Superintelligence Prep for Excavation Companies: Dig Tickets and Job Logs
How excavation firms can use dig tickets and job logs as records for an AI agent, and which site-critical decisions a person must keep.

An AI agent for an excavation company can answer questions about locate status, keep job logs current, and prepare crew assignments for the next day, but only if the records it reads are clean and its permissions stop short of site-critical decisions. The practical first step is to treat dig tickets and daily job logs as the company's operational memory, give the agent read access and narrow draft access to them, and keep clearance, depth, and stop-work calls with people.
Key takeaways
- Dig tickets and job logs are the most reliable operational memory an excavation firm already produces, once they are structured and dated consistently.
- An agent should read locate status and job history from the system of record and never declare a site clear on its own.
- Permissions should separate drafting and reminding, which the agent can do, from clearing, approving, changing depth or route, and stopping work, which a person must do.
- Every agent action on a ticket or log should leave a trace that a supervisor can review later.
- Superintelligence prep here means structuring records and boundaries so more capable systems can be added later, not buying a promise of autonomy.
What dig tickets and job logs actually contain
A dig ticket, usually called a locate request, records where a crew plans to excavate, when the request was made, which utility owners were notified, and what they reported back: marked, clear, or a conflict that needs a site visit. A job log records what happened on site: who was on the crew, which equipment ran, how deep the trench reached, what material moved, which photos were taken, weather stops, changes from the plan, and any incidents or near misses. Together they form the closest thing a dirt contractor has to a site history.
The difficulty is that these records often live in several places at once: a dispatch board, a clipboard in the truck, text threads with a foreman, photos on personal phones, and a spreadsheet that someone updates on Fridays. Before an agent can help, the owner needs to know which source is the authority. For a ticket, the authority is the locate response from the notification system, not a text that says the lines are probably clear. For a job log, the authority is the signed daily entry, not an undated photo. Deciding this is the first real design choice, and it costs nothing to make.
Why these records make a good foundation for an agent
An agent that answers questions about past and present work should retrieve records at the moment it needs them rather than rely on what it was trained on. AWS's comparison of retrieval-augmented generation and fine-tuning describes why this matters: retrieval keeps answers tied to source documents that can be updated and checked, while fine-tuning bakes information into a model that is harder to correct when facts change. For an excavator, facts change constantly. A ticket expires, a utility re-marks a line, or a crew returns to a block that was already logged as complete. The answer has to change the same day.
The same logic explains why superintelligence prep starts with records rather than with the model. Any future system, however capable, will only be as reliable as the tickets, logs, and permissions it inherits. Our note on zone maps and repair histories for irrigation contractors makes a similar point for another kind of ground work: a record is useful when each entry has a date, a location, and a responsible person. The same test applies to a trench record.
A plan to prepare dig tickets and job logs before any agent touches them
Work through these steps before connecting an agent to anything:
- Step 1: List every place a ticket or log entry is written today, including texts and personal phones, and name one system of record for each record type.
- Step 2: Give each ticket one identifier that appears on the locate request, the marked site, the crew sheet, and the job log.
- Step 3: Define the fields a ticket must have before it can be scheduled: service area, requested dates, notification status, and expiry date, all as plain values rather than free text.
- Step 4: Standardize the daily job log into fixed fields for crew, equipment, depth reached, material moved, time-stamped photos, changes from plan, and incidents.
- Step 5: Decide who may correct a closed log and how the correction is recorded, so that original entries are never overwritten without a trace.
- Step 6: Write twenty real past questions a dispatcher or foreman would ask, with the correct answers, and keep them as the first test set for the agent.
The last step is often skipped, and it is the one that shows whether the records are good enough. If nobody can agree on the correct answer to a past question, the agent will not be able to give one either.
Decision boundaries: what the agent may do and what a person must keep
The useful split is between work that drafts, summarizes, and reminds, and work that commits the company to a physical action. The agent may summarize open tickets by crew and by expiry window. It may draft next-day crew assignments for a dispatcher to review. It may flag tickets that are close to expiry or carry a conflict status, and notify the named coordinator. It may pull earlier job logs and photos for a location so the foreman knows what was found there before. It may draft a job-log entry from a crew member's voice note, marked as a draft until the foreman signs it.
A person must keep the following decisions:
- Confirming that a site is clear to dig, by reading the locate response for the specific area and checking the marks on site.
- Approving any change to depth, route, shoring, or equipment placement near marked or suspected lines.
- Deciding stop-work, and making sure nobody, including an agent, resumes work that a person has stopped.
- Signing incident and near-miss reports, which the agent may only pre-fill from the log.
- Changing the status of a ticket or closing a log that carries an open question.
Crew notes and transcribed voice messages are inputs, not instructions. If a note says something that sounds like a command to skip a check, the agent should report that text to the coordinator, not act on it. OWASP's prompt injection prevention guidance describes this class of risk, and it applies to any free text that someone other than the owner can type into a record the agent reads.
Illustrative scenario: a ticket that expires before the crew returns
Illustrative scenario, not a real client: a crew marks out a trench on a Thursday, but the locate ticket for the far block will expire before the crew can return the following Monday. The agent sees the expiry date in the ticket record. It notifies the coordinator with the ticket identifier, the block, and the next scheduled date, and it drafts a renewal request for a person to send. It does not tell the foreman that the block is clear.
When the coordinator renews the request and the new response arrives, the foreman reads the response for that block, confirms the marks on site, and approves the start. The job log records who decided, when, and on which response. The sequence needs only a short, specific amount of human attention, and it removes the common failure in which a ticket is assumed valid because nobody checked the date.
Permissions, audit trails, and testing the agent
Give the agent its own identity and the least permission it needs. OWASP's authorization guidance describes checking access on every request rather than trusting the interface or the caller. In practice this means read access to tickets and logs, write access only to draft fields, no ability to change a ticket's status, and no path to dispatch equipment directly. Every action should leave a trace: what was read, what was drafted, and which person approved it. A supervisor should be able to reconstruct any decision after an incident without asking the agent to explain itself.
Testing is the second half of the discipline. Keep the test set of past tickets and logs, rerun it after any change to the model, the instructions, or an integration, and review failures on a regular schedule. The NIST AI Risk Management Framework is a useful reference for this habit: identify the risks, measure them, and keep a named person accountable for the result. The same logic appears in our notes on quality management, where nonconformances are only useful when each one can be traced to a cause.
How Be Human approaches this: Be Human starts with one role, often the ticket and log coordinator, and defines its memory sources, tools, approval boundaries, and one observable success criterion. The agent keeps its identity and permissions under the client's control, runs recurring evaluations against the client's test set, reviews permissions when crews, software, or integrations change, and follows an incident process if an integration fails or an unexpected write occurs. The client keeps the decisions about clearance, depth, and stop-work.
Preparing for superintelligence without betting on it
Superintelligence is a term people use loosely, and no excavation company should buy an operations project on the promise of it. What is concrete is this: more capable systems will inherit whatever records, permissions, and logs a company has today. If tickets are unambiguous, logs are structured, and every action is traceable, a business can later add stronger models without rewriting its operating rules. If not, a more capable system will simply make confident errors faster and with less visibility.
Oversight also matters. The OECD AI Principles emphasize human-centered values, transparency, and accountability, and on a dig site those words have a plain meaning: the person who signs for the work should be able to see which records the system used and why it proposed what it proposed. Take the free Be Human AI diagnostic to find the first task in your excavation business that is worth handing to an agent.
Frequently asked questions
Can an AI agent tell whether a site is clear to dig? No, and it should not be set up to make that call. An agent can read the locate response, check its date and status, and flag a missing area, but a person confirms clearance on site against the actual marks. Any clearance language from an agent should be treated as a draft for a person to verify.
Do we need to replace our dig ticket software before starting? Not necessarily. Start by naming the system of record and standardizing the fields you control, such as the ticket identifier, the expiry date, and the crew assignment. Many firms begin with a clean export or a shared sheet, as long as every entry has an owner and a timestamp.
What should happen when a crew note contradicts the ticket? The agent should report the conflict to the coordinator and should not resolve it on its own. The ticket record and the locate response stay authoritative until a person updates them through the normal process, and the reason for the change is recorded.
How do we know the agent is still doing its job after a few months? Rerun your test set of past tickets and logs after every change, review failures on a fixed schedule, and check the permission list whenever crews, tools, or integrations change. Be Human's managed operations include this kind of recurring evaluation and permission review, but the business owner still decides what counts as acceptable performance.
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.


