Can AI Write a Lender-Grade Feasibility Study?

By the middle of 2026 the question arrives in one of two forms. A borrower asks whether the feasibility study the lender wants can be produced in an afternoon with a model and a spreadsheet. A credit officer asks whether the study on the desk was. Both deserve a precise answer, and the precise answer is not "no." A language model can draft prose, clean data, write the code behind a projection, and check an exhibit for arithmetic. What it cannot do, under either federal program or under the standards bank examiners apply to third-party reports, is be the qualified, independent person whose opinion the study is, verify a comparable it has never called, walk a site, or sign a certification that means anything.

This is a practitioner's reading of where that line sits, current as of 7 September 2026: what the programs say, what the regulators say, what the measured reliability of the tools is, where automated drafts fail in practice, and what a lender should ask before relying on a study that used them.

What SBA and USDA say about AI: nothing, and everything

Neither agency has published a notice, SOP provision, or FAQ on AI-generated or AI-assisted application documents, third-party reports, or feasibility studies. SOP 50 10 8, in force since 1 June 2025, and SOP 50 10 8.1, effective for applications receiving an SBA loan number on or after 1 October 2026, are silent on the subject. So is 7 CFR Part 5001. That silence is not a permission slip. The operative requirements both programs do impose are the ones a machine cannot meet.

USDA. Under 7 CFR 5001.3 a feasibility study is "a report including an opinion or finding conducted by an independent qualified consultant(s)," and a qualified consultant is "an independent third-party person possessing the knowledge, expertise, and experience to perform the specific task required." Appendix A to Subpart D requires the study to "conclude with an opinion and recommendation presented by the consultant" and to provide "a resume or statement of qualifications of the author of the feasibility study, including prior experience." For Business and Industry and Community Facilities loans over $1 million to a new business, that consultant must be "acceptable to the Agency." A person, with a resume, whose opinion it is, whom the State Office can accept or reject. A model is none of those things, and a study whose analysis was generated rather than performed has no author to name.

SBA. The SOP names no preparer standard for the feasibility study, but it is exacting about the reports it does regulate, and lenders read those rules across. The going-concern appraisal must come from a Certified General appraiser with four equivalent assignments in the prior 36 months, under USPAP, with a signed certification. The business valuation must come from a Qualified Source holding a named accreditation. From 1 October 2026 the quality of earnings report must be performed by an independent, experienced financial professional for the lender's benefit. Every one of those is a person who takes responsibility. The guaranty purchase reviewer who reads the file after an early default, with 13 CFR 120.524 open, is looking for whether the lender relied on something independent and substantiated; the Office of Inspector General's recurring finding on failed loans is that lenders "did not provide adequate documentation to substantiate financial projections." A generated projection with no verified inputs does not cure that finding. It is that finding.

Why it matters. The programs regulate authorship, independence, and substantiation. They do not regulate tools. A consultant who uses a model to draft, then verifies, corrects, and signs, has authored the study. A borrower who prompts a model and formats the output has produced a document with no author, which is the one thing neither program accepts.

What the bank regulators say

There is no finalized federal rule governing AI in credit underwriting as of September 2026. There is a framework, and it puts the burden where the programs put it.

Model risk. SR 11-7 from the Federal Reserve and OCC Bulletin 2011-12, the Supervisory Guidance on Model Risk Management of April 2011, apply to models obtained from vendors and third parties as well as those built in-house. The institution must validate them, subject them to effective challenge, and document their use, and the institution remains responsible for the output. A feasibility study whose projections come from a model the lender cannot see, validate, or challenge sits uneasily inside that guidance, whether the model is a spreadsheet or a language model.

Third-party risk. The Interagency Guidance on Third-Party Relationships: Risk Management, issued by the Federal Reserve, FDIC, and OCC in June 2023, holds that a banking organization's use of third parties does not diminish its responsibility to operate safely and soundly and to comply with law, and that due diligence should be commensurate with the risk of the relationship. Outsourcing the preparation of a report does not transfer responsibility for its soundness. Neither does automating it.

Automated valuation. The interagency Quality Control Standards for Automated Valuation Models, issued by six agencies in 2024 and effective in October 2025, are the closest binding federal analogue to "AI valuation." They require institutions to adopt controls for accuracy, data integrity, and nondiscrimination when AVMs are used on mortgage collateral. They do not reach feasibility studies or narrative market analysis, and their existence makes the point: where regulators have addressed model-driven analysis directly, they have imposed controls on the user, not blessed the output.

The appraisal profession. USPAP contains no AI-specific rule. An appraiser's use of any technology is governed by the Competency Rule and the Scope of Work Rule, and the signed certification under Standards Rule 2-3 remains personal. No Appraisal Foundation, Appraisal Institute, Appraisal Subcommittee, or state board position permits a machine to sign an appraisal, and none is expected to.

What the evidence says about the tools

The case for caution is not anecdotal. It is measured.

The Stanford RegLab and HAI study "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools," published in the Journal of Empirical Legal Studies in 2025 on tests run in May 2024, was the first preregistered evaluation of commercial legal AI tools built with retrieval grounding. The tools made by LexisNexis and Thomson Reuters hallucinated between 17% and 33% of the time; general-purpose GPT-4 hallucinated on 43% of the same queries. The errors included fabricated authorities and, more dangerously, real sources cited for propositions they did not support.

FinanceBench, the open-book financial question-answering benchmark built from SEC filings in late 2023, tested 16 model configurations on questions whose answers were in the documents provided. GPT-4-Turbo with a retrieval system answered incorrectly or refused 81% of the questions; without retrieval it reached 11% accuracy. The authors' conclusion was that all models examined "exhibit weaknesses, such as hallucinations, that limit their suitability for use by enterprises." Models have improved since; the benchmark's point, that a fluent answer and a grounded answer are different things, has not aged.

The courts have run the largest live experiment. The public database of court decisions involving AI-fabricated material maintained by the researcher Damien Charlotin held about 200 cases in mid-2025, 719 in January 2026, and 1,668 on 2 July 2026, 1,163 of them in United States courts and 653 involving licensed lawyers rather than self-represented parties. The sanctions have moved from the $5,000 fine in Mata v. Avianca in June 2023 to $15,000 per attorney imposed by the Sixth Circuit in Whiting v. City of Athens in March 2026, plus bar referrals and suspensions. The pattern in those cases is the pattern a credit officer should worry about: the output is fluent, correctly formatted, confidently stated, and wrong in ways that are invisible until someone checks the source.

Financial regulators have so far acted at the edges. The SEC's March 2024 settlements with two investment advisers for overstating their use of AI, with penalties of $225,000 and $175,000, and FinCEN's November 2024 alert on deepfake media used to defeat identity verification at financial institutions, are the closest federal enforcement has come to the subject; neither involves a fabricated third-party report, and no SBA OIG or Department of Justice matter turning on an AI-generated feasibility study has been reported.

The lenders themselves are adopting the tools at speed. Cornerstone Advisors' What's Going On in Banking 2026 reports generative AI deployed at roughly half of banks and nearly 60% of credit unions; EY-Parthenon's 2025 survey put 77% of banks at launch or soft-launch. Adoption concentrates in customer service and internal productivity, not in the credit decision, and that is the right place for it in feasibility work as well.

Where automated drafts fail a credit committee

A feasibility study that was generated rather than performed fails in predictable places. These are the ones reviewers find.

Comparables that were never verified. The comp set is the spine of the market component: the competing hotels with their rates and occupancy, the storage facilities with their rents and unit mix, the car washes with their pricing and membership. A model can assemble a plausible list from what it has read. It cannot call the property, check its own website today, confirm the room count with the front desk, or notice that the facility closed in March. Lender-grade practice checks every comp against the operator's own current source and states which figures are primary-verified and which are carried. A generated comp table is, at best, a list of leads.

Site work that was never done. Traffic counts from the state DOT, the zoning ordinance as amended, the parcel's actual access and visibility, the wetland at the back of the lot, the competitor under construction across the road. The SOP's environmental and construction provisions and Appendix A's technical component both assume someone looked. A study that describes a site from satellite imagery and a listing is a study a site visit will contradict.

Regulatory currency. Models know the rules as of their training data. The rules in this field move by notice: the 1.10x coverage floor for 7(a) Small Loans in March 2026, the ownership rule the same month, the $10 million combined limit in July, SOP 50 10 8.1 in August with a 1.25x floor and a QoE requirement, USDA's FY2026 guarantee percentages in March. A generated study that cites SOP 50 10 7.1 equity rules or the legacy 80/70/60 B&I guarantee tiers is a study written from a snapshot, and reviewers notice the vintage before they notice anything else.

Data provenance. The datasets a lender-grade study rests on, hospitality benchmarking, commercial property databases, traffic and demographic products, are licensed subscriptions accessed through the license. Their public terms uniformly prohibit automated extraction and scraping. An automated workflow that reaches them any other way has a provenance problem before it has an accuracy problem, and a study that cites a figure it cannot trace to a licensed pull cannot be defended when the number is challenged.

Internal consistency. A study is a single arithmetic object: the unit mix in the program drives the revenue build, which drives NOI, which drives coverage, which the sensitivity table stresses. Generated drafts reproduce the sections without the linkage, so the occupancy in the narrative disagrees with the pro forma, the debt service ignores the blended maturity, and the sensitivity case does not move the number it claims to move. Reviewers read across exhibits. Models write down them.

The signature. Every failure above is survivable if a qualified person catches it before delivery. The one that is not survivable is the absence of that person. A study with no independent author has no one to certify that the facts are true, the analysis is theirs, the fee is not contingent, and the projections are estimates under stated assumptions. That certification is what converts a document into evidence in a credit file, and it is the thing a lender cannot get from a model.

Where the tools belong

None of this argues for abstinence. It argues for role clarity. In a lender-grade workflow the tools earn their place in five jobs.

Drafting and editing. First drafts of narrative sections from the consultant's verified inputs, edited for house style, tightened, and checked against the numbers. The consultant writes the conclusions.

Data handling. Cleaning survey data, structuring comp tables from verified calls, converting a zoning code or an appraisal into an extractable summary the consultant then reads against the source.

Modeling. Writing and checking the code behind the pro forma, running sensitivity cases, tying exhibits to one another so that a change in the program flows through to coverage.

Quality control. Checking arithmetic across exhibits, flagging inconsistencies between narrative and tables, testing the source list against the text, catching a missing factor from Appendix A. This is the use with the best ratio of benefit to risk, because the model is checking a human's work rather than replacing it.

Research assistance. Locating candidate comparables, public filings, and regulatory documents for the consultant to verify. Every item found this way is a lead, not a finding, until a person has confirmed it against the primary source.

What the tools do not do in that workflow: select the comp set, conclude on capture or absorption, set the assumptions, decide the recommendation, or sign.

A consultant working this way should say so. A short statement in the study's methodology section, that drafting and quality-control tools were used, that every figure was verified by the author against the stated source, and that the analysis, assumptions, and conclusions are the author's own, costs nothing and answers the question before the credit officer asks it.

The disclosure question

No federal or state law as of 7 September 2026 requires a feasibility study, appraisal, or lending report to carry an "AI-generated" label. The state statutes that touch AI in financial decisions regulate something else. Colorado's original AI Act, SB 24-205, which would have imposed duties on deployers of high-risk AI systems used in lending decisions, was stayed by a federal magistrate in April 2026 after a constitutional challenge the U.S. Department of Justice joined, and was repealed and replaced on 14 May 2026 by SB 26-189, a narrower notice-based law effective 1 January 2027. Utah's Artificial Intelligence Policy Act requires disclosure when a consumer is interacting with generative AI; California's AI Transparency Act requires provenance tools from providers of large generative systems; Texas's Responsible Artificial Intelligence Governance Act, effective 1 January 2026, addresses prohibited uses and government disclosure. None labels a report. A December 2025 executive order directing the Department of Justice to challenge state AI laws makes further state mandates less likely, not more.

That leaves disclosure where the rest of the authorship question sits: in the engagement letter and the certification, as a matter of what the lender is entitled to know about how the study it is relying on was made.

What a lender should ask

Eight questions settle the AI question for any study in the file.

1. Who is the author, and does the study carry a signed certification of independence, qualifications, and non-contingent compensation?
2. Which comparables were verified against the operator's own current source, by whom, and on what date?
3. Was the site visited, and by whom? What did the visit change?
4. Which datasets were used, under what subscription, and can each figure be traced to a dated pull?
5. Which SOP edition and which USDA notice does the study cite, and are they the ones in force for the loan-number or obligation date?
6. Does the sensitivity table move the coverage number, and do the narrative, pro forma, and debt schedule agree?
7. Were generative tools used, for what, and what verification step sits between the tool's output and the study's text?
8. Would the author sign a reliance letter to a second lender on this study as written?

A study that answers all eight is lender-grade whatever tools helped produce it. A study that cannot answer the first, second, or seventh is not, however it was produced.