Septic Inspection Report: Can AI Write It From Field Notes?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Septic Inspection Report: Can AI Write It From Field Notes?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Septic Inspection Report: Can AI Write It From Field Notes?

Yes, AI can write a septic inspection report from field notes and photos: it can organise your observations, tidy the wording and caption images. It cannot decide what the findings mean. The inspector still rates each component, chooses the recommendations, checks every figure against the notes, and signs. AI shortens the write-up; the professional judgement stays with the inspector.

Most inspectors do not lose their evenings to the inspection. They lose them to the write-up: translating shorthand like "tank 1000 gal, baff in OK? scum thick, DB wet" into clean sentences, matching photos to the right sections and finding the polite way to say "this needs attention". That translation is where AI helps. The risk sits in the same place, because a model can turn a question mark in your notes into a confident statement. Everything below is about getting the speed without that risk.

Follow me on Instagram@sagnikteaches

Write notes the AI can read and the inspector can trust

The report is only as good as the notes. Shorthand is fine, but it needs a fixed structure so that nothing is silently missing. Use the same headings on every job, and write NR (not recorded) or NA (not accessible) rather than leaving a line empty. An empty line is ambiguous, and a model will fill ambiguity with a guess.

Connect on LinkedInSagnik Bhattacharya

A workable layout, one block per component:

Subscribe on YouTube@sagnikteaches
  • Job header: date, property type, bedrooms or occupancy as given by the owner, reason for inspection, who attended.
  • Tank: type and material, stated or estimated capacity (and which), number of compartments, lid and riser condition, inlet and outlet baffle or tee condition, liquid level, scum and sludge observations or measurements, signs of backflow.
  • Pump chamber or lift station (if any): pump running or not, float and alarm tested, panel condition.
  • Distribution box: found or not, level, evidence of flow, standing liquid.
  • Drainfield or soil absorption area: location, surface conditions, odour, wet patches, vegetation growth, signs of surfacing effluent.
  • Owner statements: last pumping date, known problems, recent changes in use, kept as quotes in the owner's words.
  • Limits: anything you could not see or test and why.

Voice notes work well as a first pass, spoken from the driveway or the tank lid. If you already turn voice into job sheets, the same habit is described in turning on-site voice notes into job sheets and quotes. Transcribe, then clean into the layout above before the model sees it; ten seconds of tidying prevents a paragraph of nonsense.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

A filled-in set of notes and the draft that comes back

Below is a short, illustrative set of field notes for a single-family home, with the instruction that produces a draft. Everything is invented for demonstration.

Write the findings section of a septic system inspection report from
the field notes below. Rules:
1. Use only facts in the notes. If a field says NR or NA, write
   "Not recorded" or "Not accessible" and say nothing more about it.
2. Do not give a pass/fail rating, repair cost or code reference.
   Use the rating I give you for each component.
3. Describe what was observed, then what it may indicate, using
   "may", never "is" for anything inferred.
4. Plain English a homeowner can follow. Short paragraphs.
5. End with a list of every field marked NR or NA.

NOTES
Tank: concrete, 1000 gal (owner stated, not confirmed), 2 compartments.
Lids: 2 found, 1 cracked at edge, no risers.
Inlet baffle: present, intact. Outlet baffle: NA, could not see, lid seized.
Liquid level: at outlet invert. Scum: approx 8 in (measured). Sludge: NR.
Backflow marks on inlet pipe: none seen.
D-box: found, lid off, water standing in one outlet, others dry.
Drainfield: lawn, grass greener over two lines, soft ground approx 6 ft
from D-box, no odour, no surfacing.
Owner: last pumped "about 5 years ago", slow drains in lower bathroom
since spring.
RATINGS (inspector's own): tank = needs attention; d-box = needs attention;
drainfield = monitor.

An illustrative extract of what a good draft returns, with the points an inspector would fix marked afterwards:

The concrete tank (owner stated 1,000 gallons; capacity not confirmed) has two compartments. Two lids were located; one is cracked at the edge and neither tank has risers, which makes future access harder. The inlet baffle was present and intact. The outlet baffle was not accessible because the lid was seized. Liquid was at the outlet level. About 8 inches of scum was measured. Sludge depth was not recorded. No backflow marks were seen on the inlet pipe. Rating: needs attention.

What the inspector fixes: "Neither tank has risers" should be "neither lid has a riser" (a slip in wording); "about 8 inches of scum" is correct only if the notes say it was measured, which they do; and the sentence about the outlet baffle should state the consequence the inspector decides on, such as "the outlet baffle condition is unknown and should be checked at the next pumping". The model neither knew nor should have guessed that. The recommendation is the inspector's.

Findings wording: describe, then qualify, then recommend

Reports get disputed over adjectives. "The system is failing" invites an argument; "effluent was standing in one distribution box outlet and soft ground was found about six feet away" invites a conversation. Teach the model a three-part pattern and apply it to every finding.

  1. Observation: what you saw or measured, with where and how.
  2. Qualified meaning: what it may indicate, using hedged verbs, and what you could not determine.
  3. Recommendation: the action and a time frame you choose, such as "have the outlet baffle inspected at the next pumping" or "monitor the wet area over the next season".

Before and after, using an illustrative line. A weak draft reads: "The drainfield has problems and needs repair." The corrected line reads: "Grass was noticeably greener over two lines of the drainfield and the ground was soft in one area about six feet from the distribution box. There was no odour and no surfacing effluent. Water was also standing in one distribution box outlet while the others were dry, which may mean more flow is reaching one line than the others, but the cause was not determined. Recommendation: have the distribution box checked for level and re-check the area after a period of heavy use."

The second version contains nothing the notes do not support. Notice that it says "checked for level", not "levelled": the notes record standing water, not a measured tilt, so the draft must not promote a likely cause into a finding. It is longer than the first version and much harder to dispute. You can also tell the model: "Rewrite every sentence that states a cause as a sentence that states an observation plus a possible cause", which catches the commonest overreach.

Photo captions: label what the picture shows, not what it proves

A photo caption needs three things: where it was taken, what it shows and the time or note number. Models can help you write them in bulk if you give them a list. Send your numbered photo list and notes, not the pictures themselves: "Photo 14 = outlet lid seized, as found; Photo 15 = D-box with standing water at outlet 3". The result is clean captions such as "Photo 15: Distribution box with water standing at outlet 3 at the time of inspection."

Avoid asking an AI to look at a photo and say what is wrong. It can describe colours and shapes, and it will do so with confidence, but it cannot judge flows, levels or damage inside a tank from an image. If you do use a tool that reads images, treat any claim as a prompt for you to check, as set out in what happens when ChatGPT reads photos and PDFs. Before a photo goes anywhere near a general chatbot, strip location data and crop out house numbers, vehicle plates and faces.

Recommendations the inspector writes, wording the AI can polish

Keep a short library of approved recommendation sentences, written once by you or a qualified colleague, covering the situations you meet again and again: pump due, lid damaged, outlet baffle not accessible, effluent surfacing, pump or alarm not working, distribution box unlevel, trees close to the drainfield, and so on. Each one has the same shape: what to do, how soon and who should do it. The model's job is to pick from the library by matching your notes, not to compose new ones.

That design has two benefits. The recommendations are consistent across inspectors, and you can show them to a regulator, insurer or lawyer as fixed text. Anything outside the library, the model flags as "NO MATCHING RECOMMENDATION" and the inspector writes it by hand. Be wary of letting a model estimate repair costs, because it will give you a figure with no basis in your market.

The realistic mistake: a question mark that became a fact

Picture, as an illustration, a small inspection business in which a draft said "inlet and outlet baffles are intact and functioning". The notes said "inlet baffle OK, outlet ? lid stuck". The model merged the two into a clean positive sentence, and the inspector, working through three reports in an evening, skimmed it and signed. The error surfaced later, when a pumping crew found the outlet baffle broken.

Nothing exotic caused that. The prompt did not tell the model what to do with doubt, and the review did not compare notes against the draft line by line. Two changes close the gap. First, put the NR, NA and "?" rule in every prompt. Second, review with the notes beside the draft and tick each statement against its source, instead of reading the report as prose. Reading prose is how smooth wording gets past a tired inspector.

A second case: a pump chamber and an alarm that nobody tested

Not every system is a tank and a field. Take an illustrative property with a pumped system, where the notes read: "Pump chamber lid found, bolted. Pump running on test, float lifted, alarm light on panel, horn not heard. Panel door rusted at hinge. Owner says alarm has never sounded." The temptation is for the draft to say "pump and alarm tested and working". Your notes say something narrower: the pump ran when the float was lifted, the alarm light came on, and the horn was not heard.

The honest report line is: "The pump ran when the float was lifted. The alarm light illuminated on the control panel; the audible alarm was not heard during the test. The panel door is corroded at the hinge. Rating: needs attention. Recommendation: have the audible alarm and panel checked by a qualified electrician." Every clause traces to the notes, and the recommendation is a referral, not a diagnosis. If you give the model your approved library, it should pick "refer electrical component to qualified electrician" and not invent a fix.

This case also shows why partial results need their own rating option. Use a small fixed set, such as Satisfactory, Needs attention, Monitor, Not inspected and Unable to determine. "Unable to determine" is not a failure; it is an honest statement that protects both the inspector and the owner, and it is the one most likely to be softened into "satisfactory" by a model trying to be agreeable. Put the definitions of each rating in the prompt, in your own words, so the model never has to choose one.

A plain-language summary for the owner

Owners rarely read page six. Once the full report is signed, you can ask the model for a half-page summary that the owner will read, and then review it as carefully as the report itself. The summary should repeat the ratings, not reinterpret them, and should never be more reassuring than the full document.

From the signed findings below, write a half-page summary for the
property owner in plain English. Include: one sentence on what was
inspected; a list of the items rated needs attention, each in one line
with the recommended action; a sentence on what could not be inspected
and why; and a reminder that this is a snapshot on the day. Do not
add, soften or strengthen any finding. Do not use the words "fine",
"healthy" or "no problems". Use [owner first name] as a placeholder.

[paste signed findings here]

Banning the words "fine", "healthy" and "no problems" is a deliberate trick. Models reach for them when summarising, and an owner who reads "healthy system" will remember it long after they forget the caveats. Compare the summary against the signed findings with the same line-by-line check you used on the report. If a single reassurance has crept in, remove it.

A before and after shows the difference. The model's first attempt: "Overall your system looks to be in reasonable shape with a few small things to keep an eye on." The edited version: "Two items need attention: the cracked tank lid, and standing water in one outlet of the distribution box. One area of the drainfield should be monitored. The outlet baffle could not be inspected." The second is shorter, tells the owner what to do and says what is unknown.

Nine checks before the inspector signs

CheckHow to test it
Every number came from the notesHighlight each figure in the draft and find its source line
Every NR or NA appears as "not recorded" or "not accessible"Search the draft for those fields; check none became a statement
Ratings are the inspector's ownCompare each rating with the notes sheet, not with the model's summary
Inferred causes use "may"Search for "is caused by", "due to" and "failing"; reword or justify
Recommendations come from the approved library or were written by handMatch each against the library; no repair costs unless you supplied them
Photos match the right component and captionCheck numbers against the photo list
Scope and limitations are present and unchangedThe fixed text is intact; limits section lists what could not be reached
No owner name or address leaked into a place it should not beCheck header, footers and filenames
The inspector signed and dated after reviewingSignature added by a person, never by the tool

A quick blind test builds trust in the process. Take one old report you wrote by hand, feed the original notes through your prompt, and compare the two. Anything the model changed, added or lost shows you what the prompt still misses. Repeat after any prompt change.

What it saves, as a quick sum

Say an inspector does six inspections a week and spends about 50 minutes writing each one. If the draft takes about 15 minutes of tidying notes and checking, plus 10 minutes of review, the report costs 25 minutes, saving about 25 minutes a job and roughly two and a half hours a week. Those figures are illustrative; time your own next five reports before and after. If review time starts creeping below 10 minutes, ask whether you are really checking numbers or only reading.

Getting the report into a system the office can use

Once drafts are working, the question is where they live. A report that exists as a loose document is hard to search, hard to follow up and easy to lose. Look at where you keep pumping logs and reminders, because the inspection date and findings should feed the next visit. The questions to ask a vendor about that are in septic business software: what to ask for pumping logs, reminders and AI. A clear report is also a natural point to offer an annual maintenance agreement, and septic service agreements and renewals covers how to word that offer.

Comparing this visit with the last one

Where you inspect the same property more than once, the previous signed report is the most useful input you have. Paste the earlier findings next to the new notes and ask the model to list only what has changed: a scum measurement that has grown, a lid that has since been replaced, a wet patch that has spread or gone. Tell it to say "no change recorded" only where both sets of notes contain the same observation, and to flag anything present in one and absent from the other as "not comparable". The owner gets a trend rather than a snapshot, and you get a quick way to spot something you would otherwise forget to follow up. Treat the comparison like any other draft: the inspector confirms each change against the notes before it goes into the report.

Privacy and where the notes go

A report contains an address, an owner's name and sometimes remarks about their health or finances if they were volunteered. Keep those out of prompts. Use "the property" and a job number in the notes you paste, and add the name and address when you assemble the final document. If several people draft reports, use a business plan that does not train on your content by default, rather than individual free accounts, and check the vendor's current settings page rather than assuming.

Four weeks from note sheet to a tested routine

Week one: settle your note layout and write the limits text once. Week two: write your recommendation library, even if it has only ten lines. Week three: draft five reports through the process with the notes beside you and time each. Week four: do the blind test on an old report, adjust the prompt and retire any step that never caught an error. At the end you will know whether the saving is real for your volume of work, and you will have a fixed structure that any new inspector can follow.

Questions septic inspectors ask about AI-written reports

Can AI decide whether a septic system passes?

No. A pass, fail or needs-attention rating is the inspector's professional judgement, based on what they saw and the standard they work to. AI can arrange your notes and suggest wording, but the rating, the recommendations and the signature belong to the person who inspected.

Can I upload site photos to the AI?

You can, but it should only help with captions and organisation. It cannot tell you whether a baffle is damaged or an effluent level is normal from a photo. Strip location data and anything that identifies the property or owner if you are using a general chatbot, and caption from your notes rather than from the picture.

How do I stop the AI inventing measurements?

Tell it to use only values present in the notes and to write NOT RECORDED wherever a field is empty. Then compare every number in the draft with your notes before signing. A missing depth turning into a plausible figure is one of the most damaging errors a draft can contain.

Does the report need a disclaimer or scope statement?

A clear scope and limitations section is good practice: what was inspected, what could not be accessed, and that conditions can change after the visit. Wording and any legal requirements vary, so have it written or reviewed by an adviser who knows your rules, then reuse it as fixed text.

Further reads

Sources: general inspection-reporting practice and the AI vendors' published privacy settings. The inspection standard you work to, and any licensing or reporting rules, should come from your own regulator or professional body.

Want field notes turned into reports without the evenings?

On a 1:1 call we'll look at how your notes, photos and reports flow today, and set up a draft-and-sign routine that gives you the time back but leaves your judgement alone.

Book a 1:1 call with me