Defined attributes can be extracted across annual volume while lawyers handle interpretation exceptions.
Collections: Human still decides · Queue eliminated · Regulated autonomy
Executive brief
Constrain AI to a defined extraction schema.
AI value · Processing effort
Company-reported150 attributes from 12,000 agreements extracted in seconds
Company upper-bound estimate; seconds is machine extraction, not end-to-end disposition.
Before
Read commercial-credit agreements and manually identify relevant clauses and attributes. → Use interpreted terms in servicing processes and correct interpretation errors.
After
Uses unsupervised machine learning to extract 150 defined attributes from agreements. → Review exceptions and retain legal interpretation and servicing decisions.
Human boundary
Humans own legal interpretation and downstream servicing.
Why it matters
Defined attributes can be extracted across annual volume while lawyers handle interpretation exceptions.
This case is company-reported. Use it for the operating-model shift; do not treat the numbers as independently measured.
Before
How the work ran before the change.
Lawyers and loan officers
Read commercial-credit agreements and manually identify relevant clauses and attributes.
ControlApproximately 12,000 annual agreements; company estimated up to 360,000 aggregate hours.
Loan-servicing staff
Use interpreted terms in servicing processes and correct interpretation errors.
ControlHuman legal interpretation and servicing controls.
What changed
Defined attributes can be extracted across annual volume while lawyers handle interpretation exceptions.
Decision rightAI handles the default; humans own exceptions
After
How the same work runs now.
COiN
Uses unsupervised machine learning to extract 150 defined attributes from agreements.
ControlDefined attribute schema; machine extraction reported in seconds, not end-to-end disposition.
Lawyers and loan officers
Review exceptions and retain legal interpretation and servicing decisions.
ControlUnknown or low-confidence clauses require human review; no public error-rate threshold.
Exception path
Unknown clauses or low-confidence extraction require human review.
Decision authority
Humans own legal interpretation and downstream servicing.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Lawyers and loan officers | Read commercial-credit agreements and manually identify relevant clauses and attributes. | Approximately 12,000 annual agreements; company estimated up to 360,000 aggregate hours. |
| 02 | Loan-servicing staff | Use interpreted terms in servicing processes and correct interpretation errors. | Human legal interpretation and servicing controls. |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | COiN | Uses unsupervised machine learning to extract 150 defined attributes from agreements. | Defined attribute schema; machine extraction reported in seconds, not end-to-end disposition. |
| 02 | Lawyers and loan officers | Review exceptions and retain legal interpretation and servicing decisions. | Unknown or low-confidence clauses require human review; no public error-rate threshold. |
Work that left the path
Human role before
Lawyers and loan officers read commercial-credit agreements, extracted relevant clauses and attributes, and handed interpreted terms into loan-servicing processes under legal review.
Human role after
Legal and loan staff review exceptions and own servicing interpretation.
AI role
Uses unsupervised ML to extract 150 defined attributes and clauses.
As many as 360,000 hours/year150 attributes from 12,000 agreements extracted in seconds
Initial 2016 implementation · About 12,000 commercial-credit agreements annually
Company upper-bound estimate; seconds is machine extraction, not end-to-end disposition.
Anti-pattern
Do not imply complete lawyer replacement.
Questions
Portability conditions
Reputation risk
low