Superintelligence Prep for Drain Cleaning: Emergency Calls, Camera Reports
An AI agent can take drain emergency calls, escalate urgent jobs to a technician, and link camera inspection findings to follow-up quotes.

An AI agent for a drain cleaning company works best when it links three steps that usually live in separate places: the emergency call, the camera inspection and the follow-up quote. The agent captures the facts of a backup quickly, escalates anything that could mean sewage inside the building or a failing main to a person, and keeps the findings attached to the same job record. It does not decide whether a line can safely wait until tomorrow, and it does not set the price.
Key takeaways
- Separate emergency intake from routine booking, using a written list of symptoms that trigger immediate escalation.
- Attach camera inspection findings to the same job record so the follow-up quote refers to what the technician actually saw.
- Let the agent draft quotes from approved scope language and owner-set price rules, but keep scope approval and pricing with a person.
- Define who gets paged, through which channel, and what happens when nobody acknowledges the page.
- Measure how long urgent calls take to reach a human and how often quote drafts need changes, not only how many calls the agent handled.
What an emergency drain call really contains
People who call about drains rarely describe the problem in technical terms. They say the shower is backing up, that the basement smells wrong, or that water came up through the floor drain after the washing machine finished its cycle. The first few minutes of a call decide whether the job becomes a same-day dispatch, a booking for later in the week, or a referral to another trade. An agent that is useful at this stage asks the same short set of questions every time, in the same order, and writes each answer into a structured field instead of a free-text note that nobody reads later.
The useful fields are usually simple. Where is the water, and is it leaving the property? Which fixtures are affected: one sink, several fixtures, a toilet, a floor drain, a tub on a lower level? Is sewage visible or is the smell strong? Does the caller know where the main shutoff or the cleanout is? Is there anyone in the home who is elderly, medically vulnerable, or without another working bathroom? Is the site commercial, and does it include food service? Access details such as gate instructions, pets and parking belong in restricted notes that only the technician sees.
The agent should not diagnose. It can tell the caller that a technician is being contacted and that a callback window will follow. It should not say that the problem is a main line blockage, because that claim belongs to the technician who has looked at the line. Keeping this boundary clear is what protects both the caller and the business when the first guess turns out to be wrong.
How to set escalation rules the office can trust
Escalation rules should be written by the owner or the dispatcher, not inferred by the model from a general instruction to be helpful. A workable starting list covers the situations where waiting would make the damage or the risk worse:
- Water rising through a floor drain, shower or tub on a lower level.
- A sewage smell inside the building together with more than one fixture backing up.
- Water near an electrical panel, or any sign that water is moving into a unit below.
- A commercial kitchen with a grease trap overflow during service hours.
- A vulnerable person in the home with no working bathroom.
- A caller who is unsure where the shutoff is while water is still increasing.
The rule set also needs the steps that follow a match. The agent pages the on-call technician through the agreed channel, sends the caller a plain message saying that a person is being contacted, and starts a timer. If nobody acknowledges the page within the agreed window, the case moves to a second contact, usually the owner. Routine calls follow a different path: the agent offers open windows in the schedule and states an arrival range, not an exact time it cannot control.
Dispatch and callback loops are a familiar pattern for plumbing firms, and the same logic applies here. The approach used for plumbing dispatch and callbacks shows how to keep a open case visible until someone owns it. Drain work adds one more layer: the urgency often depends on what the camera shows later, so the escalation list should be reviewed whenever the inspection tags change.
Linking the call to the camera inspection
A camera inspection produces video, a few still frames and a technician's notes. Most of the value is lost when these sit on a device or in a folder that nobody searches after the job closes. Link them to the same job identifier, the same address and the same access point that the technician used. A finding that cannot be connected to the original call is hard to explain to the customer and hard to learn from later.
The technician should tag each finding with a short, controlled vocabulary. Useful tags include blockage type, location relative to the access point, pipe material when it is visible, root intrusion, offset or sag, scale buildup, cleared or not cleared, and recommended next step. Tags are what an agent can read reliably. Long narrative notes are useful for people, but a fixed list gives the agent something it can match against approved scope language without guessing.
Illustrative scenario: A homeowner calls at night and says the kitchen sink and the basement floor drain both gurgle after a load of laundry. The agent classifies the call as urgent, pages the on-call technician, creates the job with the address and the answers, and tells the caller that a person is being contacted. The next morning, the camera run shows roots entering the line from the street side. The technician tags the root intrusion, marks the line as partially cleared, and adds a note. The agent now has the original call and the tagged findings in one record, so the follow-up is built from evidence rather than from a blank template.
Mapping a fault to a precise position on a buried line follows the same logic described for irrigation contractors who plan repairs by zone: the record should say where the problem is, not only that a problem exists.
Turning inspection findings into a follow-up quote
A follow-up quote should describe the work the footage supports, not the work the office would like to sell. Build a service menu with plain scope language for each item: clearing with a machine at a stated access point, spot repair at a named location, root removal, camera re-inspection after clearing, and a separate evaluation visit for anything that may need excavation or lining. Each item has approved scope text and a price rule that the owner sets. The agent selects items that match the tags on the job and drafts the quote from that selection. It does not invent items or combine them in ways the menu does not allow.
Ambiguity needs its own path. If the tags show an offset, a possible collapse or a note that says unclear, the quote should route to a technician evaluation with a clear statement that the scope depends on a site visit. Warranty and exclusion wording matters at this point. The customer should know what happens if roots return, and what is excluded when the line is old, the access is blocked or the condition was not visible. The way wildlife removal companies handle exclusions and warranties is a useful model for stating limits without overpromising.
Before a draft quote goes out, the office should be able to confirm five things:
- Every line item traces to a tagged finding or to a symptom the caller described.
- The scope text matches the approved menu wording.
- Exclusions for conditions that were not visible are stated.
- A technician has reviewed the footage for any finding marked ambiguous.
- The owner's price rule, not the agent, sets each amount.
Where the agent stops: permissions and approvals
Give the agent only the permissions each step needs, and review them the way you would review staff access. The OWASP authorization guidance describes the principle that each action should be checked against what the identity performing it is allowed to do. For this workflow, a reasonable starting set includes creating and updating job records inside the service area, sending acknowledgments and status messages from approved templates, paging the on-call technician and logging the response, offering open windows in the schedule, and attaching camera files to a job with a draft quote waiting for review.
Some actions should always require a person. The agent should not state a cause as fact, promise same-day completion for a complex job, set or change any price, issue refunds or credits, change the service area or hours, access payment details, or close a job that carries an ambiguous tag. These limits are easier to enforce when they are written as a list the system checks, not as a sentence in a prompt that the model may or may not follow in every case.
Callers, voicemail transcripts and field notes can contain text that reads like an instruction, such as a message telling the agent to ignore the rules and book the caller first. Treat all of that as data to be read and classified, never as a command. The OWASP prompt injection prevention cheat sheet describes this class of risk in detail. The NIST AI Risk Management Framework is a practical reference for documenting who owns each decision, how the risk is reviewed and what evidence supports the rules.
Measuring whether escalation works
Measure the path, not only the volume. For urgent calls, the useful questions are how long it took from first contact to a human being notified, whether that person acknowledged the page, and whether the caller received the message the rules call for. For quotes, measure how many drafts needed material changes before they were sent, and how many needed a technician to re-check the footage. Each of these measures depends on a log with a timestamp for every step. The OpenTelemetry concepts of traces and events give a useful vocabulary: one trace per call, with events for classification, paging, acknowledgment and quote drafting.
Build a fixed set of sample call transcripts and tag sets with known correct outcomes, and rerun that set after any change to the prompts, the escalation list, the service menu or the model. Include the difficult cases: a caller who says there is only a little water, a commercial kitchen after closing, a tag set that includes the word unclear. Add to the set whenever a real call surprises the team. This is the same discipline that the [OpenAI evals guide](https://platform.openai.com/docs/guides/evals) describes for testing model behavior against expectations.
Be Human designs and operates this kind of agent as a managed service. A first engagement defines one useful role, here the after-hours intake and escalation path, along with the memory it keeps (service area, approved scope text and the escalation list), the tools it can call, the approval boundaries and an observable success criterion, such as urgent calls reaching a technician within the agreed window. Ongoing operations include monitoring those logs, running recurring evaluations, reviewing permissions, maintaining integrations with the scheduling and paging tools, and making controlled changes when the escalation list or the service menu changes. Pricing and the promise made to customers stay with the owner.
Where to start
Start with one narrow flow, usually the after-hours call and its escalation, before adding camera-based quotes. Once the page-to-technician path has run cleanly for a period the team trusts, the quote draft can follow, using the tags the technicians already produce. The free Be Human AI diagnostic can help you find the first task worth handing to an agent in your drain business.
Frequently asked questions
Can an AI agent answer emergency drain calls after hours? It can answer the call, collect the key facts and page an on-call person, as long as the escalation list is written and the paging channel has been tested. The agent should tell the caller that a person is being contacted and should not diagnose the problem. The dispatch decision still belongs to a human.
Does the agent watch the camera footage? Usually it does not analyze the video itself. The technician tags each finding from a short controlled list and adds a note, and the agent works from those tags and notes. The technician still reviews the footage and confirms anything marked as unclear before a quote is sent.
Who sets the prices in the follow-up quote? The owner sets a price rule for each approved service item, and the agent only selects items and drafts the text from approved scope language. The agent cannot change an amount or invent a scope item. Any ambiguous finding routes to a technician evaluation instead of a fixed number.
What should we measure first? Start with how long urgent calls take to reach a human and whether those pages were acknowledged, since those show whether the most important path works. Then track how many quote drafts needed material changes before they were sent. Add other measures only after these two are stable and reviewed each month.
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.


