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

AI by industry

Sewer Inspection Reports and Quotes: Agent Follow-Ups From the Record

Link camera inspection reports and repair quotes to each property record so an AI agent drafts sewer follow-ups from the findings, not memory.

Be Human typographic cover with the title: Sewer Inspection Reports and Quotes: Agent Follow-Ups From the Record
Conceptual illustration created with AI · Be Human

An AI agent helps a sewer repair company most when it writes customer follow-ups from the camera inspection report and the repair quote stored on the property record, instead of from a technician's recollection or a sales rep's notes. The agent reads defect codes, footage references, distances from an access point and each quote line, then drafts a message that names the specific finding and waits for office approval. Done this way, every follow-up traces back to a document the customer has already received.

Key takeaways

  • Link each inspection report and each quote to one property record with stable identifiers before any agent drafts anything.
  • Let the agent write only from defects, distances and quote lines it can cite, and have it mark missing fields as gaps instead of filling them in.
  • Keep price changes, scope changes and warranty wording behind human approval, with the agent limited to drafts and tasks.
  • Track each finding through a clear decision state: accepted, deferred, declined, or no response.
  • Judge the setup by whether follow-ups name the correct finding and whether the record reflects the customer's decision within the window your team agrees on.

Why follow-ups drift after a camera inspection

A camera inspection produces several artifacts on the same day: a video file, a coded defect list, a technician's verbal summary and often a written report. The repair quote usually comes later, may be prepared by someone who did not run the camera, and carries its own line items. By the time a coordinator calls the homeowner or property manager a week or two later, the finding has been compressed into a sentence such as the line needs work. That sentence is where the follow-up fails, because it cannot say which lateral is affected, whether the problem is a root intrusion, an offset joint or a collapsed section, or which option can be deferred without consequence.

Memory also fails across people. The technician may be on another job, the estimator may not have reread the footage notes, and the office may have heard a verbal decision that never reached the file. An agent that drafts from whatever sits in the inbox will repeat the same compression, only faster and with more confidence.

What the property record needs to hold

Before an agent can draft anything useful, the record has to connect the inspection and the quote through stable identifiers. At minimum, each property record should hold:

  • A property identifier and service address, with access points such as cleanouts and manholes noted by location
  • The inspection date, the crew, and a reference to the stored footage file
  • Each defect entered as a code, a distance from the access point, and the technician's plain-language note
  • The recommended repair method for each defect, and whether the repair is required or optional
  • The quote identifier, with each line item mapped to a specific defect rather than a general description
  • The options presented to the customer and the date they were sent
  • The customer's decision and the date it was recorded, including decisions made by phone
  • Open questions, such as a missing distance, an unclear photo or a footage file that failed to upload

The mapping between defects and quote lines matters most. A quote line that reads lateral repair with no defect reference invites the agent to guess which problem the line addresses, and a guess in a customer message is hard to walk back. Pattern-based approaches to this problem are covered in general terms by the Superintelligent Agents for Plumbing Companies: Dispatch and Callbacks article, since callback-heavy trades share the same problem of decisions living in several places at once.

How the agent drafts a follow-up from the findings

The pattern is retrieval followed by drafting. When a follow-up falls due, the agent pulls the property record and only the fields it needs: the defect notes, the quote lines, the options, the last contact and the recorded decision. It then writes a draft that cites those facts in plain language and leaves out anything it cannot source. This is the basic idea behind retrieval-augmented generation, in which a model answers from supplied documents rather than from what it absorbed during training. Google's architecture guidance on retrieval-augmented generation describes how this pattern is usually assembled.

Illustrative scenario: a property manager received an inspection report showing a root intrusion at a lateral about a few meters past the cleanout, with a footage reference attached, and a quote with two options, a targeted spot repair and a longer section replacement. Ten days later, no decision is recorded. The agent drafts a short message that names the lateral, refers to the footage file, restates the spot repair option and asks which option the manager wants scheduled. It does not promise a date or describe the cause beyond what the report says. If the record also shows a phone conversation where the manager mentioned a preference for the longer option, the agent flags the mismatch for the office instead of sending either version.

The draft is only as good as the record behind it, which is why the agent should show its sources to the person approving the message. An approver who can see which defect, which quote line and which date were used can correct a mistake in seconds.

When the agent must stop and hand the message to a person

Not every follow-up belongs to the agent. A short decision guide keeps the boundaries visible:

  • Routine reminder for a quote the customer has seen and not answered: the agent drafts, an office lead approves, then the message is sent
  • Report and quote disagree on scope, location or defect count: the agent holds the draft and opens a review task for the estimator
  • Customer mentions flooding, backup or odor inside the building: the agent stops drafting and routes the conversation to dispatch
  • Request to change a price, discount or warranty term: a person writes the reply, and the agent only supplies the relevant record fields
  • Missing footage reference or missing defect distance: the agent sends nothing that claims a specific location until the gap is filled
  • Customer disputes a finding: the agent attaches the footage reference and asks a technician to review it before any further message

The last rule is worth stating directly. A customer who doubts a finding should not receive an automated reply that repeats the finding more firmly. They need the footage, a technician's review and a human response.

Permissions, tools and an audit trail

Treat the agent as a staff member with a narrow role. It should read the property record and the stored quote, create drafts, and create tasks for people. It should not edit quote totals, issue credits or change warranty terms, and it should not reach into the full field software when a single record would do. OWASP's authorization guidance recommends granting only the access a task needs and checking permissions on each request, which means the agent's access should be scoped to the record and the message types it handles.

Inspection notes and customer emails are also untrusted text. A customer comment or a free-text field in a report can contain instructions that an agent might follow if the system does not keep instructions and data separate. OWASP's prompt injection prevention guidance describes separating those inputs and validating outputs before they trigger an action. In practice, every draft and every approval should be logged with the sources it used, so an office lead can later see why a message said what it said and whether the record was the reason.

A draft that cannot show its sources should not reach a customer.

A 30-day plan to start with one message type

Starting with one message type keeps the work measurable. A reasonable first month looks like this:

  • Week 1: list the fields your inspection reports and quotes already capture, and mark which ones are missing or stored only inside a PDF
  • Week 2: link a sample of recent quotes to their property records and reconcile each pair by hand, noting every mismatch and why it happened
  • Week 3: run the agent in shadow mode on one follow-up type, so it drafts messages that nobody sends, then compare each draft with what the office actually wrote
  • Week 4: enable approval-gated sending for the one message type that matched the record well, and review every draft for a set period before widening the scope

The shadow week is the most useful. Differences between the agent's draft and the office's message usually reveal a field that exists in someone's head but not in the record, and fixing that field improves the whole process.

How Be Human runs this in day-to-day operations

Be Human designs this kind of setup as a managed service. The first engagement defines one role, here a follow-up drafter for inspection findings, along with the memory and tools it needs, the approval boundaries it must respect, and a success criterion the team can observe, such as whether drafts name the correct defect and whether the record reflects the customer's decision within the agreed window. Managed operations then cover monitoring of draft accuracy, recurring evaluations against fresh records, maintenance when the field software or quoting tool changes, periodic permission review, incident response when a message goes wrong, and controlled improvements to the drafting rules. The work is remote and available in English and Portuguese.

If you want to find the first task worth handing to an agent in your own operation, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent read the camera footage itself? In most setups it works from the coded defects, technician notes, distances and footage references stored in the record, not from raw video. If your team reviews footage and writes a structured summary, the agent can work from that summary and cite the footage file for a person to check. Anything the summary does not state should stay out of the draft.

Does the agent send repair quotes or customer messages on its own? It should not, at least not at the start. A sensible first design lets the agent create drafts and tasks, while an office lead approves each message type before it goes out. Sending can be widened only after the drafts have been reviewed over a period long enough to show consistent accuracy.

What happens when the inspection report and the quote disagree? The agent should hold the draft and open a review task for the estimator, listing the specific fields that conflict. It should not choose one version and present it to the customer. Resolving the conflict is a human decision, and the record should show who made it.

How long does it take to get this running? The timeline depends on how many systems hold your inspection reports and quotes, and on how consistently the defect and quote fields are filled in today. A small team with one field tool and clean records can start with a shadow run quickly, while a business with paper reports and several quoting tools needs more time to build the record first. A first engagement usually begins by mapping those gaps before any agent is switched on.

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 →