What 90 lending institutions told us about AI and credit work.
How they underwrite today, where the work stalls, and what they ask before buying AI. Every figure is a count of the institutions we spoke with.
THE STUDYFREE TO READ AND CITE
Who we spoke with
Banks, credit unions and CDFIs49
Private credit and non-bank lenders24
Intermediaries13
Other lenders4
Institutions in total90
Over nineteen months of recorded conversations. Not a representative sample.
Method
Ninety institutions, nineteen months, every figure a count
Over nineteen months we sat down with 90 lending institutions and asked how they underwrite, what they think of AI, and what they would need to believe before they bought any of it. We recorded most of the conversations, and every number on this page traces back to one of them.
THE 90 INSTITUTIONS, BY TYPETOTAL 90
Banks, credit unions and CDFIs49
Private credit and non-bank lenders24
Intermediaries13
Other lenders whose function did not cleanly fit those categories4
HOW TO READ THE FIGURES
This is not a representative sample. It is Texas-weighted, weighted toward commercial real estate, and built almost entirely from warm introductions.
Every figure is a count of institutions among the ones we spoke with, never an industry average. Where fewer institutions answered a question, the finding says how many.
The census of the 90 closed on September 3, 2026, and the findings were published on September 7, 2026. Our conversations have continued since: more than 100 lenders in two-way conversations over 21 months, more than half of them banks and credit unions.
Finding 1
Underwriting still lives in spreadsheets and documents
Twenty-five of the institutions we spoke with described underwriting that lives in Excel and Word, and 14 said they have no loan origination system or spreading platform at all.
INSTITUTIONSSCALE: 0 TO 90
Underwriting lives in Excel and Word25
No loan origination system or spreading platform at all14
Three institutions put a number on where the time goes: 60% to 90% of analyst time goes to data entry.
Finding 2
The memo and the first look are where work stalls
Thirteen institutions named memo writing as the bottleneck and 13 named pre-screen triage. Eighteen named post-close monitoring as a blind spot.
INSTITUTIONSSCALE: 0 TO 90
Memo writing is the bottleneck13
Pre-screen triage is the bottleneck13
Post-close monitoring is a blind spot18
About 5%
Median close rate, from the seven institutions that gave us enough of their funnel to compute one.
48 hours to three weeks
Package-to-memo turnaround, from the six institutions that told us.
Finding 3
Security came first among the questions lenders asked
Twenty-one institutions asked about data security. Twenty asked why they should not just use a general-purpose model. Eighteen asked about hallucination and accuracy. Seventeen asked for a SOC 2 report, an NDA, or references. Thirteen asked whether their data would train a shared model. Eleven asked about explainability, and 10 raised what their regulator would think.
INSTITUTIONSSCALE: 0 TO 90
Data security21
Why not just use a general-purpose model20
Hallucination and accuracy18
A SOC 2 report, an NDA, or references17
Whether their data would train a shared model13
Explainability11
What their regulator would think10
Finding 4
Lenders raised credit policy without being asked
Eleven institutions asked whether AI can read a credit policy at all. Six named policy exception detection as the thing that most impressed them.
INSTITUTIONSSCALE: 0 TO 90
Asked whether AI can read a credit policy at all11
Named policy exception detection as what most impressed them6
Finding 5
AI is already inside the building, and vendors have a record to answer for
Eighteen of the institutions we spoke with use a general-purpose model informally today. Sixteen have an internal AI build underway. Twenty-two described vendor fatigue or a failed implementation.
INSTITUTIONSSCALE: 0 TO 90
Described vendor fatigue or a failed implementation22
Use a general-purpose model informally today18
Have an internal AI build underway16
Have a board or executive mandate on AI8
Eight institutions told us they have a board or executive mandate on AI. Every one of them is dated 2026.
Finding 6
Diversified books outnumbered CRE-only ones
Roughly 40 of the institutions described their own loan mix. Sixteen of them are diversified or C&I-heavy. Four are CRE-only.
INSTITUTIONSSCALE: 0 TO 90
Diversified or C&I-heavy16
CRE-only4
What the findings mean for a credit team
What a credit team evaluating AI can take from the figures above.
FINDING 1
Count the hours before you count the software
Where analyst time goes in your shop, and where files wait, is the baseline any tool will be judged against. Measure it first.
FINDING 2
Look past the closing
More institutions named post-close monitoring as a blind spot than named either bottleneck before closing. Ask how covenants and reviews are handled for the life of the loan.
FINDING 3
Expect the security review on day one
Data security, general-purpose models and accuracy topped the questions lenders asked. Put them to every vendor in writing, early.
FINDING 4
Test any tool on your own credit policy
Eleven institutions asked whether AI can read a credit policy at all. Ask to see it work on your policy and your files.
FINDING 5
Compare against the general model you already use
Many teams already use a general-purpose model or are building their own. A vendor should show what it adds to that, on your work.
FINDING 6
Plan for a shop that does many things
C&I and diversified books outnumbered CRE-only ones. A tool that fits one loan type covers part of the work.
The figures on this page, in one table for citation
Each figure in the wording the published study uses. Counts are institutions among the 90 we spoke with.
SAMPLE
90 institutions over nineteen months
49 banks, credit unions and CDFIs; 24 private credit and non-bank lenders; 13 intermediaries; 4 other.
THE STACK
25 and 14
25 described underwriting in Excel and Word; 14 have no loan origination system or spreading platform.
ANALYST TIME
3 institutions
Put 60% to 90% of analyst time on data entry.
BOTTLENECKS
13, 13 and 18
Memo writing (13) and pre-screen triage (13) as bottlenecks; post-close monitoring (18) as a blind spot.
FUNNEL
7 institutions
Median close rate about 5%.
TURNAROUND
6 institutions
Package to memo from 48 hours to three weeks.
QUESTIONS
21, 20, 18, 17, 13, 11, 10
Data security; why not a general-purpose model; hallucination and accuracy; SOC 2, NDA or references; training a shared model; explainability; the regulator.
POLICY
11 and 6
Asked whether AI can read a credit policy (11); named policy exception detection as most impressive (6).
ADOPTION
18, 16, 22 and 8
General-purpose model in informal use (18); internal AI build (16); vendor fatigue or failed implementation (22); board or executive mandate, all dated 2026 (8).
LOAN MIX
Roughly 40 answered
16 diversified or C&I-heavy; 4 CRE-only.
CITE AS: LENDERBOX, "WHAT 90 LENDING INSTITUTIONS TOLD US ABOUT AI AND CREDIT WORK," PUBLISHED SEPTEMBER 7, 2026. FIGURES DESCRIBE THE INSTITUTIONS WE SPOKE WITH.
Questions about the study
How many lenders were in the LenderBox study?
90 lending institutions, over nineteen months of conversations from February 2025 to September 2026. Forty-nine were banks, credit unions and CDFIs, 24 were private credit and non-bank lenders, 13 were intermediaries and 4 were other lenders.
Is the sample representative of US lenders?
No. It is Texas-weighted, weighted toward commercial real estate, and built almost entirely from warm introductions. Every figure is a count among the institutions we spoke with, never an industry average.
What did lenders ask about AI most often?
Data security (21 institutions), why not just use a general-purpose model (20), hallucination and accuracy (18), and a SOC 2 report, an NDA or references (17).
Where do lenders say underwriting work stalls?
Thirteen institutions named memo writing as the bottleneck and 13 named pre-screen triage. Eighteen named post-close monitoring as a blind spot.
How many lenders has LenderBox spoken with since the study?
More than 100 lenders in two-way conversations over 21 months, more than half of them banks and credit unions. The study figures cover the 90 institutions in the census, which closed on September 3, 2026.
Can we cite these figures?
Yes. Cite LenderBox, "What 90 lending institutions told us about AI and credit work," published September 7, 2026, and describe each figure as a count of the institutions in the study.