Draw inspection fraud: 6 ways progress photos lie

This post in 30 seconds.

• The problem: lenders release draw funds against photos nobody can place in time or space. An emailed JPEG proves almost nothing about the site today.
• The taxonomy: progress photos lie in six distinct ways, from recycled shots to AI edits, and each one defeats a different naive check.
• Who this is for: credit risk managers and draw administrators at construction and renovation lenders who fund some or all draws on borrower-submitted photos.

In 2016, a Vermont developer borrowed over $210,000 to build a house on a lot in Swanton. The loan agreement required photos of construction progress before each payment, so he sent them. The photos showed other people's buildings, and in one case he pasted a house and a garage from two different lots into a single frame. The financed property was, and remains, a vacant lot. He was sentenced to 10 months in federal prison in December 2020, four years after the money started moving (DOJ, District of Vermont).

That case is draw inspection fraud in its plainest form. The money leaves on the strength of an image, and the image answers to nobody. This post walks through the six ways progress photos misstate a site, how each one slips past the checks most draw teams run today, and which controls actually catch them.

If you administer draws at a construction lender, you already know the uncomfortable part: on small-balance and renovation loans, a site visit per draw rarely pencils, so photos carry weight they were never designed to carry. Venta Capture, a product of VentaVid, was built for exactly this gap, and a 15-minute demo will show you what verified capture looks like mapped to your own draw workflow.

Why a draw photo is weak evidence

An emailed photo arrives with no capture context: no proof of when it was taken, where, on what device, or whether those pixels ever saw the site at all. That is the root problem every fraud type below exploits.

Most draw teams run some version of the same three checks:

• Ask for photos per line item. Reasonable, but the request says nothing about where the photos come from. A folder of old files satisfies it as easily as a phone on site.
• Glance at the metadata. EXIF fields (date, GPS, device) are editable with free tools in under a minute. Worse, email compression and messaging apps strip or rewrite metadata in transit, so even honest photos often arrive blank.
• Compare against last month's draw package. Useful for catching the lazy cases. Useless against a borrower who photographs a different unit, a different elevation, or a neighbour's site.

The timestamp inside a photo file deserves its own warning. The device sets it, not an independent party, which means the person holding the device sets it. A draw inspection exists precisely because self-reported progress needs an independent cross-check, and a self-dated photo is still self-reported progress.

And the pressure on this weak point is growing. Deloitte's Center for Financial Services projects that generative AI could push US fraud losses to around $40 billion by 2027, up from $12.3 billion in 2023 (Deloitte). Progress photos are a soft target inside that number: high dollar value per image, low scrutiny per image.

For the general mechanics of checking any customer-submitted image, see our guide to photo verification for customer-submitted photos. The rest of this post applies those mechanics to the draw file specifically.

The six ways progress photos lie

Every type below defeats a different naive check, which is why a single countermeasure never closes the gap. Each case cited here is a real prosecution or indictment with a linked source; none of it is hypothetical.

1. Recycled or re-dated photos

The borrower resubmits photos from an earlier phase of the same project, or from a previous project entirely, with the file dates freshened up. Framing drawn at draw two reappears at draw four. The drywall photos from last year's flip document this year's.

Why the naive checks miss it: re-dated EXIF passes a metadata glance, and photos from a different project won't match anything in this loan's file. Unless the same reviewer handled both projects and remembers the shots, nothing connects them.

What catches it: reuse detection. A system that fingerprints every submitted image can flag that a photo closely resembles an earlier submission, or that the same photo was submitted before, with a click-through to the earlier case. The match does the remembering, not the reviewer.

2. Borrowed-site photos

The Fairbanks case above is the template: photos of structures the borrower does not own, presented as the financed property. Sometimes it is the neighbour's build at a flattering angle. In the Vermont case it went further, with a house and garage from two separate lots composited into one image.

Why the naive checks miss it: a lender 300 miles away has no mental picture of the site. If the photo shows a plausible house at a plausible stage, it reads as progress. Metadata, even when intact, gets ignored when the picture itself looks right.

What catches it: location and time signals captured at the moment of recording, not read from an editable file afterwards. When capture happens inside a controlled session, GPS and time come from the session itself, and a photo taken somewhere else simply does not fit.

3. Staged partial work shot to look complete

The work is real but thin. One finished wall shot from four angles stands in for a finished room. Flooring laid in the two metres nearest the door reads as a floored unit. Materials stacked on site get photographed to imply they are installed.

Why the naive checks miss it: every individual photo is genuine. Metadata is clean, the site is the right one, and nothing in the file contradicts the claim. The lie lives in what was not photographed.

What catches it: guided capture. When the lender defines the shot list (every elevation, every room, wide shot then detail, in a fixed order with example photos), the borrower cannot choose the flattering angles. Gaps become visible because the flow records that step 7 of 12 was skipped, not silently absent.

4. Inflated percent-complete claims

The photos are honest and the arithmetic is not. The draw request claims 75 percent on a line item the site shows at 45. In the Kansas City case, a developer drew $652,784.34 against costs that were never incurred, backed by fabricated expense documentation, and was sentenced to 21 months (DOJ, Western District of Missouri). An Indiana indictment alleges a Muncie contractor collected around $188,000 through nine forged lien waivers for work that was never performed (Coleman Report).

Why the naive checks miss it: percent complete is a judgment call, and a few photos give a reviewer almost nothing to judge with. One inspector reviewing a project manager's submissions put it bluntly on a professional forum: beyond the obvious, there was no real way to verify the percentages.

What catches it: structured capture tied to line items. When each budget line has its own required photos and questions, the reviewer compares a claim against evidence for that specific line, not against a general impression of the site. The paperwork fraud (fake invoices, forged waivers) still needs its own controls, but the visual record stops being the accomplice.

5. Photos of screens or printouts

The borrower photographs a rendering on a monitor, an old photo on a tablet, or a printed brochure page. The resulting file is a genuine, freshly taken photo with clean metadata, which is exactly what makes it dangerous.

Why the naive checks miss it: every date, device and location field is real, because the photo really was taken just now. It just wasn't taken of the site.

What catches it: screen detection and session context. Image analysis can recognise the artefacts of a photographed screen or print and record it as a fact on the item. A session code displayed in frame during capture adds a second barrier: a screen playing old footage cannot contain a code that was issued seconds ago.

6. AI-generated or AI-edited progress photos

The newest entry, and the cheapest. Generative tools can produce a plausible "after" image of a renovated kitchen from a text prompt, or edit a real photo of the actual site so that unfinished work looks finished. The Vermont composite took manual effort in 2016; the same edit is now a one-line instruction.

Why the naive checks miss it: there is nothing to cross-reference. The image is unique, so reverse-search finds nothing. The site is correct, so location checks pass if metadata is spoofed to match. At phone-screen size, current fakes are hard to spot by eye.

What catches it: provenance over inspection. You will not reliably out-stare a generated image, but you can require that images are recorded now, on this device, inside a controlled session, and sealed on receipt. An AI edit has to enter the file somewhere, and material that arrives as an upload instead of a session recording gets labelled and checked for editing traces rather than trusted by default.

What verifiable capture changes

The common thread in all six types is that the lender has no control over how the image was made, so every control has to be bolted on after the fact. Verifiable capture moves the control to the moment of creation.

Here is the model, using Venta Capture as the concrete example. The lender builds the capture flow once: which elevations, which rooms, which line items, in what order, with example photos and plain-language instructions per step. At each draw, the borrower or site contact receives a secure, personal capture link and opens it on their phone. No app, no account. They film and photograph the site with the flow steering them, answer the attached questions, and submit.

What arrives is not a folder of JPEGs. It is a structured evidence package, and the difference shows against each fraud type:

A few mechanics behind that table deserve one line each:

• Recorded in the session, on this device. Media captured through the in-session camera carries that provenance, confirmed server-side. The receipt time is verified by the server, not read from the file.
• Sealed on receipt. Each submission carries a SHA-256 fingerprint, a visible stamp on session-recorded media, and an independent timestamp from an external timestamping authority (RFC 3161), so the question "when did this evidence exist" is answered by a party outside both lender and borrower.
• Uploads are allowed, honestly labelled. A borrower who already has a photo (say, storm damage from last night) can upload it where the flow permits. It arrives marked "Provenance: Not verified" and is checked on metadata, editing traces, and whether it already exists on the internet, with each finding carrying its own caveat.
• Signals, not verdicts. Findings surface as green, amber or red observations for a human reviewer. Multiple recording attempts, for instance, are normal behaviour in themselves; the system raises a flag, a person decides.

Now the money question: is this worth adding to your draw cycle? Walk one illustrative example. Say you fund 200 draws a month at an average of $35,000, and photo-only review lets one materially misstated draw per quarter through at an average overadvance of $20,000. That is $80,000 a year of loan proceeds ahead of collateral, before workout costs, on assumptions you should replace with your own portfolio's numbers. Against that, a guided capture request costs the borrower about ten minutes per draw and costs your team nothing to review but the time they already spend on photos, now spent on evidence that can be placed.

You can test the flow on your own draw checklist before the next funding cycle: start for free and build your first capture flow the same day.

What verification does not prove

A credit committee should trust this control for what it proves, and nobody should oversell what that is. Three limits matter, and they are worth stating as plainly as the fraud types.

First, a genuine recording can still show a staged situation. Verified capture proves the pixels were made now, here, on this device. It does not prove the plumbing behind the drywall exists. Harder to manipulate does not mean impossible to deceive.

Second, the timestamp proves receipt, not the event. "The system received this submission at 14:02 on Tuesday" is a strong, independently checkable fact. "The concrete was poured Tuesday" is a conclusion a human draws from the whole file.

Third, this is remote-first, not remote-only. High-value draws, final draws, and anything flagged amber or red still deserve boots on the ground. The point of verifiable capture is not to end site visits; it is to make sure the draws you fund without one rest on evidence rather than trust, and to tell you which files need the visit. The end-to-end process question (schedules, inspectors, disbursement mechanics) is its own topic, and we cover it separately in our guide to construction draw inspection.

I read a year's worth of DOJ construction-fraud press releases while researching this post, and the pattern that repeats is not sophistication. It is distance: a lender who could not see the site, deciding on evidence the borrower controlled. For what makes submitted images hold up when a file is later challenged, our piece on inspection photo evidence standards goes deeper on the evidentiary side.

Frequently asked questions

What is draw inspection fraud?

Draw inspection fraud is misrepresenting construction progress to trigger a loan disbursement, most often through photos, invoices or lien waivers that overstate or fabricate work in place. Prosecuted forms include recycled photos, photos of other properties, fabricated costs and forged subcontractor waivers.

How common is fraud in construction draw requests?

No reliable prevalence statistic exists, because most inflated draws are resolved as billing corrections rather than prosecuted. The prosecuted cases run from a $210,000 single-house loan in Vermont to $33 million drawn against fake invoices in Texas, so the exposure scales with the book, not the loan size.

Can EXIF data prove when a progress photo was taken?

No. The device writes EXIF fields and free tools edit them in seconds, so EXIF is a hint, never proof. Email and messaging apps also strip or rewrite metadata in transit, which is why even honest photos often arrive with nothing to check.

Do AI-generated progress photos actually fool lenders?

At the size and speed a draw reviewer works, a current AI image of a renovated room is plausible enough to pass an eyeball check. The workable control is provenance (requiring images recorded in a controlled session and sealed on receipt) rather than training reviewers to spot fakes.

Does verified photo capture replace third-party draw inspections?

No, and it should not. It raises the evidence floor on every draw and surfaces the signals that tell you which draws justify the cost of an independent site visit, so inspection budget lands where the risk is.

What if the borrower has no mobile signal on site?

Venta Capture keeps working when the connection drops: every photo, video and answer is saved on the phone as it is made and sent automatically once coverage returns. The case records how each file reached the server, so a rural site is a circumstance in the record, not a red flag against the borrower.

Talk to us

If your draw file still rests on photos that arrive by email, every type in this taxonomy is an open door, and you will usually find out the way the Vermont lender did: years later, from a prosecutor. Changing that does not require new inspectors on every draw. It requires that the photos you already collect become evidence you can place in time, space and session.

What it takes to start.

• Free plan, no credit card. Build a draw capture flow and run a test submission today.
• Live in 10 minutes. Your borrowers need no app and no account, just the link.
• Stuck? Book a free setup call and we build your first flow together.

A real person from the VentaVid team builds your first draw flow with you on a free setup call.

Book a demo to see a draw workflow end to end, or start for free and send your first capture link before the next draw cycle.