SearchTrading & Execution

Evaluate · software and agents

Derivatives documentation software: the agents and what none may do

Derivatives documentation software either stores agreements, abstracts them into fields, or reads them on demand. Derivatives Documentation Search is the third kind: eight AI agents, each with a job and a written list of things it may never do, and two lawyers who sign every close-out answer.

Updated · 13 min read

See one case, screen by screen →
Watch with transcript →

Derivatives documentation software helps legal, risk and operations teams answer questions about their ISDA Master Agreements, Schedules, credit support annexes, repo and prime brokerage agreements and the amendments to them, without reading every document by hand. Products fall into three kinds: repositories that store and index executed documents, abstraction tools that turn key terms into fields at signing, and reading systems that read the documents themselves when a question is asked. Each fails differently, and a buyer should know which kind is on offer. Botminds' Derivatives Documentation Search is a reading system built from eight AI agents: they plan the reading, read every document in its own language, apply amendments in order, flag documents that disagree, read every cited value a second time, join exposure, draft notices and keep the estate filed, and each has a written list of things it may never do. Senior counsel verifies each row and a second lawyer approves any answer used in a close-out.

The screens and figures below come from the solution's fictional sample estate, Quillfield Capital and the Hollins Brothers Bank group, and its sample settings. Regulatory texts are real and linked.

Eight agents, three people: who reads, who checks, who decides

  1. Agent

    Intake agent

    Files new executed documents from the document store, email archive and e-signature records into their agreement families; flags missing signature pages and poor scans.

    Never: Overwrites an executed copy, or files a scan below 150 dpi; it goes back to the sender.

  2. Agent

    Planner

    Turns a plain-language question into a reading plan: document families, clause types, fund entities, as-of time.

    Never: Answers itself, or reads before an ambiguous question is settled; it asks back first.

  3. Agent

    Reader

    Reads every planned document in its own language, scans included, and finds the clause and value with page and paragraph.

    Never: Infers a value that is not written; when nothing is found, it says where it looked.

  4. Agent

    Amendment agent

    Orders each family's chain and applies the latest effective change, keeping the original value visible.

    Never: Treats an unsigned or one-sided document as effective.

  5. Agent

    Conflict agent

    Flags documents that disagree, searches the email archive, proposes a resolution with both passages side by side.

    Never: Resolves a conflict without a written rule; everything else goes to a lawyer.

  6. Agent

    Checker

    Re-reads every cited passage blind, on a different model family, and confirms or disputes the value with a score.

    Never: Approves anything; a close-out reading below 0.90 goes to a lawyer.

  7. Agent

    Exposure agent

    Joins mark-to-market and collateral per agreement from the collateral system and proposes a decision by the firm's rule.

    Never: Decides, or shows exposure to roles not entitled to it.

  8. A person decides

    Senior counsel, then Deputy General Counsel

    Verifies each row, decides the flagged ones with a reason; a second lawyer approves the answer.

    Never: Approves an answer they prepared.

  9. A person decides

    Head of counterparty risk

    Records a decision per agreement, by the firm's rule or with a reason; the Deputy General Counsel approves it.

    Never: Approves their own decisions.

  10. Agent

    Notice agent

    Drafts Early Termination notices and reservation-of-rights letters to the latest address, by a permitted method, every reference cited.

    Never: Sends; drafts stay locked until two lawyers sign.

The agents, their limits and the people who decide, as set in the fictional sample firm's configuration of Derivatives Documentation Search. Legal operations sends signed notices; internal audit reads the trail.

What does derivatives documentation software do?

It answers questions about executed agreements faster than reading them, and the three kinds do it in different ways. A repository finds documents; an abstraction tool returns the fields someone filled in; a reading system returns what the documents say now. The choice matters most on the questions nobody anticipated and on agreements amended after the abstract was made.

Three kinds of derivatives documentation software, judged on a downgrade morning

Repository or abstraction toolReading system with agents
What you get back

Repository or abstraction tool: Documents to read, or fields keyed in at signing.

Reading system with agents: An answer per agreement, each value cited to its page.

Amendments

Repository or abstraction tool: Only if someone filed and re-keyed them; abstracts drift from the documents.

Reading system with agents: Read and applied in date order every time, with the original shown.

A question nobody set up

Repository or abstraction tool: No field, so no answer; back to reading by hand.

Reading system with agents: The Planner builds the reading from the question; the answer arrives as a new column.

German or French masters

Repository or abstraction tool: Often abstracted from a translation, or not at all.

Reading system with agents: Read in the original, with the English reading beside it.

What can go wrong

Repository or abstraction tool: A confident field that no longer matches the executed document.

Reading system with agents: A misreading; which is why every value is read twice and low scores go to a lawyer.

Our comparison. Many firms combine kinds: a repository remains the place executed documents live, and a reading system reads from it.

Where Derivatives Documentation Search fits, and where it does not

It fits after signing, when the questions are "what does every agreement with this counterparty say, as amended?" and "who may terminate, on what terms, with what exposure?". It does not negotiate agreements; drafting counters and tracking concessions is the job of Botminds' ISDA Negotiation solution, whose agents are described in ISDA AI tool: agents draft the counter, lawyers make the concession. It does not keep the collateral system's margin data or make margin calls; Botminds' ISDA and CSA Data Extraction and Margin Call Management solutions do that. And it does not replace the collateral system: the Exposure agent reads mark-to-market and collateral from it, for entitled roles only.

Supervisors have asked for this kind of automation for a long time. The interagency guidance tells banks to "Automate processes to track and manage legal documentation, especially when there is a large volume of legal agreements" (SR 11-10 guidance). The guidance does not say how; the rest of this page is about how to do it without handing judgment to a machine.

Can derivatives documentation software read scanned and foreign-language agreements?

It must, because an estate's hardest rows are scans and local-law masters. In the sample the Reader reads every planned document in its own language, scans included, and shows German and French originals beside the English reading. The Hollins estate holds 214 documents in English, German and French; one German Rahmenvertrag has a rating trigger typed as "BBB-", struck by hand and replaced with "BBB", initialled by both parties. The Reader scored its reading 0.84 and the blind Checker 0.86, both below the 0.90 a close-out answer needs, so the row waited for a lawyer even though both agents read "BBB".

Scan quality is set as a rule, not left to chance: the Intake agent sends scans below 150 dpi back to the sender. A system that silently reads a poor scan produces a confident value nobody can verify; one that refuses it produces a request for a better copy.

The agents at work: a plan first, then each agent reports as it finishes

Ask screen during reading, with the question Hollins Brothers Bank was downgraded to BBB-. Which agreements let us terminate? A progress panel shows 152 of 214 documents read, English 150, German 1, French 1, amendments applied 60, while reading Currency Alpha Fund with Hollins. The agents panel lists the Planner (61 families, rating-downgrade clauses, latest amendments, 0.7 seconds), the Reader (214 documents in English, German and French, 2.1 seconds), the Amendment agent (84 amendments and 3 protocols, 3 triggers changed in 2023), the Conflict agent (2 found, 1 resolved by rule, 1 for a lawyer), the Checker (213 of 214 agree, 1 below 0.90) and the Exposure agent (joined from the collateral system).
Derivatives Documentation Search on the fictional sample estate. Each agent's line is a claim a lawyer can check: how many documents, which amendments, which conflicts, which readings fell short.

How does derivatives documentation software handle amendments?

Properly, only by reading the chain every time. The Amendment agent orders each family's documents, applies the latest effective change, and shows the original value struck through beside the term in force, citing the amendment. Its rule is a written setting: latest effective amendment wins, protocols apply from their effective date, initialled handwritten changes count. It never treats an unsigned or one-sided document as effective. The guide to ISDA amendment chains explains why each of those choices is there.

Can derivatives documentation software send termination notices?

It should not, and this one does not. The Notice agent drafts Early Termination notices and reservation-of-rights letters to the address in the latest notices clause, by the delivery methods each agreement allows, with every reference cited; the drafts stay locked until senior counsel and the Deputy General Counsel have both signed, and legal operations sends them. The table sets out each agent's limits and the risk each limit closes.

The 'never' list, and the risk each one closes

AgentWhat it may never doThe risk it closes
Notice agentSend a noticeAn irreversible step taken without the two people accountable for it
Amendment agentTreat a one-sided amendment as effectiveA termination right claimed on a document the counterparty never executed
Conflict agentResolve a conflict without a written ruleA silent choice between two versions of an agreement
CheckerApprove anythingA machine certifying its own colleague's work
Exposure agentDecide, or show exposure to roles not entitled to itA decision made by arithmetic; exposure leaking to staff who do not need it
PlannerRead on a guess when a question is ambiguousA confident answer to a question nobody asked
Limits as set in the fictional sample firm's configuration of Derivatives Documentation Search. Agents also never change a value without a person; corrections and threshold changes go to a second person.

The Financial Stability Board's final sound practices on AI put unauthorised actions first among the risks of agentic AI: an agent working toward a goal by its own route may do something illegal, unethical or simply unapproved before any person has looked (FSB, Sound Practices for Responsible Adoption of AI, 15 October 2026). A notice that terminates billions of dollars of trades is the clearest case of an action that should never be within an agent's reach.

Who approves an answer from derivatives documentation software?

A person who did not prepare it. In the sample, senior counsel verifies each row, deciding the flagged ones with a reason code, and the Deputy General Counsel approves any answer used in a close-out; the approve button is locked for the person who prepared it. The same split applies to close-out decisions recorded by counterparty risk. An everyday answer, such as an operations analyst's question about a minimum transfer amount, needs a 0.80 confidence rather than 0.90 and is not used in a close-out unless a lawyer verifies it.

The FSB's sound practice on human oversight explains why the signature has to mean something: oversight counts only when the people doing it have the ability, the authority and the reason to step in; a sign-off given for form's sake does not (FSB, Sound Practice 10). Citations to the page are what give a reviewing lawyer the ability; the locked approve button and the reason codes give the authority and the accountability.

How is derivatives documentation software audited?

Through an append-only trail of every question, reading, decision and signature, and through release tests for each agent version. US rules already expect documentation controls to be audited: a swap dealer must have an independent internal or external auditor audit its documentation policies and procedures periodically, often enough to catch any material weakness (17 CFR 23.504(c)). A fund using agents needs the same evidence for its own processes, and for its auditors and board.

Every agent's version, model and test result before release, and the thresholds only the Deputy GC can change

Settings screen, Agents and thresholds, for the administrator: version 14 in force and version 15 awaiting approval. A table lists each agent with version, model, threshold and last test before release: Planner v2.3, plans match counsel on 240 of 250 questions; Reader v3.4, 0.90 close-out and 0.80 everyday, 98.1% on 1,200 lawyer-checked clauses; Amendment v1.9, 99.2% on 310 amendment chains; Conflict v1.4, all 64 seeded conflicts found; Checker v2.1 on a second model family, caught 46 of 48 seeded misreadings; Exposure v1.2, rule D1 $15 million, reconciled to the collateral system daily; Notice v1.1, 60 of 60 test letters correct address and method; Intake v1.6, 150 dpi minimum, 99.6% of 2,000 filings linked correctly. Confidence needed for a close-out answer 0.90; for an everyday answer 0.80, with 0.85 proposed and awaiting Daniel Okafor. A note says changes to thresholds, rules and roles go to the Deputy General Counsel for approval.
Derivatives Documentation Search, sample settings. The release tests are run on seeded and lawyer-checked sample data before each version goes live; they are design checks, not measured customer results.

What the trail holds in the sample

41,286

trail entries since 1 October in the sample estate, people and agents alike

10 years

retention set for the trail and every export

2

lawyers on every close-out answer: one verifies, one approves

0.90

confidence a reading needs before a close-out answer can use it without a lawyer deciding the row

Fictional sample estate and settings in Derivatives Documentation Search. Every entry carries the agent version that produced it and the reason given by any person who acted on it.

The trail: who asked, who read, who decided, and why

Audit trail screen, append-only, showing the 26 most recent of 41,286 entries since October 1, retained 10 years, with an export button. Entries on October 19 include Priya Raman's what-if question at 9:13, her answer to the Planner choosing to read below investment grade as each agreement's own words at 9:12, the Planner agent asking back before reading at 9:11 with its version, her set-off question at 9:10, Daniel Okafor's approval at 9:09 of designating an Early Termination Date on 4 agreements and reserving rights on 5 with the reason Verified against source and the event evidence, and Hannah Cho recording the decisions at 9:05 with the reason Committee instruction.
Derivatives Documentation Search on the fictional sample estate. The Planner's question back and the lawyer's answer to it are entries like any other.

Does derivatives documentation software replace the collateral system?

No, and it should not try. The collateral system calculates and moves margin; the documentation software reads what the agreements say about it. They meet in two places: the Exposure agent reads mark-to-market and collateral per agreement to size decisions, and operations corrections to collateral terms, such as a minimum transfer amount, go through a maker and a checker before anyone relies on them. The useful division is that the documents remain the source of the terms, and the system remains the source of the numbers.

What should you ask a derivatives documentation software vendor?

Ask for evidence on your own documents, not a demonstration on someone else's. The checklist below is the set of questions we would ask any provider, including ourselves; the guide to measuring legal document search AI turns the accuracy questions into a test you can run.

Eight questions for any derivatives documentation software

  1. Does every value cite its document, clause and page?

    And does an empty answer list where the system looked?

  2. How are amendments, side letters and protocols applied?

    In what order, from what date, and is the original value shown?

  3. What happens with a one-sided or unsigned amendment?

    The only acceptable answer is that a person decides.

  4. Is every reading checked independently?

    By a second reading, ideally on a different model, with a threshold below which a lawyer decides.

  5. Can it answer a clause question nobody set up?

    Without a new field or a project; and does it ask back when the question is ambiguous?

  6. Can it act on its own?

    Send, approve, change a value: each should be impossible for the software.

  7. Who sees exposure, and who can approve?

    Entitlements by role, fund entity and region; maker and checker enforced, not suggested.

  8. What does the audit trail hold, and for how long?

    Questions, readings, versions, decisions, reasons and signatures, append-only, exportable for auditors.

Our buyer's checklist. In a pilot, put the same questions to the software that your lawyers have already answered by hand, and compare the citations as well as the answers.

To see the agents on the sample estate, start from Derivatives Documentation Search.

Questions people ask

Is there free derivatives documentation software?

Repositories and generic document search can be cheap or bundled with other systems, but they return documents rather than answers. The cost that matters is the lawyer time spent reading after the search: by our estimate about four lawyer-days for one question across one dealer group's 61 agreement families.

How long does it take to go live?

The solution's first-time setup connects the document store, email archive, collateral system and rating feed, maps fund entities and regions, and runs a first full read of the estate. A pilot on one dealer group, tested against questions your lawyers already answered, is the sensible first step.

Can AI agents be trusted with legal documents?

With reading, under checks; with decisions, no. The design that earns trust is the one where every value is cited, read twice, and decided by a person when the readings disagree or fall short, and where the agents cannot send, approve or change anything.

Does the software work for repo and prime brokerage agreements?

Yes. The Reader reads GMRAs, prime brokerage terms and local-law masters alongside ISDAs; in the sample estate it reads 14 GMRA 2011 agreements when asked about close-out under the repo agreements.

What data does it need?

The executed documents from the legal document store, the email archive for countersigned copies and notices, the collateral system for exposure, the counterparty database and rating feed for entities and ratings, the industry protocol adherence list, and NAV by fund entity from the fund administrator for NAV-decline tests.

Derivatives Documentation Search.
See it on your own files.

An agentic solution from Botminds. Ask one question of every ISDA, CSA, repo and prime brokerage agreement at once; agents read each in its own language, apply amendments and protocols in order, flag documents that disagree, read every value twice and return a grid of the whole estate, each cell cited to its clause. Book a demo, or explore how it works.