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

AI by industry

AI for Sewer Inspection Firms: Camera Logs and Repair Quotes

An AI agent that indexes sewer camera logs and drafts repair quotes for inspection firms. The inspector confirms each finding before anything is sent.

Be Human typographic cover with the title: AI for Sewer Inspection Firms: Camera Logs and Repair Quotes
Conceptual illustration created with AI · Be Human

An AI agent for a sewer inspection company can index camera inspection videos together with the operator's logs, so the office can search defects by street, pipe segment and inspection date. It can then draft a repair quote from findings the inspector has confirmed. The agent prepares the draft, and the inspector reviews every finding before any quote, note or message leaves the company.

Key takeaways

  • Start with one defect workflow, such as lateral defects found during municipal or commercial runs, instead of the whole office.
  • The agent indexes footage metadata, operator notes, defect codes and tagged frames so past runs can be searched by segment.
  • Repair quotes are drafted from confirmed findings and the company's own rate sheet, never from prices the model remembers.
  • The inspector confirms each finding before anything is sent, and the agent holds draft-only permissions.
  • Success is judged by how often drafted findings survive inspector review, not by how quickly a draft appears.

Why the video and the paperwork drift apart

A sewer camera run produces two records that rarely live together. The footage sits on a drive or in a cloud folder, while the operator's log sits in a field form, a spreadsheet or a notebook in the truck. A defect that matters to the customer, such as an offset joint or a root intrusion at a lateral connection, is described in one place and timestamped in another. When the office builds the report and quote days later, someone has to scrub the video again to match a note to the right frame.

The drift gets worse with history. A segment inspected on separate occasions may have notes from two operators, one who used the company's defect codes and one who wrote free text. Whether a defect is new, already repaired or recorded differently changes the repair decision. In many firms that memory lives in a senior technician's head, which means it leaves when that person changes roles.

What the agent indexes and what it must not invent

The agent's first job is an index, not a judgment. For each run it records the asset identifiers the company uses, the start and end of the run, footage timestamps, the operator's observations, the defect codes applied, and the pipe material and access points where the record names them. Tagged frames and photos are stored with links back to the original files, so anyone can open the source.

When the office asks a question, such as which laterals on a given block showed root intrusion in the last inspection, the agent retrieves matching entries and answers with source references attached. Grounding answers in stored records rather than in general knowledge is the central pattern of retrieval-based systems, and it is what makes the answer checkable. Inspection records also follow the discipline of nonconformance records, a pattern covered in AI for Quality Management: Nonconformances and Audit Prep.

The rule is simple: if the record does not contain a fact, the agent says so. It does not infer a diameter, a material or a depth from a similar pipe nearby. Missing fields appear as explicit gaps so the inspector can fill them during review.

A step-by-step plan for the first indexing project

  • Step one: pick one asset group, such as one municipal contract or one commercial property, and export the most recent complete set of run records and footage metadata.
  • Step two: agree on the defect code list, then map the free-text terms operators use to those codes. The lead inspector signs off on this mapping.
  • Step three: index the records with source links, and mark incomplete entries as gaps rather than filling them.
  • Step four: prepare a set of questions whose answers the office already knows, run them against the index, and check each answer against its source.
  • Step five: log every correction the inspector makes with a short reason. These corrections become the evaluation set for later changes.
  • Step six: add the next asset group only after corrections stop clustering around one defect type or one operator.

Drafting repair quotes from confirmed findings

A quote draft starts only after the inspector marks findings as confirmed. The agent groups confirmed defects by segment and proposes a repair method from the company's own options, such as spot repair, a liner or excavation. Each proposal is tied to the recorded defect. Line items and unit rates come from the company's approved rate sheet, which the agent reads but never supplements from general knowledge. The output is a draft in the company's template, with a short list of assumptions and an explicit note on any item the record does not support.

A practical decision guide for the office:

  • If a defect is confirmed and the record cites a source frame, the draft can carry it as a line item.
  • If a defect appears only in free text with no frame, the draft marks it for the inspector's review and does not price it.
  • If the repair method depends on access, depth or material that is missing, the draft asks for that data instead of assuming it.
  • If the index shows a prior repair on the same connection, the draft states the earlier work and does not present it as new.

Callbacks after a repair follow a similar pattern of linking records to work orders, a logic explored in Superintelligent Agents for Plumbing Companies: Dispatch and Callbacks.

Illustrative scenario: a lateral defect with an earlier repair

The following scenario is illustrative and does not describe a real client. An inspection run on a residential block notes a root intrusion at a lateral connection and a possible offset. The operator tags two frames and applies a defect code. The agent finds, in the index, an earlier run on the same segment that recorded a spot repair at that connection. Its draft does not propose a new spot repair as if the problem were new. It lists the earlier repair with its source, flags the current observation as a possible recurrence, and asks the inspector whether the frames show the same connection.

The inspector reviews both frames and corrects the offset note, which the frames do not support. The inspector confirms only the root intrusion. The agent regenerates the draft with one confirmed line item, notes the prior repair as context, and sends nothing. The correction is logged, and it becomes a test case for the next evaluation.

Approval boundaries and security

The agent has read access to the run index and footage metadata, draft write access to quote templates, and no permission to send anything. Sending a quote, changing a confirmed finding or editing the rate sheet belongs to named staff. Those permissions should be reviewed on a schedule, not only at setup.

Operator notes, customer messages and text pasted into the system are inputs, not instructions. A note that says to ignore the rate sheet must be treated as data to be reviewed, never as a command. The OWASP guidance on prompt injection explains why content from external sources should not change how an agent behaves. In practice that means narrow tool scopes, logs for every draft and correction, and traceability from every quote line back to a source. The NIST AI Risk Management Framework offers a useful vocabulary for assigning each of these risks to an owner and a control.

  • The agent cannot send email, messages or documents to customers.
  • Every quote line links to a confirmed finding and a source frame.
  • Changes to the rate sheet require a named approver.
  • Each correction is logged with its reason.
  • Access to footage is limited to the asset groups the current job covers.

Choosing where to start and how Be Human runs it

Be Human designs this kind of agent as a defined role: its memory, its tools, its approval boundaries and one observable success criterion, such as how often drafted findings are confirmed without edits. Managed operations cover monitoring of the index, recurring evaluations against inspector corrections, maintenance of connections to the inspection software and file storage, periodic permission review, incident response when a source fails or a draft goes wrong, and controlled changes to prompts or templates that are tested before reaching the office. The service is remote and available in English and Portuguese. Zone-based records and repair logic follow a similar structure, described in Superintelligence Prep for Irrigation Contractors: Zones and Repairs.

The best first workflow is usually the one with the most repeated questions and the clearest source records. Take the free Be Human AI diagnostic to find the first task in your operation worth handing to an agent.

Frequently asked questions

Can an AI agent send repair quotes to property owners on its own? No, not in this design. The agent drafts quotes from confirmed findings and has no permission to send. A named staff member reviews the draft, edits it if needed and sends it through the company's existing process.

Does the agent replace the inspector's judgment on defect codes? No. The agent indexes what the records say and flags gaps and conflicts. The inspector decides which findings are confirmed, and those decisions are the basis for every draft.

What does the agent need before it can start? It needs an export of run records, footage metadata and operator notes for one asset group, plus the company's defect code list and approved rate sheet. Without a clear code list, the first step is agreeing on the mapping from free text to codes.

How is the agent's accuracy checked over time? Accuracy is checked against a set of known questions and against the inspector's logged corrections. Each change to prompts, templates or connections is tested on that set before it reaches the office, and the results are reviewed as part of the managed service.

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 →