ISDA data extraction you can question, check and sign: every credit support annex term in force, cited to its clause
Collateral and risk systems run on the terms in force, each one cited to its clause.
See one case, screen by screen ↓ISDA and CSA Data Extraction is ISDA data extraction software for collateral, legal and risk teams, run by seven AI agents you work with in plain words. They read every executed ISDA Master Agreement and Schedule, credit support annex (1994 and 2016 CSAs, IM CSAs), GMRA and GMSLA with every amendment and protocol letter, in the language it was signed in, and turn them into the terms in force today: CSA digitization with each value cited to its clause. Legal confirms the terms, and the collateral system and risk engine are corrected only when their owners approve.
AI in asset management: a buyer's guide to agentic AI for investment managers →
Updated
ISDA data extraction is where every margin call starts. The call is computed from a handful of numbers in the credit support annex — the threshold, the minimum transfer amount, the independent amount, eligible collateral and haircuts, the valuation agent, the notification time and rounding — plus the NAV triggers in the ISDA Schedule. Those numbers were keyed by hand into the collateral management system and the risk engine when the agreement was signed, and every amendment since has to be re-keyed by someone who knows it exists.
Nothing checks the systems against the paper again until a call is disputed or an auditor samples a file. An amendment signed as an emailed PDF and filed beside the Schedule, a protocol adherence letter never linked to the agreement it amends, a haircut in a German or French annex keyed from the English cover page: each one leaves a system computing calls on yesterday's terms. The uncleared margin rules made it harder — ISDA reported that their final phases significantly increased margin calls, settlements and collateral disputes. And machines alone are not the answer: in ISDA's own test, language models read CSA clauses well but rarely perfectly, so the reading has to be checked twice and signed by a person.
Estimated split of the sample firm's 3 h 40 min median per agreement by hand, against the 12-minute target with the solution. Hours.
Seven agents read twice, apply amendments, compare with systems and size drift; legal confirms, system owners approve.
Quillfield Capital runs 2,400 executed ISDA, CSA, GMRA and GMSLA agreements and 9,000 amendments and protocol letters with 46 counterparties, across 318 funds and mandates. In March 2023, Amendment No. 3 cut Brennmoor Bank AG's threshold on the Strategic Income Fund's 1994 CSA from USD 10,000,000 to zero. Nobody told collateral operations. Here is how the agents find it, answer the questions people ask, and wait for legal and the system owner to decide, screen by screen, in the working solution.
Hannah's home screen shows 14 agreements in legal review, $41.2M of open drift and 9 loads waiting for system owners, each with one next step. First on her list: “Brennmoor × Strategic Income Fund: threshold changed, systems did not.” Beside it, what the agents did since Friday: the Drift agent compared 62,410 terms in force with three systems at 11:40 p.m. on Sunday, and the Extraction and Checker agents finished the last back-catalog batch — 312 legacy agreements, 8,112 terms, the two readers agreeing on 96.8%.
“Agents read, compare and size. They never change a system value: legal confirms the terms in force and the system owner approves every load.”
On the drift page the Drift agent has already left its summary: 37 open, 4 new, $41.2M at stake, each sized from the risk engine's exposure, starting with Brennmoor. Hannah asks which drifts are over $1 million. Nine, $28.2M together: Brennmoor × Strategic Income Fund, threshold, $6,820,000; Kestrel Harbour × Global Macro Opportunities, ineligible collateral, $4,950,000; Nordhaven × Asia Bond Fund, fund threshold, $3,200,000. The agent filters the list to match.
“I never correct a system myself: legal confirms, the owner approves the load.”
Collateral elections down the side, Brennmoor's eight agreements with Quillfield funds across the top: a 1994 CSA, two 2016 VM CSAs, a 2018 IM CSA, a GMRA 2011, a GMSLA 2010, and German and French collateral annexes — 84 terms in force, 7 needing attention. Each cell is the term in force today with a clause chip. One cell stands out: zero in the agreement, USD 10M in the collateral system, marked Drift.
The rail: “At stake $6,820,000 · Delivery Amount not called.” Collateral system and risk engine both hold USD 10,000,000.
A clause chip opens the passage itself. For the Euro Credit UCITS fund's Besicherungsanhang, Nr. 4 Abs. 2 on page 6 sets a “Bewertungsabschlag von 2 %” — a 2% haircut on German government bonds of five to ten years — shown beside the English and beside UCITS guideline 4.2, which asks for at least 4%. The Policy agent has already marked it as outside the guideline.
Every value is read in the original language, and opening a source is logged.
On the agreement page, Hannah asks the agents. The Drift agent answers with four citations: the record was keyed on June 30, 2014 from the original annex and never changed; Amendment No. 3 replaced USD 10,000,000 with zero, but no load request was raised; the risk engine has the same gap. Below, the terms in force list each election with the document it came from and how strongly the two readers agreed.
Answers cite the page; the agents act only with a person's say-so.
The 2002 Master Agreement and Schedule, the 1994 CSA, Amendment No. 1 (minimum transfer amount USD 250,000, notification 11:00 a.m. New York), the 2018 U.S. Resolution Stay Protocol adherence letter, the 2020 IBOR Fallbacks Protocol adherence letter, Amendment No. 2 (Federal Funds to SOFR on cash collateral) and Amendment No. 3. The stay protocol letter was filed in 2022 under the Schedule folder; the agent places it by its effective date.
When two documents conflict, the Amendment agent stops and asks legal instead of choosing.
Exposure to Brennmoor at Friday's close: $6,812,400 across 41 trades, with no collateral held. With a zero threshold, the Credit Support Amount is $6,812,400; it exceeds the $250,000 minimum transfer amount and rounds up to the nearest $10,000 under ¶13(b)(iv)(D): $6,820,000. The collateral system, working from $10,000,000, called nothing. Since the amendment the exposure was above the minimum on 418 business days, peaking at $9.42M on June 3, 2025.
“The agent sizes; it never makes a call. The call follows only after legal confirms and collateral operations approves the load.”
Hannah picks the reason code “Amendment executed and in force” — signed by both parties, effective date passed — and ticks “I have read the cited clause and the amendment chain.” The Load agent prepares two loads, one to the collateral system and one to the risk engine, each to be approved by its owner. Nothing changes in a system yet.
Ruth's loads page opens with the Load agent's note: 7 loads wait for her, largest first, and approving three of them before 11:00 a.m. New York adds $10,840,000 of calls to today's cycle. She opens LD-2026-0327 — threshold, dealer: USD 10,000,000 to zero, confirmed by Hannah Brooks, a delivery call of $6,820,000 on the 11:00 a.m. New York cycle — and approves with “Matches the agreement in force, confirmed by legal.”
The person who confirms the terms never approves the load.
The Load agent checks legal's confirmation and Ruth's approval, pushes zero to the collateral system, keeps the earlier record as a version (version 3 to 4, restorable), reads the value back — zero, matching the agreement in force — and hands it to the next margin cycle: a $6,820,000 delivery call to Brennmoor at 11:00 a.m. New York. Every step lands on the audit trail with its fingerprint.
The Load agent never writes without legal's confirmation and the owner's approval, and never overwrites without a version.
The same Brennmoor amendment, seen by five of the people who carry it: what it looked like by hand, and what it looks like now.
Seven agents read twice, apply amendments, compare with systems and size drift; legal confirms, system owners approve.
Watches the legal email intake, dealer portals, the negotiation platform export and the document repository; classifies each document, matches it to its agreement by parties, dates and references, splits bundles by signature page and spots duplicates by content.
Reads ISDA Schedules, 1994 and 2016 CSAs, IM CSAs, GMRAs, GMSLAs and local-law masters, and extracts the collateral elections and Schedule NAV triggers with document, paragraph and page, in the original language with the English beside it.
Reads the same clauses blind, with a second model family and a different method, and reports that both readings agree or shows the two side by side with both passages.
Collects every document in an agreement's chain, orders amendments and protocol adherence letters by effective date, applies each change and records which document and paragraph set every term in force.
Compares each term in force with the negotiation guideline for the fund type — '40 Act fund, UCITS fund, hedge fund, separate account, private credit fund — tracks approved exceptions and raises expired ones again.
After the risk engine's end-of-day run, compares every term in force with the collateral system, the risk engine and the counterparty database, and sizes each difference with the latest exposure, threshold, minimum transfer amount and rounding.
Prepares a corrected load for each confirmed difference, pushes it only after the system owner approves, keeps the earlier record as a version and reads the value back from the system.
Verify, confirm, approve. The agents propose; a named person decides.
Every screen has a place to ask in plain words — on the home screen, beside a cell in the election grid, on an agreement, on the drift list and on the loads page. Every answer cites the clause or the system record, and the agents act only with a person's say-so.
Which drifts are over $1 million?
Nine drifts are over $1 million, $28.2M together. The largest: Brennmoor × Strategic Income Fund, threshold $6,820,000; Kestrel Harbour × Global Macro Opportunities, ineligible collateral $4,950,000; Nordhaven × Asia Bond Fund, fund threshold $3,200,000. Filtering the drift list.
Which NAV triggers are tighter than our guidelines?
Four one-month NAV triggers are tighter than their guideline. One is under an approved exception that expires December 31: Brennmoor × Global Macro Opportunities at 20% against at least 25%. Three '40 Act funds, at 15%, 20% and 22.5%, are open.
Who approves the load?
Two loads, two owners. After legal confirms, Ruth Okafor approves the collateral system and Marcus Lindqvist the risk engine. I push only after each approval, keep the 2014 record as a version and read the value back.
Why was LD-2026-0323 escalated?
LD-2026-0323 carries $2,400,000 of under-called exposure and has waited more than 4 hours since Yuki Sato confirmed it. The rule in Settings sends anything over $1 million to Grace Whitfield after 4 hours; she is watching, but you still approve it.
13 screens from the working solution, on its sample data. Pick one to see it large.
The confidence levels the agents work to, drift materiality and your negotiation guidelines are settings. A change is proposed by one person, approved by another, and recorded.
| Setting | Default | Choose from |
|---|---|---|
| Both readers must agree | On · locked | Locked on; any disagreement goes to a person |
| Shown as agreed at or above (still verified by a person) | 95% | A percentage; a change needs a second approver |
| An analyst re-reads the page below | 85% | A percentage; a change needs a second approver |
| Intake auto-matches at or above | 90% | A percentage; a change needs a second approver |
| Nothing is ever auto-confirmed | Locked | Every value is verified by an analyst and confirmed by legal before any load |
| Drift materiality | High above $1,000,000 · escalate after 4 hours | High above $1,000,000 · Medium $250,000 to $1,000,000, legal within 1 business day · Low below $250,000, legal within 2 business days |
| Drift sweep | Nightly at 11:40 p.m. New York | After the risk engine's end-of-day run, and on demand |
| Negotiation guidelines | By fund type, versioned | '40 Act fund 5.0 · UCITS fund 4.2 · Hedge fund 3.1 · Separate account 2.3 · Private credit fund 1.4 |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate (the split of the sample firm's 3 h 40 min median per agreement) · “typical” = published figures (ISDA, paper exploring the use of generative AI to extract and digitize CSA clauses, May 15, 2025: eight language models tested on five CSA clauses, several over 90% accuracy with CSA-specific context, 100% rarely achieved on nuanced clauses, the largest gains on the minimum transfer amount and threshold; ISDA derivatiViews, August 19, 2021: the final phases of the uncleared margin rules significantly increased margin calls, settlements and collateral disputes). Quillfield Capital, Brennmoor Bank AG and the other people, dealers, funds and agreements named on this page are fictional: characters and sample data in the working solution.
ISDA and CSA Data Extraction is ISDA data extraction software for collateral, legal and risk teams, run by seven AI agents you work with in plain words. They read every executed ISDA Master Agreement and Schedule, credit support annex (1994 and 2016 CSAs, IM CSAs), GMRA and GMSLA with every amendment and protocol letter, in the language it was signed in, and turn them into the terms in force today: CSA digitization with each value cited to its clause. Legal confirms the terms, and the collateral system and risk engine are corrected only when their owners approve.
The Extraction agent reads each CSA clause and records the value with its document, paragraph and page; the Checker agent reads the same clause blind with a different method. Terms both agree on at 95% or above go to an analyst to verify, and any disagreement goes to a person with both passages side by side. Legal then confirms the terms in force.
ISDA's 2025 paper found several language models above 90% accuracy on CSA clauses with CSA-specific context, but 100% rarely reached on nuanced clauses. That is why the solution uses two independent readers, never auto-confirms a value, and has a person verify and legal confirm. In the working solution's back-test on 50 agreements abstracted by hand, the readers' disagreements caught all 14 agent errors, and the cited pages showed 31 errors in the hand abstraction.
The Amendment agent orders every amendment and protocol adherence letter by effective date, not filing date, applies each to the base agreement and records which document and paragraph set every term. When two documents conflict — for example a side letter signed by only one party — it stops and asks legal instead of choosing.
Yes. Every night the Drift agent compares each term in force with the collateral system, the risk engine and the counterparty database, and sizes each difference with the risk engine's exposure. In the sample estate one sweep found 37 drifts worth $41.2M, ranked by dollars at stake, and you can ask the Drift agent about any of them.
Yes. Any agreement's terms in force export as CSV, Common Domain Model JSON or a term sheet in PDF with every source passage, and the election dictionary maps each election to its ISDA clause, its CDM field and each system's field, so loads land in the right place.
Yes. The analyst verifies, legal confirms the terms in force and the system owner approves each load — never the same person. Agents never verify, confirm or approve, and the Load agent writes only after both approvals, keeps the earlier version and reads the value back.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your agreements, negotiation guidelines and system fields — and it goes live once every quality gate has passed. First-time setup connects your systems, loads the estate, maps elections to system fields and runs a back-test on agreements your team abstracted by hand.
Target 20–30% fewer fails, each fix cited and checked before the market’s cut-off.
1-min demoStandard Settlement Instruction ManagementEvery SSI change proven at the source and called back before a dollar moves.
1-min demoPosition and Cash ReconciliationEvery break explained from its source before the NAV is struck, every fix proved first.
1-min demoWe'll run ISDA and CSA Data Extraction on a sample of your own ISDAs, CSAs and amendments, against what your collateral system holds today.