Home Energy Audit Report: Turn Findings Into a Client Report With AI

Lead with a one-page summary of the three actions that matter most, then give a findings table with each problem, its evidence, the fix, a cost range, a saving range with its assumptions and a priority. Write it in plain English, promise no savings, and trace every figure to your own measurements before the report leaves your desk.

AI can sort field notes into that table and rewrite it for a homeowner, but it must only use figures you supply. Give it a note that says "approx 100mm mineral wool" and a first draft will happily add a target thickness and a yearly saving that appear nowhere in your notes. Every house detail, price and saving in the example that follows is illustrative.

Follow me on Instagram@sagnikteaches

One house, one visit: what the auditor brings back

The example is a detached four-bedroom house from the 1970s, extended once at the back, with a family of four who are home most evenings. The auditor spends a morning on site: a blower door test, thermal images, a look in the loft and at the heating, and a count of the lamps. She also asks for 12 months of energy bills, because estimates are much better when the billing history is available, and age of the home, climate and the way people live in it all change the result. What she brings back is a page of rushed notes, a photo log and the bills.

Connect on LinkedInSagnik Bhattacharya

JOB 0417. Detached, 4 bed, built 1970s, one extension rear (2000s). Family of 4, home most evenings.
Blower door: 9.8 ACH50 before sealing. Thermal: cold lines at loft hatch, 2 ext doors, kitchen corner near sink, rear ext wall low down.
Loft: approx 100mm mineral wool, patchy, gaps round tank (P-07). Walls: cavity, nothing found at 2 inspection holes (P-08, P-09).
Boiler: gas, 17 yrs per data plate (P-10), 1 room stat, 3 of 9 rads no TRV. Cylinder jacket thin.
North bed: dark patch ~30cm at corner of ext wall, window wet. Possibly mould, didnt test.
No CO alarm anywhere (checked landing, hall, boiler cupboard).
2 windows misted between panes (P-11). 14 halogen downlights kitchen + lounge. Unused chimney in lounge, open.
Bills (12 mo): gas 14,200 kWh, elec 3,900 kWh. Latest prices gas 0.08, elec 0.22 per kWh.

On the prices in the last line, the gas bill comes to about $1,136 a year (14,200 x $0.08) and electricity to about $858 (3,900 x $0.22). Those two numbers matter later, because they set the ceiling on what any heating measure could save.

Subscribe on YouTube@sagnikteaches

Field notes into candidate findings, marked unverified

The first AI job is the dullest and the most useful: turn the notes into one row per issue, without judging, ranking or costing anything. Use a job code, not the client's name or address. The prompt forbids the things AI most wants to add.

You are helping a home energy auditor turn field notes into candidate findings. Use ONLY the notes below. Do not add any number, standard, recommended thickness, cost, saving or health claim that is not in the notes.

For each distinct issue write one row: ID | Location | What was observed (my words, no judgement) | Evidence (measurement, photo ID or "none") | Suggested action (the type of work only) | Status: UNVERIFIED.
If a note is ambiguous or incomplete, write [ASK AUDITOR: question] in the row instead of guessing. Do not rank, cost or estimate savings.

NOTES:
[paste notes here]

A first pass is usually tidy and quick, and it still contains problems that read well. Four from the example follow, each with the fix. Check every row against the photos and the measurement log before you let it into the next step.

The pattern across all four is the same. The AI supplies a plausible fact where the notes are quiet, and it turns an observation into a judgement. The "ASK AUDITOR" instruction helps, but it does not remove the need to check every row. Keep the status column on UNVERIFIED until you have ticked the row against its evidence, and treat that column as the gate to the next step.

The findings table for this house

Once the rows are verified, add what only you can add: the priority, the cost range from your own price data or recent quotes, and the saving range from your model. This is the table for the example. It has two kinds of "not applicable", which are safety items that have no payback and a damp issue that cannot be costed until its cause is known.

Do the payback sums in a spreadsheet, not in the chat. The lowest payback is the low cost divided by the high saving, and the highest is the high cost divided by the low saving. For the loft top-up that is $900 divided by $200, which is 4.5 years, and $1,300 divided by $90, which is 14.4 years. Wide ranges look unhelpful until you picture the alternative, which is a single figure that looks precise and is wrong. A range says honestly how much the answer depends on weather, heating habits and the final price. If the plan involves a heat pump, the quote side of that conversation is in the heat pump quote template tutorial.

Priority order: safety, quick wins, fabric, systems, comfort

A homeowner reads a report from the top and acts on the first thing that sounds sensible, so the order of the table is a recommendation in itself. The order that works for most homes follows five rules.

1. Anything affecting safety or health comes first, whatever its payback. Alarms and a damp patch are not investments, so they carry no payback and still sit at the top.
2. Cheap measures with short paybacks come second. Draught-proofing and lamps are quick to do and show quickly, which builds trust in the rest of the report.
3. Fabric jobs come before systems. Insulation and air sealing change how much heat the home needs, so it is better to size or upgrade the heating after those, not before.
4. Systems follow, with the age and condition shown as evidence. A boiler's age is a fact; its efficiency is only a fact if you tested or looked it up.
5. Comfort-only items go last and say so. Replacing two sealed window units may make a room quieter and warmer, but the savings are small, and the report should not dress it up as an energy measure.

The ordering also manages interactions. Sealing draughts in a house that already has a damp problem can make the damp worse if ventilation is not adequate, which is why F2 sits above F3 and says "find the cause before treating or sealing". Whether to add ventilation advice is a judgement for the auditor, not for the AI. The same logic is why savings overlap: once the loft and walls are insulated, the heating controls have less heat to save. For writing up the contractor's side of insulation with the same restraint, see the insulation proposals tutorial.

Savings as ranges: wording that promises nothing

The five gas measures in the example have saving ranges of $340 to $770 when you add the lows and the highs, which is 30 to 68 per cent of the $1,136 gas bill. That is a useful sanity check, and it fails as an answer. The same house also uses gas for hot water, and the measures overlap, so a package of all five cannot save the sum of the parts. Suppose the auditor's modelling software puts the whole package at $300 to $520 a year in gas, which is 26 to 46 per cent of the bill. That figure may go in the report as a modelled range. The per-measure ranges stay in the table as what each does on its own, and the report says plainly that they do not add up.

Put an assumptions block beside every savings figure so a reader can see where it came from. For the example: energy prices of $0.08 a kWh for gas and $0.22 for electricity, taken from the latest bills; 12 months of bills as the baseline (14,200 kWh of gas and 3,900 kWh of electricity); a heating pattern as the family described it; typical weather; work done to a good standard; costs are ranges from recent quotes, not quotes; savings overlap and are not additive.

Then use wording that cannot be read as a promise. These pairs show the swap.

Keep a "words to avoid" list in the prompt for the summary: will save, guaranteed, pays for itself, cuts your bills by, and "up to" with no basis. An AI drafting from your table will reach for these because they sound helpful.

The plain-English summary: draft, faults and the edited version

Only the verified table goes into this prompt, so every number the AI can use is one you have already checked. The summary is the part most homeowners read, so it gets the most care.

Write a one-page summary of this energy audit for a homeowner who does not know building terms. Use ONLY the table below. Do not use any number that is not in the table, do not add totals, and do not promise savings: use the words "estimate", "range" and "depends". Structure: (1) the main weak points in two sentences, (2) the first three actions in order, one sentence of reason each, (3) one sentence saying every saving is a range that depends on weather, heating habits and the price when quoted, (4) one sentence saying the savings overlap and have not been added together. Short sentences. No jargon without a plain explanation. Avoid: will save, guaranteed, pays for itself, cuts your bills by.

TABLE:
[paste the verified findings table and the assumptions block]

A first draft typically makes three claims the table does not support. It says the measures "will cut your bills by about a third", which is a number that appears nowhere. It says "your boiler is inefficient and should be replaced", which the table contradicts, since efficiency was never tested. And it says "loft insulation pays back in 4 years", a single figure where the table has a range of 4.5 to 14.4. Delete or correct all three. The edited version reads as follows.

The main weak points we found are draughts, a thin and patchy layer of loft insulation, and walls with no insulation inside them. Two safety jobs come first: fit carbon monoxide alarms near the boiler and on the landing, and have the damp patch in the north bedroom looked at before anyone paints over it. Next come the cheap jobs, draught-proofing and LED lamps, which have the shortest payback ranges in the table, about 1 to 6 years. The loft top-up and the wall insulation cost more and take longer to pay back. Every saving in this report is an estimate and a range, and it depends on the weather, how you heat the house and what the work costs when quoted. We have not added the savings together, because the measures overlap.

Read it as the homeowner would and ask whether they could say what to do first. A tutorial on checking readability is in how to check AI content for plain English, and it is worth running on the summary before it goes out. Better still, hand the page to someone outside the trade and ask for the three actions in their own words.

Measurements the homeowner cannot see, in one sentence each

A report that says "9.8 ACH50" to a family has told them nothing. Give each measurement a plain line that says what it is and what it means for them, and a line on what not to claim. Four terms come up in most audits.

Thermal images depend on conditions: reflections, sunshine and wind can all change what the camera reads, so state when and under what conditions the images were taken. The thermal imaging survey tutorial covers turning scans into findings in more depth. When you ask an AI to explain a measurement, give it your plain line from this table and tell it to rephrase for reading level only, not to add facts.

The figure audit: trace every number before the report goes out

This is the step that earns the reader's trust and protects you. Take the finished report and ask an AI for an inventory, not an opinion: "List every number in this report as a table: Number | The full sentence it appears in | Page or section. Include costs, savings, paybacks, dates, measurements and counts. Do not judge whether any number is correct." Then tick each row against the figures table and the measurement log yourself. The AI is good at finding numbers and cannot check them.

On the example report, the inventory held 63 numbers, and five failed. Each failure is a type you will recognise.

Five slips in 63 numbers is about 8 per cent, and every one would have gone to a client as fact. Check the arithmetic as well: redo each payback in the spreadsheet, confirm that units agree (kWh against dollars, years against months), and make sure each finding ID mentioned in the summary exists in the table with the same wording. Add a line at the end of the report that names who checked the figures and when. If you want to go further with invented numbers, how to catch made-up figures in AI-drafted proposals covers more checks.

Photos, captions and the limits statement

Every finding that has a photo should name it, and every photo should carry a caption that says what the reader is looking at. An AI can draft captions from your photo log. For example, from the log entry "P-07 loft, NE, insulation depth ~100mm, gap round tank" it can write "P-07. Loft, looking north-east: insulation about 100 mm deep, with a gap around the cold water tank." You confirm each caption against its photo, because an assistant that cannot see the photo will happily describe a good one.

Protect the household in the photos. Crop out family pictures and letters, use the job code in file names, and send nothing to an AI tool that shows a name or address. Use a business or team plan where the vendor says content is not used for training by default, or switch off the training setting on a personal plan.

Close the report with a limits statement. It protects you and tells the reader what the report is. Adapt this to what you actually did.

This report is based on a visit on [date] and the measurements and photographs listed in the appendix. We did not open walls or floors, test the boiler's combustion [if true], or check electrical or structural safety. Estimated costs are ranges drawn from [your source] and are not quotes. Estimated savings are ranges based on your bills and the assumptions listed, and they are not guarantees. Figures checked by [name] on [date].

If you keep a library of standard wording for repeated findings, the tutorial on building a report comment library shows how to build one so the AI starts from your approved phrases rather than inventing new ones each time.

Report time, and four lessons from the first run

The hours for this report, by hand and with AI drafting, break down as follows. These figures are illustrative, so time your next two reports before you draw conclusions.

Three hours saved on one report, and the figure audit got longer on purpose. That is the pattern to expect, and it comes with four lessons.

• Use AI for two jobs only: turning notes into a draft table, and turning the verified table into plain English. Every number comes from you.
• Keep a figures table as the single source. The AI only ever sees verified rows, and the audit compares the report to that table.
• Ranges and assumptions beat precision. A range with its assumptions is accurate and defensible; a precise figure is neither.
• Test the report on a real reader. Ask someone outside the trade for the three actions and what to do first. Count the clarification emails you get in the first month and fix the wording that causes them.

Questions auditors ask about AI-assisted reports

Should an energy audit report include savings figures at all?

Yes, as ranges with the assumptions beside them, because homeowners need some sense of scale to decide what to do first. Never as a promise. State the energy prices and the bills you used, say the measures overlap so the figures do not add up, and say that real bills depend on weather and how the home is heated.

Can AI estimate savings from photos or notes?

No. Savings come from your own model or calculation tool, using the home's bills. An assistant given photos or notes will still produce a confident-looking number, and it will be invented. Use AI to format and explain figures you have already calculated, and never to produce them.

How long should the plain-English summary be?

About one page: the three actions that matter most, in order, with the reason for each in a sentence, plus one line on what is not covered. If the homeowner cannot say what to do first after reading it once, it is too long or too vague, so test it on someone outside the trade.

Is it safe to put a homeowner's details into an AI tool?

Keep the personal details out. Use a job code instead of a name or address, crop photos so family pictures and letters do not show, and use a business or team plan where the vendor says content is not used for training by default. Check the setting on any personal plan before you paste anything in.

Further reads

• Damp Survey Report: Write Findings and Quote Lines With AI — Write damp findings and quote lines in plain terms.
• Can ChatGPT Write Home Inspection Report Narratives? What to Check — What to check before AI-written report narratives go out.
• Heat Pump Installer Software: Surveys, Heat Loss and Quotes, AI Compared — Surveys, heat loss and quotes for the installer's side.
• Insulation Leads: What They Cost and How AI Qualifies Them — How insulation leads are qualified once reports are sent.
• Mould Inspection Report Template: Findings, Photos and Limits With AI — Findings, photos and limits wording for mould reports.
• Structural Engineer Report Template: A Plain-English Summary You Can Check — A plain-English summary you can check, from another report type.
• EPC Assessor Software: What to Check First, and Where AI Helps — Check scheme approval, approved-list status and update speed before you pay, then use AI for client emails and wording, never for lodgement.
• Insulation Quote Template: Material, Thickness and Area Lines With AI — A line-item insulation quote with thickness, area and exclusion wording, plus a prompt that flags missing measurements instead of inventing them.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: standard building-science definitions of blower door results (air changes per hour at 50 pascals), thermal imaging, and R- and U-values; OpenAI and Anthropic business-plan data-use pages. House details, prices, costs and savings in the example are illustrative.