A problem pattern in credit and compliance, and neuro-symbolic AI as an approach

About this paper

The framing and assertions are my own, except where a source is cited. The appendix gives full citations.

This paper contains no confidential or proprietary information from FIS, any other employer, or any working relationship. I invest in one early-stage regtech company in this field. It is described but not named.

The argument

In several segments I work in, some business problems look alike even though the domains differ. Language models and machine learning help with them, and neither solves them on its own. Neuro-symbolic AI is a class of technology designed for problems like these. Production versions in finance approximate a single-system target, and the symbolic side takes specialized IP to build. The pattern is worth recognizing, and the approach is worth evaluating on its own terms.

The pattern

Across several familiar segments, some business problems share some traits:

• Rule- or contract-bound: the answer is set by written rules or contract terms, with definitions, cross-references, and amendments.
• Repeatable: the same inputs have to give the same answer, today and at the next review.
• Explainable to an outsider: an examiner, auditor, or counterparty will ask how the answer was reached, step by step.
• Document-heavy: the inputs are long documents, often in several versions.

Examples from my own work:

Segment	Decision	What sets the answer	Who asks how
Regtech	Which rules apply to this bank, and which control meets each one	Regulations, guidance, and the bank's own policies	Examiners
Syndicated lending	Whether a borrower passes a covenant test	The credit agreement and its amendments	Agent bank, lenders, auditors
Private credit and funds	Which side-letter terms a fund owes each investor; which loans count toward a borrowing base	Side letters, the LPA, the facility agreement	Investors, lenders, fund auditors

The same shape shows up in other segments: insurance claims against policy wording, tax treatment, and trade documents checked against a letter of credit. There are many more.

Where general approaches fall short on this pattern

Approach	What it does well here	Where it falls short
Language model alone	Reads long, messy documents; drafts summaries	Doesn't guarantee the same answer twice or a path an outsider can check
Classic machine learning	Scores patterns in structured data	Needs labeled outcomes, which are rare here; returns a score rather than a reason
Rules engine	Exact and repeatable	Someone has to write every rule from the text, and rewrite it when the text changes

Each of these approaches covers some of the traits. None covers all of them.

LLMs are, arguably, catching up. Recent evidence, trait by trait:

• Document-heavy: largely handled. On clause identification in commercial contracts, most models performed "at a level comparable to junior legal assistants" (ContractEval, 2025).
• Repeatable: solvable, though not by default. Run 1,000 times with identical input at temperature zero, one large model gave 80 different completions, because server load changes how requests are batched. Batch-invariant processing made all 1,000 identical, at about 1.6 times the run time (Thinking Machines Lab, Sep 2025).
• Rule- or contract-bound: still weak. On detecting planted errors in contracts, the best model scored an F1 of about 63%. On matching contract terms to the statutes they depend on, the best score was below 14% (CLAUSE, EACL 2026).
• Explainable to an outsider: still weak. When reasoning models used a hint to reach an answer, their stated reasoning mentioned it 25% of the time (Claude 3.7 Sonnet) and 39% of the time (DeepSeek R1) (Anthropic, Apr 2025). A model's account of its reasoning is not a reliable record of it.

Neuro-symbolic AI as an approach

Neuro-symbolic AI is a class of technology that combines learning (reading, recognizing patterns) with symbolic reasoning (explicit knowledge and logic that can be checked). It produces answers that are flexible, exact, and explainable.

Production systems in finance today are sophisticated approximations. They pair strong components, and the pairing is where much of the engineering is:

• Droit. FIS bought Droit on 9 March 2026. Its Adept platform encodes regulatory obligations as executable logic and decides in real time whether a trade or product complies with the rules of each jurisdiction (A-Team Insight). Droit added a generative "Decision Decoder" to explain Adept's decisions in plain language (Droit).
• Nammu21. Its NEL Protocol turns credit agreement provisions into structured data. Finastra built it into the Loan IQ Nexus Build API on 5 May 2026 (Finastra). Nasdaq Ventures, State Street, and UBS invested in March 2025, for syndicated loans and private credit (PRNewswire).

The target state is one coherent system, where learning and reasoning work together inside the same architecture rather than side by side. Unlikely AI, founded in 2019 by Alexa co-creator William Tunstall-Pedoe, is building toward that. It describes combining "generative AI, statistical AI, with symbolic algorithmic methods, [and] conventional software methods" for reliability and explainability (TechCrunch, July 2024). Its investor Amadeus Capital describes it as combining "machine learning with the reliability of symbolic computing" (Amadeus Capital).

The distance between the approximations and the target state is itself worth tracking. A system that pairs a model with a rules layer can be exact where the rules exist and fluent where they don't. The open engineering question is how gracefully it handles the space between.

The symbolic side is the moat

The learned side is increasingly available to everyone. The symbolic side takes three things that are slow to build:

• Domain ontologies: a model of what a covenant, an obligation, or a side-letter term is, and how they relate.
• Curated content: rule text and contract language, kept current as it changes.
• People who know both the domain and the technology.

That combination is specialized IP, and it is a moat for the providers who have it.

Two developments show the moat from different sides:

• Content owners are consolidating. CUBE completed its purchase of Thomson Reuters Regulatory Intelligence and Oden in January 2025, bought Kodex AI in October 2025, and bought 4CRisk in February 2026 (CUBE; PRNewswire; CUBE). That puts rule content, change tracking, and agents under one owner.
• Shared standards can narrow the moat. In derivatives, ISDA's Common Domain Model, now run by FINOS and in production at version 7, is a shared machine-readable model of trades and legal agreements (FINOS; A-Team Insight). Where the industry builds the structure in public, less of it is proprietary.

Three cases

	Regtech	Syndicated lending	Private credit and funds
Where the structure is	The chain from rule text to applicability, obligation, control, evidence, and examination	The credit agreement: definitions, covenants, tests, and amendments	Direct loan agreements, side letters, LPAs, and investor reporting
What it buys	Lineage an examiner can walk backward	Faster onboarding now; covenant tests and amendment votes run off the data later	Tracking obligations across documents negotiated one at a time
Reference points	Droit (FIS); CUBE; a small field of AI-native entrants	Nammu21 in Loan IQ; Versana	ILPA templates; few providers at scale
Shared standards	Few	LSTA and LMA forms; Versana's agent-sourced data	ILPA reporting templates and model LPA

Syndicated lending shows two routes to the same data. Versana publishes structured facility data from agent banks through a shared hub, which reduces the need to interpret documents downstream (Versana). Nammu21 creates persistent information objects from the agreement itself. Arguably the second is the more durable and extensible technology approach.

Evolving standards may reduce friction in private credit. AIFMD II applies from 16 April 2026, with leverage caps of 175% of NAV for open-ended and 300% for closed-ended loan-originating funds, and 5% retention on loans they originate and sell (Dechert). ILPA publishes reporting templates and a model LPA (ILPA). How closely LPAs follow the model is not measured.

Evaluating an approach built for this pattern

Questions I'd ask of any technology (or vendor):

1. Given the same inputs twice, does it give the same answer? How does it show that?
2. Can each answer be traced to the clause or rule that drove it, in a form an outsider can follow?
3. Where does the symbolic content come from, who maintains it, and how fast does it change after a rule or contract changes?
4. How do learning and reasoning connect in the system: one architecture, or components passing results to each other?
5. What happens when the text is ambiguous or the structure has no rule for the case?
6. What does the buyer still have to build or maintain?

Contrasting positions

Question	One position	The other position
Will better language models close the gap?	Repeatability is now an engineering choice (Thinking Machines Lab), and document handling is largely solved (ContractEval)	Contract-bound reasoning and faithful explanations remain weak (CLAUSE; Anthropic)
Where does the explanation come from?	The structure: the decision path is the explanation	A model, afterward: Droit added a generative explainer to Adept
Who builds the structure?	Vendors, one firm at a time	The industry, one market at a time: ISDA CDM, Versana, ILPA templates
Is regulation a forcing event?	The EU AI Act treats credit scoring as high-risk, with explanation and documentation duties	The AI Omnibus, in force since 27 July 2026, moved those duties to 2 December 2027 (European Commission)
Where is the value?	In the reasoning structure and its lineage	In the curated content under it
One system or paired components?	One coherent system is the goal, and pairing is a stage on the way	Paired components are good enough for many production needs

Open questions for the reader

1. Which decisions in your own work have these traits, and what are you using for them today?
2. As models improve, does the symbolic layer shrink to a thin audit trail, or does the need for outside checking keep it thick?
3. How long until single-system approaches reach production in finance, and what would prove they're ready?
4. In each field, does the structure end up as a vendor product or a shared standard?
5. Who maintains the structure after the sale, and at what cost?

Appendix

Not sourced

Judgment and framing:

• The traits, and the claim that they recur across my segments.
• The examples table, and who asks how in each case.
• The assessment of where each general approach falls short, and that none covers all the traits.
• The reading of the language model evidence by trait, including that document handling is largely solved.
• That production systems are approximations of a single-system target state.
• That the symbolic side is the moat, and what it takes to build.
• That shared standards narrow the moat.
• The three-case table.
• That Nammu21's approach is more durable and extensible than a data hub.
• That covenant tests and amendment votes could later run off structured agreement data (inference).
• That a structure built for one private-markets deal carries over less well to the next (inference).
• That a data hub reduces the need to interpret documents downstream (inference).
• The evaluation questions.
• Rows 3, 5, and 6 of the contrasting positions table.

Exists, but not used or not found:

• Finastra's claim that 70% of syndicated loans are serviced on Loan IQ. It is a vendor marketing figure with no published method.
• CUBE's customer count, for the same reason.
• Unlikely AI's accuracy claims. They are the company's own and not independently tested.
• Benchmarks of language models on covenant tests or side-letter obligations specifically. I found none.
• How closely fund LPAs follow the ILPA model. I found no measurement.
• Adoption of document-extraction tools by private credit funds. I found no buyer-side survey.

Sources

Claim	Source
Most models comparable to junior legal assistants on clause identification	"ContractEval: Benchmarking LLMs for Clause-Level Legal Risk Identification in Commercial Contracts," arXiv 2508.03080, August 2025. Uses the CUAD dataset; 4 proprietary and 15 open-source models.
80 unique completions from 1,000 identical prompts; identical after batch-invariant kernels; about 1.6x run time	Thinking Machines Lab, "Defeating Nondeterminism in LLM Inference," 10 September 2025
Best F1 about 63% on contract discrepancy detection; below 14% on legal citation matching	"CLAUSE: A Discrepancy Benchmark for Auditing LLMs," Findings of EACL 2026. More than 7,500 perturbed contracts, 10 error categories.
Reasoning models mentioned hints they used 25% (Claude 3.7 Sonnet) and 39% (DeepSeek R1) of the time	Anthropic, "Reasoning models don't always say what they think," 3 April 2025; Chen et al., arXiv 2505.05410
Unlikely AI: founded 2019 by William Tunstall-Pedoe; approach described in its own words	TechCrunch, "Alexa co-creator gives first glimpse of Unlikely AI's tech strategy," 9 July 2024
Unlikely AI described as combining machine learning with symbolic computing	Amadeus Capital, portfolio page
FIS acquired Droit, 9 March 2026; Adept encodes obligations as executable logic	A-Team Insight, "FIS Acquires Droit: Computational Law Moves Into the Core of Capital Markets Infrastructure," March 2026
Droit's generative "Decision Decoder"	Droit, "Introducing Decision Decoder: Explaining Adept Trade Decisions using Generative AI"
Nammu21 built into Loan IQ Nexus Build API, 5 May 2026	Finastra press release, "Finastra and Nammu21 Team Up to Cut Loan Onboarding Time," 5 May 2026
Nasdaq Ventures, State Street, and UBS invest in Nammu21	PRNewswire, 17 March 2025
Versana's shared hub for agent-sourced facility data	Versana news release, 30 April 2026
CUBE acquisitions: Thomson Reuters Regulatory Intelligence and Oden, Kodex AI, 4CRisk	CUBE news releases, January 2025 and February 2026; PRNewswire, 16 October 2025
ISDA CDM, run by FINOS, version 7 in production	FINOS, CDM overview; A-Team Insight, "FINOS Puts Common Domain Model (CDM) 7 into Production"
AIFMD II loan origination: 175% and 300% leverage caps, 5% retention, from 16 April 2026	Dechert, "AIFMD 2.0: Focus on Loan Origination," January 2026. Directive (EU) 2024/927.
ILPA reporting templates and model LPA	ILPA, Templates, Standards and Model Documents
AI Omnibus in force 27 July 2026; Annex III high-risk duties from 2 December 2027	European Commission, "AI Omnibus enters into force," July 2026