The File That Clears Everything Except the Building
A condo loan can have a perfect borrower and still die on the project. The income is calculated, the assets are sourced, the appraisal supports the value, and then the file sits for two weeks while someone chases an HOA management company for a questionnaire, a reserve study, and a master insurance certificate. When the questionnaire comes back, it says the association voted a special assessment for balcony repairs. Now the question is not whether the borrower qualifies. It is whether the project is eligible to be sold to Fannie Mae or Freddie Mac at all, and that question is the one most lenders staff worst.
I build the underwriting agents on our origination side, and the condo project review is the workflow where I see the widest gap between how important it is and how little attention it gets. The borrower-facing math gets the engineering. The project review gets a processor, a shared inbox, and a spreadsheet of management-company contacts. So it is the slowest step in the file, and after the Surfside collapse it became the step most likely to turn a clearable loan into a denial. An AI agent helps here, but only if it is built to understand that project warrantability is a different kind of decision than borrower qualification, and that the rules changed in a way that made the old questionnaire logic wrong.
What Changed in 2023, and Why Your Old Checklist Is Out of Date
The rules in force now came out of Fannie Mae Announcement SEL-2023-06, which took the temporary requirements Fannie issued after Surfside in Lender Letter LL-2021-14 and wrote them permanently into the Selling Guide. They apply to loans in condo and co-op projects with five or more attached units, whatever the review type, for applications dated on or after September 18, 2023. Freddie Mac runs a parallel set in its Guide, because both ran the change at FHFA's direction. If a lender is still screening projects on a pre-2023 questionnaire, it is screening against a rule that no longer exists.
The substance is in B4-2.1-01, General Information on Project Standards. A project is ineligible if it needs critical repairs, and Fannie defines critical repairs as work that significantly affects the safety, soundness, structural integrity, or habitability of the project's buildings, plus any mold, water intrusion, or potentially damaging leaks that have not been addressed, all life-safety hazards, and the results of deferred maintenance that has become significant. A project under a current evacuation order is ineligible. A project with a special assessment is eligible only if the lender documents the reason for it, confirms the association has the funds to complete the work, and confirms the work is not critical. Routine maintenance and repairs that do not rise to the critical bar do not make a project ineligible, which is the distinction that decides most files and the one a keyword-matching checklist gets wrong.
There is a reserve signal underneath all of this. Fannie flags any project that budgets less than 10 percent of assessment income for reserves, treating it as at increased risk of significant deferred maintenance and a future special assessment. A project under the 10 percent line is not automatically ineligible, but it is a project the reviewer has to look at harder, usually with a reserve study, because the rule is telling you where the next special assessment comes from.
What the Agent Reads, and What It Is Reading For
The inputs to a project review are a small pile of documents that arrive in no fixed format: the lender's project questionnaire (Fannie Form 1076 or the limited-review short form), the HOA's operating budget, a reserve study if one exists, the master insurance certificate, the covenants, and sometimes meeting minutes. The processor's job today is to read all of them, find the handful of facts that decide eligibility, and notice when two documents disagree. That last part is where humans slip, because the facts are scattered and the disagreements are quiet.
The agent reads the whole pile against the loan. It pulls the eligibility-determining facts into the Loan Graph as structured fields: owner-occupancy percentage, single-entity ownership concentration, the share of commercial space, delinquent HOA dues as a percentage of units, the presence and amount of any special assessment, the reserve contribution as a percentage of the budget, the answers to the critical-repairs and deferred-maintenance questions, and the insurance coverage and deductible against the guide limits. Then it does the thing the shared-inbox process does not do reliably: it cross-checks the documents against each other. If the questionnaire says no special assessment and the budget shows a line item for one, that is a contradiction the agent surfaces rather than a box it quietly checks. If the questionnaire is signed by someone whose authority the form itself requires, the agent confirms the signer is not an interested party where the guide prohibits it.
The questionnaire changed too, which matters for the agent's logic. When SEL-2023-06 took effect, Fannie added questions to Condo Project Manager covering critical repairs, material deficiencies, significant deferred maintenance, inspection reports, evacuation orders, and special assessments, and those answers can change a certification result. An agent that extracts the questionnaire has to read the current fields and know which answers are disqualifying on their own, not the fields the form carried three years ago.
Why This Is a Governed Model, Not a Document Reader
It would be easy to frame this as extraction: read the questionnaire, populate the fields, done. That framing is how a lender gets a finding. A project review is a determination that a loan is eligible for sale, and a wrong one is a repurchase risk the same way a wrong income calculation is. So the component that reads the questionnaire and proposes an eligibility conclusion is a model in the credit path, and we govern it the way SR 11-7, the interagency model-risk guidance, expects: validated against a labeled set of real questionnaires before it touches a live file, monitored for the error that matters, and bounded by a confidence floor below which it stops.
The error that matters here is asymmetric, and the agent's thresholds reflect that. A false ineligible, where the agent flags a project that is actually warrantable, costs a review cycle and annoys a reviewer. A false eligible, where the agent clears a project that needed a critical repair the questionnaire buried, is a loan sold on a representation that was not true. So the confident-clear threshold on the critical-repairs and special-assessment questions sits high, and anything the agent is not sure about on those specific questions routes to a human rather than clearing. The agent is tuned to escalate exactly the files where being wrong is expensive, which is the inverse of tuning it to maximize how many projects it clears without help.
Where the Agent Stops
The agent does not make the warrantability call on a close file, and it does not decide a project is eligible on documents that contradict each other. It assembles. On a project that clears cleanly on every determinative field, with the documents agreeing and the critical-repairs answers negative, it produces the warrantability conclusion with every field cited to the document and page it came from, and a human reviewer confirms. On a project with a special assessment, a reserve contribution under the 10 percent line, an evacuation order, a positive answer to any critical-repairs question, or a contradiction between documents, it stops and hands the file to the project-review specialist with the issue named and the supporting pages attached. It also stops when the inputs are incomplete, because the most common real-world failure is a questionnaire that came back with questions blank rather than a genuinely hard project, and the agent's job there is to go back to the management company and get the missing answers, not to guess them.
What the agent never does is substitute its own judgment for the appraiser's or the lender's on whether a repair is critical. Fannie's own guidance says relying solely on the appraisal to complete a project review is not recommended, and the same caution applies to the agent. The determination that a given repair significantly affects structural integrity is a judgment the lender owns. The agent brings the reviewer the appraiser's comments, the questionnaire answer, the budget line, and the reserve study finding, lined up so the human can make that call in minutes instead of hours. It does not make the call.
The Thing That Actually Gets Faster
The reason to put an agent here is not that it reads a questionnaire faster than a person, though it does. It is that the project review is a waiting problem before it is a reading problem. A loan spends days in project review not because the reading is slow but because the documents arrive in sequence, someone has to notice each one landed, and the file goes back into a queue between each step. A long-horizon agent holds the goal of a complete project review and works it continuously. It requests the full document set from the management company the day the condo is identified, not after the appraisal. It wakes when the budget arrives and checks whether the reserve line clears the 10 percent signal. It wakes again when the questionnaire lands and runs the critical-repairs logic. When a document is missing or an answer is blank, it goes back to the source on its own rather than waiting for a processor to notice the gap at the next file touch. The days it gives back are the days the file used to spend sitting between touches, which on a condo is most of them.
At Sei we treat the project review as a desk inside the system rather than a document-reading feature bolted beside underwriting, because the review reads the same loan model the rest of the file does and the determinative facts belong in the Evidence Trail next to every other decision. The file we hand an examiner or an investor on a condo loan shows the questionnaire, the budget, the reserve finding, the special-assessment documentation, and the critical-repairs answers, each cited, with the eligibility conclusion and the human who confirmed it. That is the file the post-Surfside rules were written to produce, and it is the one a project review run out of a shared inbox almost never assembles in time.
Ramkumar Venkataraman
CTO & Co-Founder