The AI-lite problem: why most mortgage vendor AI stops at the LOS door

“Most lenders play it safe. They confine AI to front-end chat or lead nurture, keeping it far away from loan origination systems, servicing databases, or investor reporting workflows. It’s ‘AI lite’ — safe, but shallow.”

That’s AngelAi CEO Pavan Agarwal in National Mortgage Professional, November 2025. He named the pattern most mid-market mortgage shops are now living inside.

The AI adoption numbers look like growth: 15% of lenders in 2023, 38% in 2024 per STRATMOR’s 2024 Technology Insight Study, Digital Innovation module. Underneath that headline, 63% of the lenders using AI rely on vendor tools and only 21% develop their own internal capability. The shape of the spend matters more than the volume of the spend. The shape is AI-lite.

I’d argue the framing of “safety” is downstream of a vendor business model, not an engineering choice. Vendors who can’t integrate with Encompass sell that limitation as caution. Lenders pay for the marketing layer because the workflow layer requires engineering capacity neither side has on staff. Here’s the gap, what it costs, and what closes it.

1. What “AI-lite” actually means in mid-market mortgage

The pattern repeats across vendor categories. Chatbots that answer prospect questions before they hit the form. Lead-nurture sequences that ping borrowers between conditional approval and clear-to-close. Application “smart fill” that pre-populates the 1003 from a driver’s-license scan. All real. All shallow.

What it does not include: doc validation against the specific condition the underwriter asked for. Pre-qual logic that runs outside Encompass and writes results back. TRID redisclosure detection that catches an APR change before the next batch ships. Condition-clearing inbox automation that emails the borrower with the right doc request and validates the response.

The line is the LOS boundary. STRATMOR’s recent commentary calls AI-lite “tourist AI” — visiting the operation, taking photos, leaving without changing anything underneath.

Optimal Blue CEO Joe Tyrrell, on the cost of the wrong kind of automation: “The only thing worse than a bad process is an automated bad process.” That’s the AI-lite tax. The shop bought the tool, ran it for six months, then quietly stopped using it because the front-end automation revealed that the back-end process was the actual problem.

2. The cost is structural, not cyclical

MBA Q2 2025 puts mortgage production cost per loan at $10,965. Top-quartile lenders spend $6,900; bottom-quartile spend $16,500 — a 2.4× productivity gap on the same loan. Q1 2025 critical defect rate hit 1.31% per ACES, and income/employment defects jumped 42.5% year over year. Average mortgage repurchase rate is 0.49% and each repurchase costs $32,288 per Reggora and STRATMOR. Cost-to-produce has climbed for 20+ consecutive years.

Twenty years. Front-end chatbots did not move that line. They will not.

The work that moves the cost-per-loan line lives inside the LOS, or upstream of it with write-back. Doc classification + income calculation cuts 5–8 hours per file per Areal.ai. Freddie Mac’s data: lenders using its full LPA digital capabilities save $1,700 per loan, and using multiple automation tools cuts defect rates by up to 75%. None of that is AI-lite. It is LOS-integrated workflow automation that requires a real build.

TRUE leadership, quoted in 2025: “I have talked with many CEOs and owners of lenders who are frustrated with high technology spending but have little to no ROI for it. The spending amounts to hundreds of dollars per loan but the cost to manufacture the loan keeps going up every year, a pattern that has been consistent for more than 20 years. There is this jaded view that is impacting the adoption of AI mortgage tech.”

The jaded view is correct. The fix is not more AI-lite.

3. What “operator AI” looks like at scale

Better.com’s Betsy AI agent placed 1.89 million calls in 2025 and saved loan officers 1,666+ hours per month, automating 35.5% of borrower inquiries end-to-end, per the ElevenLabs case study. That is voice AI inside the inbound funnel, not next to it. The agent qualifies, schedules, and writes results to the system of record.

Ocrolus customers process 95%+ of borrower documents through its classification and extraction pipeline — Better.com being the public reference. The pipeline runs against the actual file the underwriter will read, in the LOS the file lives in. Not a separate intake tool.

Pennymac took an equity stake in Vesta in October 2025 to migrate to an “AI-native” LOS, with CEO David Spector saying the partnership is “already producing benefits through faster speeds and lower costs to close.” Building or partnering for an AI-native LOS is not an AI-lite move. It is a thesis on which layer of the stack the next decade of mortgage automation actually lives in.

MeridianLink launched “Millie,” its role-based AI agent suite, in 2025 and reported 38% YoY customer mortgage volume growth versus the MBA’s 19% industry growth. The lift comes from automation that lives inside the platform, not on top of it.

The pattern is consistent: the ROI lives in the LOS-integrated work. The AI-lite layer (chat, nurture, intake widgets) lives on top because it is what vendors can ship in 12 weeks without integration partnerships.

4. Why mid-market shops keep buying it anyway

Two reasons. First, the procurement story. AI-lite vendors run a $30K–$120K/year SaaS contract that any line-of-business owner can sign without IT involvement. LOS-integrated work requires a real engineering effort, IT review, and sometimes a Capital Project budget line. The friction asymmetry pushes spend toward the shallow side.

Second, the demo story. Voice agents, chatbots, and lead-nurture flows demo beautifully. A 30-minute Zoom shows the chatbot answering a question, the email sequence firing, the smart-fill populating a 1003. The LOS-integrated work demos badly — the value is in incidents that did not happen (a TRID re-disclosure caught, an exception that did not stack, a buyback that was not triggered). That value shows up on a 90-day retrospective, not a sales call.

Result: 63% of mortgage lenders using AI rely on vendor tools per STRATMOR, and the cost-per-loan line stays flat. The 21% who develop internal capability — Pennymac, Better, MeridianLink in part, a handful of mid-market shops with senior AI engineering on staff — pull ahead.

The middle option is fractional. Someone who has shipped operator AI inside Encompass before, embedded for 8–12 weeks, hands off the system with a runbook. Same architectural pattern as the build-it-yourself route, without the 6-month senior-AI engineer hiring cycle. We covered the math in You can’t hire a senior AI engineer fast enough.

5. The four questions to ask any AI vendor before signing

If you are evaluating a mortgage AI vendor right now and the framing feels like marketing, here are the four questions. The framework comes from auditing 14 vendor pitches over the last twelve months. The ones that pass these four typically ship operator AI. The ones that dodge ship AI-lite.

1. Where does the output write to in Encompass (or our LOS)? Specific field paths, not “we integrate via API.” If the answer is “the user copies the output into the LOS,” it is not LOS-integrated.

2. What happens to your output when an underwriter overrides it? Real operator AI surfaces the override, logs the disagreement, and feeds back to the model. AI-lite leaves the override invisible and the model untouched.

3. Can you show me a customer where you have replaced a manual workflow, not added a new screen? Adding a new screen is AI-lite. Replacing a workflow means the team stopped doing the manual version because the AI version is reliable enough.

4. What’s your defect rate on production output, dated and named? A vendor that ships LOS-integrated AI tracks defect rates by category. A vendor that ships AI-lite cites accuracy on synthetic benchmarks.

If the vendor’s answers are vague on any of these, the product is probably AI-lite. That is not a moral failing on the vendor’s part — it is the rational shape of a SaaS business that has to sell to 200 shops with one shared codebase. It just is not what mid-market lenders need.

What this looks like in production

A mid-market broker shop we know bought three AI products in 2024: a voice intake agent, a lead-nurture sequencer, and a borrower portal chat layer. Total spend, ~$90K/year. Twelve months in, the voice agent was forwarding 80% of calls to humans because the qualifying logic could not be tuned for the shop’s specific overlay rules. The nurture sequencer was sending borrowers redundant reminders because it could not see status changes in Encompass. The chat layer answered FAQ-level questions and routed everything else to an LO.

Cost-per-loan was unchanged. Defect rate was unchanged. The COO told us the spend felt like “renting marketing.”

The shop kept the chat layer and dropped the other two. Replaced them with a two-month build: a condition-clearing inbox agent that read the underwriter’s actual conditions list, sent borrowers specific doc requests, OCR-validated the response, and wrote cleared docs back to the eFolder. Same engineering team, fractional, 9 weeks end-to-end. Condition cycle dropped from a median of 11 days to 4. Cost-per-loan finally moved.

That is the gap between AI-lite and operator AI in a single shop. The difference is not the model. It is whether the system actually touches the system of record.

Where this goes next

The 38% of lenders running AI in 2024 will look closer to 70% by end of 2026 if STRATMOR’s trajectory holds. The 63% leaning on vendor tools will become structurally split: the ones with budget and patience will partner for LOS-integrated builds. The ones without will keep paying AI-lite vendors and wonder why the cost-per-loan stays at $11K.

The shops that win the next 36 months are the ones whose ops people can answer the four questions above for every tool already in their stack. That audit takes a week. The shops that have not run it are buying ghosts.