I have watched more construction projects stall over money than over materials. Not because the money was not there, but because it was stuck. Committed in a lender's system, locked behind a review process that nobody designed for speed, waiting on a human being to compare a stack of lien waivers against an AIA G702 application that was faxed, scanned crooked, and uploaded as a 47-page PDF where pages 12 through 18 are rotated 90 degrees. Seven business days. That is the industry average for a construction draw to move from borrower request to funded disbursement. Seven days of your framer wondering whether to start another job because this one stopped paying.
Built Technologies, a Nashville company whose platform now serves 185,000 contractors and processes loans through partners including US Bank and Citi, deployed an AI system called Draw Agent earlier this year that reviews and approves construction draws in as few as three minutes.
Not three days, not three weeks, but three minutes for a process that used to consume the better part of a loan administrator's work week.
Why Draws Take So Long in the First Place
If you have never built a custom home, here is how the money works. Your construction lender does not hand your builder a check for $450,000 on day one. Instead, the loan is divided into draws, typically five to seven, each tied to a construction milestone. Foundation poured, framing complete, mechanicals roughed in, and so on. Before each draw is released, the lender must verify that the work was actually done, that the right people were paid, that insurance is current, that nobody filed a lien, and that the budget still makes sense against the remaining scope.
That verification requires documents. Lots of them. Invoices from every sub and supplier. Conditional and unconditional lien waivers from each party. Inspection reports. Updated insurance certificates. Progress photos. Change orders, if the scope shifted. A sworn construction statement or application for payment, usually on AIA G702 and G703 forms, which are industry-standard documents so dense with line items that reading one carefully takes an experienced reviewer forty-five minutes.
Now multiply that across a lender processing hundreds or thousands of active construction loans simultaneously, each with its own draw schedule, its own cast of subcontractors, its own jurisdictional insurance requirements, and one or two loan administrators trying to keep up by toggling between email, spreadsheets, PDFs, and whatever document management system the bank chose in 2019.
Thomas Schlegel, VP of Engineering at Built Technologies, described the old process with the kind of weariness that comes from watching it for eight years: "It's just one of the most inundated, labor-intensive, inefficient processes that exists in construction."
How the Draw Agent Works
Built Technologies partnered with MightyBot to build what they call an "AI exoskeleton," a system that wraps around Built's existing loan management platform without replacing it, and the architecture matters because lenders are, justifiably, paranoid about ripping out systems that handle billions in loan value. Draw Agent went from concept to production in three months, but nothing in Built's core infrastructure got replaced; the AI reads the same data, through the same APIs, that human reviewers used yesterday.
When a new draw package arrives, the agent classifies every page independently with confidence scores. A 25-page submission might contain an AIA G702 application on pages one through three, a G703 continuation sheet from four to twelve, conditional lien waivers from three different subcontractors on pages thirteen through eighteen, an inspection report on nineteen, and insurance certificates from twenty to twenty-five, all in mixed formats with poor scan quality and rotated pages that the system parses without complaint.
Then it cross-references every data point against every other data point in the package. Are the dollar amounts on the lien waivers consistent with the payment application? Does the inspection report confirm the milestone the draw is requesting payment for? Is the insurance certificate current, or did it expire last Tuesday while the document sat in someone's inbox? Are there any subcontractors listed on the budget who have not submitted waivers, and if so, which ones and for how much?
A human reviewer good at this job catches most of those discrepancies. A tired human reviewer on her fortieth draw package of the week misses the insurance expiration date because the certificate was buried on page twenty-three and the renewal date font is eight points. Built reports that Draw Agent catches four times more compliance risks than human-led reviews in production.
The Dead Capital Problem Nobody Talks About
Here is an original calculation that neither Built Technologies nor the lending industry surface publicly, because it does not flatter anyone.
A typical US construction loan for a new single-family home runs around $350,000, according to NAHB data. At a 7% interest rate, every day that money sits in "committed but unfunded" status costs the lender approximately $67 in lost interest. That sounds trivial until you multiply it across a draw schedule. A seven-draw project where each draw takes seven business days to process means 49 cumulative days of dead capital per loan. That is $3,283 in lost revenue per loan, just from processing delays.
For a mid-size regional bank processing 500 active construction loans, that is $1.64 million per year in interest the bank could have been earning but was not. Schlegel confirmed the incentive directly: "If a lender can disperse money faster, they're typically making more money."
Now look at it from the builder's side, which is where it actually hurts people. Every day a draw sits in review is a day the builder is financing the project out of pocket or, more likely, out of another project's cash flow. "Very frequently, builders are using money from Peter to pay Paul," Schlegel told HousingWire. "And that's when things come sideways."
The Trust Ladder
Built did not hand the keys to an AI on day one. They built a three-stage adoption model that any project manager would recognize as change management done correctly.
Stage one is audit mode, where the AI reads everything, analyzes everything, flags everything, and does nothing about any of it. It produces a report that sits next to the human reviewer's own assessment, and lenders compare the two, learning what the system catches that they miss and what it flags that turns out to be a non-issue over weeks of parallel operation that builds calibration on both sides.
Stage two is assist mode, where the AI handles routine tasks like scheduling inspections, sending borrower communications, and pre-populating compliance checklists while the human still approves every draw. The workload shifts dramatically: instead of spending 20 hours processing a package, the reviewer spends 90 minutes checking the AI's work.
Stage three is full automation, where the Draw Agent independently completes the review and approval process for straightforward draws while complex cases still route to humans. Three Built customers were live in this mode as of the MightyBot case study, processing thousands of draws monthly.
This phased approach is the only reason it works, because construction lending is a domain where a single error can mean a lien on someone's half-built house, an insurance gap that leaves a burned structure uninsured, or a draw released for work that was never completed. You do not automate that overnight, and anyone who tells you otherwise has never sat across from a homeowner explaining why their project is three months behind schedule because the bank funded a draw against a lapsed insurance certificate that nobody caught.
What This Does Not Fix
Built reports 99%+ accuracy in production, and the investor presentation version of this story ends there, but the project manager version keeps reading.
All performance metrics come from Built Technologies, and no independent third-party audit has verified the 95% time reduction, the 4x risk detection improvement, or the 99% accuracy figure. At scale, even 99.5% accuracy means that one out of every 200 draws contains an undetected error. For a lender processing 2,000 draws per month, that is ten errors that the system was supposed to catch and did not, and some percentage of those errors involve real money disbursed against work that was never completed or insurance that had lapsed without anyone noticing.
We also do not know the false positive rate. If Draw Agent flags 400% more "risks" than human reviewers, how many of those flags are genuine compliance issues versus artifacts of imperfect document parsing? A system that over-flags creates a different bottleneck: the human reviewer now spends time adjudicating AI alerts instead of reviewing draws directly. Built's phased rollout mitigates this during stages one and two, but in full automation mode, over-flagging could slow the very process the tool was designed to accelerate.
And concentration matters because Built handles loan management for 185,000 contractors through major banks. If a systematic error in the document classification model, say, a failure to recognize a new lien waiver format from a major title company, propagates across thousands of simultaneous draw reviews, the blast radius is not one loan but every loan processed that day.
What Builders and Homeowners Should Know
If you are building a custom home on a construction loan, ask your lender whether they use Built Technologies or a comparable AI draw platform, and whether Draw Agent is active on your loan. If it is, your draws should process measurably faster than the seven-day industry average, and that acceleration compounds in ways most homeowners never consider: faster draws mean your builder can pay subs on time, which means your subs show up on time, which means your project stays on schedule, which means your interest reserve lasts longer and your final cost stays closer to the number you signed.
If your lender still processes draws manually, you are absorbing the cost of their operational inefficiency in the form of delayed disbursements and the downstream schedule slippage they cause. That cost is invisible on your closing statement but very real in your carry budget.
For builders running multiple projects: the cash flow math changes when draws clear in days instead of weeks. A builder carrying three concurrent projects with $100,000 in pending draws across them is financing $100,000 out of operating capital or, more commonly, out of the next project's budget. Clearing that float faster does not reduce the total cost of the project. It reduces the total cost of running the business, which is the cost that ultimately gets passed to you, the person buying the house.
For lenders reading this: the competitive advantage window is narrow. Your construction lending desk will either adopt this technology or lose builders to competitors who already have, because the builder does not care which bank gives them a slightly better rate. They care which bank gets them paid this week.
Frank DeLuca has managed commercial and residential construction projects for twenty years. He writes about the systems, processes, and overlooked operational details that determine whether a project finishes on time and on budget. The dead capital calculations in this article use published NAHB average loan amounts and prevailing interest rates; actual figures vary by market, loan size, and lender terms.