures become part of the next evaluation.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →