Ask six fixed questions and request four photos per problem area. Have AI describe only what is visible and what is missing, never a cause, severity or cost. Apply your own engineer-referral rule before booking the free inspection, and track cost per sold job against a ceiling you set.
The reason to screen is not to turn enquiries away. It is to send the right person with the right questions answered, to catch the enquiry that should go to an engineer before anyone quotes a fix, and to know what each lead really costs. The AI is a note-taker with good eyesight. It does not know why a wall is cracked, and if it ever sounds sure, delete that sentence. A structural diagnosis is an engineer's judgement, made on site with a professional's liability behind it, so no assistant in your process, whichever vendor makes it, should be set up to give one. On the engineer's side of this, what is safe and what never is for structural engineers draws the same line.
Six questions that sort an enquiry before a photo arrives
Put these on your web form, or ask them in the first text. They take under two minutes to answer, and each one changes what you do next. Keep the wording fixed so the answers can be compared across leads.
| Question | Why you ask it | Sample answer |
|---|---|---|
| 1. Where is the problem: basement wall, floor, crawl space, outside wall, ceiling? | Tells you what to bring and what the photos must show. | Basement, north wall |
| 2. When did you first notice it, and has it changed since? | Growth matters more than size. A crack that has not moved in years is a different enquiry. | About eight months ago; longer since spring |
| 3. What else have you noticed: doors or windows sticking, gaps at frames, sloping floors, damp? | Several signs together change the referral rule. | Front door sticks in wet weather |
| 4. Anything recent nearby: heavy rain or drought, tree removal, a leak, digging next door? | Context for the estimator, not a diagnosis. | Tree removed last autumn |
| 5. Do you own the property, and is this linked to a sale or a survey report? | Who can authorise work, and whether a deadline exists. | Owner; no sale |
| 6. Has anyone else looked at it, and do you have a report or quote? | Separates a comparison shopper from a customer with a report to act on. | No one yet |
Question 6 is easy to leave out, and it pays for itself. A customer holding an engineer's report wants a quote against a specification. A customer with nothing wants an opinion first. Those are two different conversations, and your reply should say which one you are about to have.
Four photos per problem area, requested in a text people can follow
Customers send one wide shot taken from the stairs, or a blurred close-up with nothing for scale. Give them a script that produces usable images. Ask for four photos per area, for at most three areas, taken in daylight with the flash off.
- Wide: the whole wall or floor from a few steps back, showing both ends and the nearest corner.
- Mid: the crack or area, with a coin or ruler beside it for scale.
- Close: the widest part of the crack, coin or ruler in frame, held square to the wall.
- Context: the same spot from outside, or a door or window that sticks.
Hi [first name], thanks for getting in touch. To send the right person with the right kit, could you text us four photos of each problem area (up to three areas)? 1. The whole wall or floor from a few steps back. 2. The crack with a coin beside it. 3. A close-up of the widest part with the coin in frame. 4. The same spot from outside, or any door or window that sticks. Please do not fill or paint over anything before we visit, and keep people and papers out of frame. We will reply within the hour with the next step.
One optional step that costs the customer thirty seconds: ask them to draw a short pencil line across each end of the crack and write today's date beside it. When your estimator arrives, any growth is visible without anyone relying on memory.
What AI may say about a photo: describe, flag, never diagnose
Paste the six answers and the photos into an AI assistant that can read images, and give it a job with tight edges. It describes what is in view, notes what the customer reported, ticks the flags from your own list, and says what is missing. It does not name a cause, rate a severity, suggest a repair method or estimate a price.
You are a screening assistant for a foundation repair company. You do not diagnose.
You will receive a customer's answers to six intake questions and up to twelve photos.
Output, in this order:
1. Photos received: how many, and which of wide / mid / close / context are present per area.
2. Scale: is a coin or ruler visible in the mid and close photos? yes/no per area.
3. Visible features: neutral description only (where it is, direction it runs, whether it follows
joints, approximate length relative to the frame). No causes, no words like settlement,
shrinkage, minor, serious, cosmetic or safe.
4. Reported by the customer: list each item and mark it REPORTED, not confirmed.
5. Referral-rule flags: for each item on the list below, say TRIGGERED (with the evidence and
whether it came from a photo or from the customer), NOT TRIGGERED, or CANNOT TELL.
6. Missing information: exactly what to ask for next.
7. Suggested route: one of A, B, C or D from the table below, marked FOR OFFICE DECISION.
If a photo is blurred or ambiguous, say so. Never guess.
REFERRAL RULE: [paste your rule table here]
ANSWERS AND PHOTOS: [paste]A trimmed sample response for the basement enquiry above: "Photos: three received for the north wall, no outside photo. Scale: coin visible in the close photo only. Visible features: a crack running diagonally from the lower corner of the window opening towards the floor, following the mortar joints in a stepped line; length about two thirds of the frame height. Reported: crack longer since spring (REPORTED); front door sticks in wet weather (REPORTED). Flags: growth TRIGGERED (customer); door sticking TRIGGERED (customer); horizontal crack CANNOT TELL; bowing CANNOT TELL (no photo along the wall). Missing: outside photo, a photo from one end of the wall looking along it. Route: B, for office decision."
The fix I would make to that draft is one phrase. "Following the mortar joints in a stepped line" is a description, and the estimator will read it correctly. If the AI had written "stair-step crack, typical of settlement", that is an interpretation, and the prompt forbids it. After a few weeks, scan drafts for the banned words and reject any that contain them.
Write the engineer-referral rule before the first lead arrives
This is the part only you and your engineer can write. Sit down with the engineer you refer to, agree the triggers and the wording, and put the result in one table the office can follow without judgement calls. The thresholds belong to them, not to me and not to an AI. A starting structure:
| Route | When (agree the real triggers with your engineer) | What happens |
|---|---|---|
| A. Standard visit | Few signs, stable according to the customer, photos usable, nothing in the engineer-first list. | Book the free inspection; the estimator brings the photos and notes. |
| B. Visit, and recommend an engineer | Two or more signs together, growth reported, or the customer is worried. For example, a growing crack plus sticking doors. | Book the visit to measure and document; tell the customer an independent engineer is recommended before any repair is quoted. |
| C. Engineer first | Triggers you and your engineer list as engineer-first, such as a wall that appears to bow or lean, a horizontal crack in a below-ground wall, or sudden new movement after excavation nearby. | Do not quote from photos. Give the customer the engineer-first message and offer to visit only to document. |
| D. Not a foundation enquiry | Surface cracking in paint or plaster, a different trade, or a service area you do not cover. | Send the polite redirect and invite a re-send if anything changes. |
Two principles keep the table honest. When the AI says CANNOT TELL on a trigger, treat it as triggered until a person has looked, because the cost of an unnecessary visit is a few hours and the cost of a missed warning is much larger. And never use the table to talk someone out of an inspection they are anxious about. If the customer insists, book the visit.
Four reply scripts: book, ask again, engineer first, not our job
Each route needs words that sound like a person and promise nothing about the cause. These drafts keep placeholders for the details. The AI can produce them from the route and the customer's answers, and you read before anything goes out.
Route A or B, booking: Hi [first name], thank you for the photos. They help us decide who to send and what to bring, but they cannot tell us what is behind the wall, so the free inspection is where we find that out. We can come on [day] between [time] and [time]. [If B: Because you mentioned the crack has grown and a door is sticking, we would also recommend an independent structural engineer looks before any repair is agreed, and we can explain why on the visit.] Reply YES to confirm.
Needs more photos: Hi [first name], thanks, the close-up is clear. We could not see the whole wall or the outside, so could you send one photo from a few steps back and one from outside at the same spot? If photos are difficult, tell us and we will book a visit without them.
Route C, engineer first: Hi [first name], thank you for sending these. From what you describe, we do not think anyone should quote a repair until a structural engineer has looked at the cause, and we would rather say so now than sell you the wrong fix. We can suggest how to find an independent engineer, and we are happy to visit and document what we see for them. Which would you prefer?
Route D, not our job: Hi [first name], the photos show what looks like surface cracking in the plaster or paint. A decorator or plasterer is the person to ask. If the crack grows, new ones appear or doors start sticking, send us fresh photos and we will have another look.
Notice the phrase "we would rather say so now than sell you the wrong fix" in the engineer-first script. It earns trust, and it is also the sentence a decent firm would want quoted back to it in a dispute.
Wiring it up: forms, photo folders and a manual first month
You do not need automation on day one, and I would not start there. For the first 20 leads, run the process by hand so you learn where it jams.
- One intake form. The six questions as required fields on your web form, with a photo upload that accepts several images. Add a reply-by-text option for customers who would rather message.
- One folder per lead. Name it with the lead reference and the date, and save every photo there. The AI step, the estimator and the customer record all point to it.
- One saved prompt. Keep the screening prompt and your referral table in a ChatGPT or Claude Project so every lead starts from the same instructions, and name one person as the owner of those instructions so nobody changes them on a busy afternoon.
- One person approves. Whoever is in the office reads the AI output, picks the route and sends the reply. Nothing leaves without a human reading it.
Phone enquiries need the same six questions, read as a call script. Ask them in order, type the answers straight into the lead record, and text the photo request while the customer is still on the line. A filled-in note from such a call: "Basement, east wall. Noticed about a year ago; no change since. Door and windows fine. Nothing recent nearby. Owner, no sale. No report. Photo request sent 10:42." That note takes a minute to write and gives the estimator everything the web form would have given. Once the manual month has run, Zapier or Make can create the lead folder and draft the first reply from a new form entry. Keep the AI step set to produce a draft for approval, not to send.
Cost per lead, per inspection and per sold job
Screening changes your costs only if you measure them. Here is one invented month at a company buying leads from a source that charges $85 each. Replace the figures with yours.
| Stage | Count | Cost or value |
|---|---|---|
| Leads bought | 40 | $3,400 at $85 |
| Valid (right service, owner, in range) | 29 | |
| Reached within 15 minutes | 22 | |
| Free inspections booked | 14 | $243 each in lead spend alone |
| Inspections held | 12 | $142 each (2.5 hours at $48, plus $22 vehicle) = $1,704 |
| Quotes issued | 9 | |
| Jobs sold | 3 | $28,200 at $9,400 average; gross profit $10,152 at 36% |
| All-in acquisition cost | $3,400 + $1,704 = $5,104 | |
| Cost per sold job | $1,701, which is 50% of the $3,384 gross profit per job |
Look first at the gap between the first two rows. Eleven of the 40 leads failed the most basic check (wrong service, outside your area, or a tenant with no owner involved), which is $935 of the month's spend. Whether any of that comes back depends on your source's rules, so ask it in writing which of those categories it credits and how long you have to claim. The screening note is your evidence: it records what the customer said and when, so file any claim the same week rather than reconstructing it from memory a month later.
Now set a ceiling in advance. Mine would start at 40% of gross profit per sold job, because the rest of that margin must pay for the van, the office and the insurance, but yours should come from your own overheads. 40% of $3,384 is $1,354. At $1,701 the source fails the ceiling, even though it brought in $10,152 of gross profit. That is the trap with bought leads: the month looks profitable until you subtract everything it took.
Here is what AI screening and faster replies could do in the same month, as assumptions to test and not as a promise. Screening removes three inspections that were never going to sell (two that the photos showed as cracking in paint or plaster, and a garden-wall crack on a structure you do not repair), which saves $426. Faster first replies bring in three extra inspections, and one of them sells. Inspections held are back at 12, jobs sold are four, and the cost per sold job drops to $1,276, which is 37.7% of gross profit and under the ceiling. The AI did not change the lead price. It changed what the same leads turned into. If the second assumption does not happen after 60 to 90 days, the answer is to negotiate the lead price, change the source or stop buying, and the sums above tell you which.
For a wider comparison of lead sources by cost per booked job, the handyman leads tutorial works through three platforms. Where a storm or event produces a rush of enquiries, the roofing storm leads tutorial adds a triage routine that carries over.
Two ways the screen fails, and how each one shows up
The flat photo that hides the problem. A customer sends a sharp close-up of a hairline crack in a basement wall, plus a wide shot taken square-on. Both are fine. The wall also leans in at the top, which no square-on photo can show. The screen scored it route D. A week later the customer sends a better photo to another firm. The fix is a fifth photo request for walls: stand at one end of the wall, hold the camera level with the surface, and photograph along it. Add one question, "Does the wall look straight when you sight along it?", and treat any hesitation as a trigger.
The reassurance that slipped through. The AI drafted the reply to a customer and wrote "this looks like a minor shrinkage crack, so there is likely nothing to worry about". You forbade the words in the prompt, but a reply-drafting step is a separate prompt, and nobody put the rule in it. The fix is structural: one policy block, pasted into every prompt that touches customers, that bans cause, severity, safety and cost language, and a last-step scan of any draft for words like minor, shrinkage, settlement, safe, serious and cosmetic. A person reads the draft anyway, but a scan catches what a tired person misses at 5pm.
Keeping home photos tidy, private and attached to the right lead
Customers are sending images of the inside of their homes, which can show faces, letters, screens and valuables. Ask them to keep people and papers out of frame, as the photo script does. Save the photos to the lead record the day they arrive, with the six answers, so the estimator opens one place and not three inboxes. Use the same job reference everywhere.
When you paste into an AI tool, remove names, phone numbers and the street address; the assistant needs the answers and the images, not the identity. Use a business plan where the vendor states it does not train on your content by default, check the model-training switch on any personal plan, and tell customers in your privacy notice that photos are reviewed with the help of software. Keep the AI output in the record labelled "screening note, not an assessment", so nobody later mistakes it for one.
Thirty days in: what to audit
After a month, check five numbers and one sample. The numbers: how many valid leads had photos within two hours, how many bookings became held inspections, how many held inspections became quotes, how many routes were B or C, and the cost per sold job against your ceiling. The sample is the important part: pull ten enquiries that were routed D or declined and have an estimator look at the photos without the AI note. If any of the ten should have been a visit, tighten the rule that let them through, and do it before another month of leads arrives.
A good sign is the office stopping to ask "what does the rule say?" instead of "what do you think?". A bad sign is estimators ignoring the screening note because it is long. If that happens, cut the AI output to the flags, the missing items and the route, and keep the rest in the record. A short screen that gets used beats a thorough one that does not. And if you want to see the same pattern on a trade with a very different risk profile, screening loft conversion leads by roof, head height and budget is a useful comparison.
Questions foundation repair owners ask about screening leads with AI
Can AI tell me whether a crack is serious from a photo?
No, and your process should make sure it never says so. A photo shows a surface at one moment; it does not show movement, soil, drainage or what is behind the wall. Use AI to describe what is visible, list what is missing and match the enquiry against your own referral rule. The assessment stays with a person who visits.
Should I refuse enquiries whose photos look cosmetic?
Not outright. Reply that the photos show surface cracking you would normally pass to a decorator or plasterer, ask the customer to re-send if it grows or other signs appear, and offer a visit if they are worried. Declining silently on a photo is the one move that can cost you both a good customer and your reputation.
What if the customer cannot take good photos?
Offer to book a short visit without them, or ask them to send what they have and add a phone call. Screening should speed up good enquiries, not block anyone who is not comfortable with a phone camera. Tag the lead as visit-first so nobody waits on photos that will never arrive.
Is it safe to put customers' home photos into an AI tool?
Treat them as personal data. Use a business plan, switch off model training where the setting exists, strip names and addresses from anything you paste, and ask customers not to include people or documents in frame. Keep the photos in your own job record, and keep the AI output there as a screening note.
Further reads
- Garage Door Leads: Qualify and Answer Them Faster With AI — The same lead-cost sums applied to garage door enquiries.
- Carpet Cleaning Leads: Buy Them or Reply Faster With AI? — Buy leads or reply faster: the same decision elsewhere.
- Insulation Leads: What They Cost and How AI Qualifies Them — Real price ranges for insulation leads by source, how to turn them into cost per booked survey, and an AI qualifying script that protects survey slots.
- Paving Leads: Answer Commercial and Residential Enquiries With AI — How a paving firm answers residential and commercial enquiries fast, books site visits and tests whether replying beats buying more leads.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.