Use casesFinancial institutions & fintech
Corporate credit memo and financial statement analysis
Analysts spread statements and write the memo by hand, in a fixed format, for every renewal.
- Kinds of AI
- SLMs
- Who decides
- Credit analyst
- Where it runs
- Your servers, banking data kept in Türkiye
SLM · draft sections
- met3. Financial analysis: 12 figures, all source-linkeddrafted
- flagged4. Repayment capacity: interim not auditeddrafted
- met6. Risks: foreign-currency debt vs. lira revenuedrafted
- Figure check1 mismatchinventory in note 9 does not match the balance sheet line; sent to the analyst
- SLM
Why a person decidesThe analyst edits and signs; the credit authority decides the loan.
DecidesCredit analyst
1. The desk
Today
- ReadAudited statements, interims, and notes arrive as PDFs and scans; the analyst retypes them into the spread template.
- WriteThe memo follows the bank's format section by section, and most of each renewal repeats last year's with new numbers.
- SignReviewers check figures against the PDF by hand before the analyst signs; a typo in a ratio can reach committee.
With the system
- ReadThe SLM spreads the statements into the bank's template, each line tied to its source and page; ratios and covenant tests are computed in code with the bank's own formulas.SLM
- WriteThe same SLM drafts each memo section in the bank's format with every figure linked to its source, and a self-check flags mismatches; the proposal is left to the analyst.SLM
- SignThe credit analyst opens the flagged figures, edits, and signs, as before; the credit authority decides the loan.Person decides
What lands on the desk
The analyst opens a spread and a memo draft in the bank's own format, each figure one click from its page in the statements and each ratio from the bank's own formula. Mismatches and unaudited figures are flagged, and the proposal is left to the analyst.
The analyst edits and signs the spread and the memo; the credit authority decides the loan. The model drafts only; it writes no proposal and no recommendation.
SLM · spread
- PeriodsFY2025 audited · H1 2026 interim, unaudited
- Net salesUp year on yearFS-2025 p.6 l.3
Code · ratios, the bank's formulas
- EBITDA marginrule RT-02lower
- Net debt / EBITDArule RT-05above last year
- metDebt service cover: thinner headroomcovenant met
SLM · draft sections
- met3. Financial analysis: 12 figures, all source-linkeddrafted
- flagged4. Repayment capacity: interim not auditeddrafted
- met6. Risks: foreign-currency debt vs. lira revenuedrafted
- 7. Proposalleft to the analyst
- Figure check1 mismatchinventory in note 9 does not match the balance sheet line; sent to the analyst
- SLM
- Code
Why a person decidesThe analyst edits and signs; the credit authority decides the loan.
DecidesCredit analyst
2. The system
For your technical team — skip it if you like.
How it works
In
- Audited statements and notes
- Interims and the auditor's report
- PDFs or scans
SLM · understands
Statement spreading
Schema-guided extraction into the bank's spread template
Reads
- Everything under In
Out
- Each line with source, page, and quote
- Unaudited figures marked
Code · calculates
Ratio calculator
The bank's own formulas, in deterministic code
Reads
- The spread
- Last year's spread
Out
- Ratios and trends
- Covenant tests
- Computed in code, never by the model
SLM · understands
Memo drafting
The same SLM: one task per section of the bank's format
Reads
- The spread and the ratios
- Last year's memo
- Notes
Out
- Company, group, financials, cash flow
- Repayment capacity, collateral, risks
- Every figure linked to its source
SLM · understands
Figure-to-source check
The same SLM: a self-check pass over the draft
Reads
- The draft
Out
- Mismatches flagged for the analyst
Decision
Credit analyst
Reads
- The draft and its flags
Out
- The analyst edits and signs; the proposal section is left to the analyst
None: one SLM with fixed tasks, called from the bank's credit workflow, is enough: one kind of AI on its own, with no agent and no new machine-learning model.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Small language model (SLM)
Test set
100–200 past memos and spreads graded by senior analysts before training: field-level spread accuracy, figure-to-source match, and section quality on a written rubric.
Dataset
Past statements paired with the analyst's final spread and the signed memo, with personal data removed.
Continued pre-training (CPT)If needed
On financial-reporting language (TFRS statements and notes, Turkish accounting terms), only if the base model fails vocabulary items on the test.
Supervised fine-tuning (SFT)
On spreading, and on each memo section in the bank's format.
Reinforcement learning (RL)If needed
Only with a checkable reward: every figure in the draft matches its cited source and the ratio calculator. Skipped if SFT already passes the test.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Analyst hours per renewal.
- Fewer figure errors reaching committee.
- Consistent memos across analysts.
What your institution measures
- Analyst hours per memo
- Spread field accuracy
- Figure-to-source mismatches caught before committee
- How much analysts edit the drafts
- Committee return rate
Moody's reports that Taishin Bank cut its credit report production time from 20 hours to 4.
Moody's, "How Taishin Bank used AI to amplify human expertise in credit analysis" · vendor-reported
Measured in the Decision Blueprint
A time study of 20–30 recent renewals sets the baseline; your institution sets the target, confirmed on held-out memos.
Conditions
The rules, the data, and the known error, with its guard.
- On premises, under the banking regulator's (BDDK) rules and KVKK Art. 9.
- Scanned statements need text-recognition (OCR) quality checks.
- Over-reliance: the analyst must open the source of every flagged figure.
- Unaudited figures are always labeled.
Start here
We time your analysts on recent renewals and test the spreading task on your own statements, scans included, before anything is built.
Similar desks
- SLMs
TodayThousands of messages a day land in one queue; misroutes miss legal deadlines.
With the systemEach message routed to the desk that owns it; the model never closes one.
Head of operations decides
- ModelsSLMsAgentic AI
TodayReferred files wait while analysts rebuild the firm from four systems and last year's committee notes.
With the systemOne file with the rating, its reasons, and the authority level.
Credit manager decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/corporate-credit-memo · Printed copy; the website is the current version.