You need to enable javascript to run this app
Documentation menu

Docs / AI

How LendrTech uses AI

What AI does inside LendrTech, what it is never allowed to do — produce a number your borrower could be quoted — and what happens to a document you upload.

Before you drop a borrower's Loan Estimate into any piece of software, two questions deserve a straight answer: is a machine going to invent a number you'll have to defend, and where does that document go? Neither question is paranoid. Both have specific answers, and they're on this page.

The one rule

AI never produces a number your borrower could be quoted.

Every figure LendrTech shows you — the monthly payment, the cash to close, the APR, the mortgage insurance, every subtotal on a fee sheet and every bar on a chart — comes out of LendrTech's calculation engine. That engine is ordinary code: the same inputs produce the same outputs every time, on your screen, on the report, and in the PDF your borrower opens. No model is asked to compute it, estimate it, sanity-check it, or round it.

That isn't a promise about intentions — it's how the software is put together. The AI-assisted parts of LendrTech hand fields to the calculator. The calculator does the arithmetic. Nothing crosses that line.

Where AI is used today

One place: reading a lender's document into a fee sheet.

When you import a fee sheet PDF, a model reads the file you uploaded — a lender fee sheet, a Loan Estimate, a Closing Disclosure — and reports back what it found: loan amount, rate, term, the closing cost line items, prepaids, taxes and insurance, credits. Those values land in a new fee sheet as starting text in form fields. Everything downstream of them — the totals, the cash to close, the payment, the APR — is then computed by the same engine that handles a fee sheet you typed by hand.

Reading a document and pricing a loan look similar from a distance and are completely different jobs. One is transcription: the number is printed on the page, and the work is finding it and putting it in the right box. The other is judgment about money, and it belongs to you and to the calculator. Only the first job is given to a model.

The instruction that model works under is blunt about the difference: return a value only when it is clearly present on the page — never guess, never infer, never compute a value you cannot read directly. Absent means absent. A field the document doesn't state comes back empty and arrives on your fee sheet empty, rather than filled with something plausible. The one piece of arithmetic it may do is arithmetic the page has already done for you — a prepaid line printed as a number of months at a stated amount per month. Anything that would require adding up unrelated line items comes back blank on purpose.

You review before anything is saved

Nothing an import produces is saved, sent, or shown to a borrower until you look at it.

An import drops you on a fee sheet with a Review needed banner and every imported field highlighted, so the fields a model touched are visually distinct from the ones you filled in yourself. You confirm them against the original document, correct anything that's off, fill in whatever the PDF didn't cover, and click Save. Until you do, nothing is stored.

The review step is the point of the design, not a formality bolted on the end. Treat AI output the way you'd treat a fast assistant who has never seen this lender's fee sheet before: quicker than typing, still yours to sign off.

What happens to what you upload

What leaves LendrTech is the file you chose, plus a fixed instruction naming the fields to look for. Nothing else travels with it — not your client list, not your other fee sheets, not your saved reports, not your branding, not anything about your other borrowers.

That request goes to a third-party AI model provider over a paid business API, not a consumer chat product, and under those terms what you send is not used to train models. That distinction is the reason for the paid tier. The document is processed to return the fields and nothing more: no second use, no analysis of your book of business, no model getting better at your expense.

LendrTech doesn't file the upload away either. It reads the PDF, fills in the form for your review, and the upload is finished. What persists is the fee sheet you save — stored under the same protections as everything else in your account (see your account and security).

And you never have to use it at all. Enter manually and Use template sit beside Import PDF in the same dialog, and they build exactly the same fee sheet. An account that never uploads a PDF never sends anything to a model.

Where your AI usage shows up

Open Settings from the sidebar and choose the Billing tab. AI usage lives there, on its own cards, so it's never a surprise on an invoice:

  • PDF imports this month counts imports against your plan's monthly allowance. One upload is one import, whatever the document.
  • AI assistance, where your plan includes AI-assisted work beyond importing, is a separate monthly pool — shown as the percentage left this month, the date it refills, and a breakdown of what's drawn on it so far.

The two are independent — an import never draws down the AI assistance pool, and a full bar on one card says nothing about the other. A failed extraction isn't counted at all, and the error message says so. If a monthly pool does run out, the AI-assisted step pauses until it refills; everything you build, price and send by hand keeps working exactly as it does now.

How much each plan includes is on the pricing page; usage and limits explains how the meters behave, and plans and billing covers changing plans.

What AI gets wrong

Extraction is good, not perfect, and the failures are worth knowing:

  • Unusual layouts. A vendor template the model hasn't seen much of may come back partly filled, with a fee or two in the wrong row or missing.
  • Scans and photos. A picture of a page reads poorly or not at all — a PDF with real text behind it works far better.
  • Documents showing several scenarios. Where a page prints a comparison column beside the proposed loan, a figure can come from the wrong column.
  • Fees under house names. A line item your lender calls something unusual may land in a neighboring row.

The most important one: it does not know when it's unsure. There's no confidence score behind an imported value, which is exactly why every imported field carries the same "check this" highlight instead of a reassuring percentage on some fields and not others. Uniform doubt is more honest than confident-looking doubt.

If a lender fee sheet that's common in your market consistently comes back wrong, send it to team@lendrtech.com — real documents are how the importer improves. Troubleshooting covers an import that misses badly, and the mortgage glossary explains any field name on the review screen you'd rather not guess at.

Still have a question about how any of this works? Email team@lendrtech.com — we'd rather answer it than have you guess.