On June 16, the Department of Justice charged an Indiana contractor with nine counts of bank fraud. Richard Turner allegedly submitted falsified lien waivers claiming five subcontractors had completed work they never performed, forging their signatures on documents nobody thought to verify. He stole $188,000 from a $785,225 construction loan, the kind of money that was supposed to pay plumbers and electricians and roofers for work on a house that someone was waiting to move into, and he spent it at casinos.
Turner didn't need AI. A pen and a printer were enough.
That should terrify you, because the people who come next will have tools he didn't, and those tools are already free, already trivial to operate, and already producing output that fools professionals whose entire job is spotting fakes. Open ChatGPT, upload a photo of bare drywall, type "show this kitchen finished with granite countertops and hardwood floors," and in under a minute you have something that would sail through any draw review conducted by someone staring at a screen across the country from the job site. Truepic, a photo verification company that works with private lenders, demonstrated exactly this process using freely available AI tools. Ten minutes, zero dollars, and no technical skill required.
It gets worse. You can generate fake video walkthroughs from a single still image, multiple walkthroughs showing different stages of completion, all fabricated from one photograph of a half-framed room that has not changed since the last draw was approved.
How the Draw Process Actually Works
If you have a construction loan, your lender releases money in stages called draws. Pour the foundation, submit photos and paperwork, get the next chunk of funds. Frame the walls, submit again. Rough-in plumbing and electrical, submit again. Four to six checkpoints on a typical residential project, each releasing $55,000 to $85,000 depending on the loan size and schedule, each one a moment where someone decides whether the work being paid for has actually been performed.
Lenders verify these requests in one of three ways. Large commercial lenders send a third-party inspector to walk the site, which is expensive and slow but genuinely hard to fake. Community banks and credit unions often rely on internal staff who may or may not visit the property, sometimes approving draws based on photos and paperwork alone. Then there are the private and hard-money lenders, the ones funding fix-and-flips and spec builds, who overwhelmingly depend on borrower-submitted photos reviewed remotely by someone who has never set foot on the property and never will.
That third category is the soft target, and it is a larger share of the market than most homeowners realize, because private lending has expanded aggressively into residential construction over the past decade as traditional banks tightened underwriting standards after the 2008 crisis. A borrower snaps photos on their phone, emails them to the lender, and someone at a desk decides whether the work matches the draw request. Before generative AI, faking those photos meant staging a job site, borrowing progress from another project, or learning Photoshop well enough to fool a trained eye. All of that cost real money and carried real risk of detection.
Now it costs nothing. Zero.
Four Million Draw Events Per Year
I ran the numbers because nobody else seems to have, and the math is not complicated, just depressing. Census Bureau data puts U.S. residential construction starts at roughly 1.35 million annually. About 60 to 70 percent of new single-family starts involve construction financing, which gives us approximately 800,000 active construction loans per year. Five draws per loan on average, which produces roughly four million draw events per year.
At a conservative 0.1 percent fraud rate for detected fraud, that implies about 4,000 fraudulent draws per year. At an average draw amount of $70,000, the annual exposure lands at $280 million, a figure that sits remarkably close to the FBI's Internet Crime Complaint Center reported $275.1 million in real estate fraud losses for 2025, which was itself up from $173 million in 2024 and $145 million in 2023.
Something drove that acceleration, and the FBI's own report names it explicitly: AI-referencing complaints exceeded 22,000 in 2025 with adjusted losses surpassing $893 million across all categories, a number that captures everything from deepfake romance scams to fabricated loan documentation. A May 2026 Celent survey of 115 U.S. financial institutions found that 93 percent of lenders say fraud now contributes directly to their credit losses. Eighty-two percent say those losses increased year over year. Nobody is confused about the direction.
My calculation has real limitations, which I will get to, but the directional math is hard to dismiss: four million annual trust checkpoints, each protected by nothing more than the assumption that photos are genuine.
What Lenders Are Building
Two companies illustrate the emerging response, and their approaches diverge in ways that reveal a fundamental disagreement about what the actual problem is and where the defenses need to be built.
Built Technologies launched its Draw Agent in November 2025. It is an AI system that reviews the entire draw package: loan agreement, construction plans, budget, inspection photos, historical draw data, insurance certifications, all of it cross-referenced against patterns that suggest something does not add up. The company claims 95 percent reduction in review time, 60 percent faster draw turnaround, and a striking 400 percent increase in detected risks compared to human-only reviews. Zions Bancorporation, Anchor Loans, and AgSouth Farm Credit are early adopters, and Randy Stewart, an executive vice president at Zions, said what once took hours of manual review now happens in minutes.
Useful, and genuinely so, because it would have caught Turner's forged lien waivers in Indiana before a single dollar moved, would have flagged the discrepancy between claimed subcontractor work and the absence of corresponding insurance certificates, and would have correlated the draw request against the project timeline in ways that a human reviewer processing sixty files before lunch simply cannot. But Built's approach treats the draw as a document-intelligence problem, and it assumes the photos themselves are real, which means the one attack vector that generative AI makes trivial is the one this particular tool was not designed to address.
Truepic attacks the photo layer directly, which is the harder and considerably more interesting problem because it addresses the one assumption that the entire draw process has always depended on without ever formally validating: that a photograph of a construction site is an honest representation of what actually exists at that location on that date. Their Vision platform requires inspection photos to be captured in-app, cryptographically signed at the moment of capture, geolocated, and timestamped with data that cannot be spoofed after the fact. Upload from a camera roll? Blocked. Feed in an AI-generated image? Blocked. Reprocess a downloaded stock photo through a metadata editor and try to pass it off as a jobsite capture? Also blocked. If the photo was not taken on-site in real time through their controlled capture system, it simply does not enter the draw package, and there is nothing a borrower with a generative AI subscription can do about that constraint. Jacob Sherick, a senior manager at Kiavi, one of the largest private lenders in the fix-and-flip space, said the platform gives them faster turnaround without sacrificing integrity.
Truepic has also partnered with the American Association of Private Lenders to develop an educational course called "Media Fakes & Misrepresentation," specifically designed for lending teams who rely on photo evidence in draw processes.
Why In-Person Inspection Does Not Solve This
If you are thinking that a real inspector walking the site eliminates the problem, you are half right. And entirely impractical for the part of the market most at risk. In-person inspections cost $150 to $400 per visit. On a five-draw loan, that adds $750 to $2,000 in verification costs per project. For a community bank funding $50 million in construction loans, fine. For a private lender running 200 concurrent fix-and-flip projects across twelve states? The math craters.
Private lending moved toward borrower-submitted inspections precisely because they are cheap and fast, and the competitive dynamics of the industry punish any lender who adds friction to the draw process because borrowers will simply find one who does not. Market pressure pushes verification toward cheaper and faster methods at the exact historical moment when AI photo fabrication exploits that preference with devastating precision, a coincidence so perfectly timed that you would think someone designed it as a fraud tutorial.
What the Counterargument Gets Right
The strongest objection to this analysis is that draw fraud, even AI-enhanced draw fraud, remains a small-dollar problem compared to the total volume of construction lending. Most residential construction loans are originated by banks and credit unions that use some combination of in-person inspection and internal review. Private and hard-money lenders are a smaller slice of total origination, and they accept elevated risk as part of their business model.
Fair enough. An inspector standing in an unfinished kitchen will not be deceived by a generated image of that same kitchen with granite countertops, because the inspector can see the bare drywall with his own eyes and touch the studs with his own hands. And most construction fraud still follows Turner's template: falsified documents, inflated invoices, misrepresented credentials, all low-tech and effective. The pen-and-printer fraud vector will remain dominant for years because it works and it is simple and nobody has bothered to digitize the lien waiver verification process at most community banks.
But the asymmetry matters. Before generative AI, fabricating convincing construction photos required hundreds of dollars in effort and significant detection risk. After generative AI, it requires nothing, ten minutes and a free account that anyone's grandmother could set up. When the cost of committing fraud drops to zero while the cost of detection remains high, the volume of attempts increases even if the detection rate stays constant, because you are dealing with a population of potential fraudsters whose decision calculus just shifted from "not worth the risk" to "why not try." Lenders who rely on photo-based draw verification without provenance tools are now exposed to a category of fraud that did not meaningfully exist eighteen months ago.
If You Have a Construction Loan
Ask your lender how they verify draw requests, and not in general terms. Specifically: are inspection photos verified for authenticity at capture, or simply reviewed visually by staff? If the answer is visual review of emailed photos, your draw process has no defense against AI-generated fabrication whatsoever, and the person reviewing those photos could not distinguish a real kitchen from a generated one even if you paid them to try. That does not mean your contractor is committing fraud, but it does mean the entire verification system was designed for a world where faking a convincing photograph of a finished kitchen required real skill, real money, and real risk of getting caught.
That world ended.
If you are a general contractor managing subcontractors on a project with construction financing, document your own progress independently. Timestamped photos from your phone, taken daily and stored in a cloud folder you control. This creates a parallel record that protects you if a subcontractor's draw documentation is ever questioned, because when draw fraud succeeds, the people who did actual work are often the last to get paid. Turner allegedly stole $188,000 from a loan meant to pay legitimate plumbers and electricians. Those subcontractors presumably did no work, because Turner forged their waivers, but on projects where subs have completed real work and a dishonest GC submits fabricated draw docs for phases they managed, it is the honest subcontractors who eat the loss.
For lenders evaluating verification platforms: the minimum viable defense is controlled photo capture with cryptographic signing and geolocation, which means photos that are locked to a specific device, a specific time, and a specific set of GPS coordinates at the moment of capture, with no ability to substitute images generated elsewhere. If your draw review process accepts photos uploaded from a camera roll or sent via email, you are trusting that nobody with access to a free AI app has figured out what a finished kitchen looks like. Truepic's approach of locking capture to a controlled in-app environment addresses this specific threat vector. Built's Draw Agent solves a different and equally important problem. The two are complementary, not competing, and a lender deploying one without the other has addressed half the attack surface.
Limitations of This Analysis
My fraud surface area calculation uses rough estimates because actual construction loan data is fragmented across thousands of institutions with no central reporting. Census housing starts data covers permits, not funded construction loans, and the 60 to 70 percent financing rate is an industry estimate rather than a surveyed figure. The 0.1 percent fraud rate represents detected fraud only; the actual rate is higher by an unknown margin because undetected fraud is, by definition, uncounted. The convergence of my $280 million estimate with the FBI's $275 million figure may be coincidental, since the FBI's real estate fraud category includes investment fraud and rental scams, not just draw fraud specifically.
Built Technologies' performance claims (95 percent time reduction, 400 percent more risks detected) come from internal analyses of early adopter programs and have not been independently audited. Truepic is a vendor selling verification services, and their risk framing has commercial motivation. No peer-reviewed study has yet quantified the specific contribution of AI-generated photos to construction loan fraud, and the novelty of the threat means historical data cannot predict future fraud volumes with any confidence.
I could not determine what percentage of draw events nationally are approved based on photos alone versus in-person inspection, and that number would dramatically change the actual exposure calculation.