HARRY STAHLSTAHL SOLUTIONS

Analysis paper · Shared on request

AI meets deterministic systems: rules, workflow, and compute

Harry Stahl ·

Where AI on its own falls short on rules, workflow, and calculation, how that shows up by business domain, and why the risk collects where they meet.

About this paper

This is the longer analysis behind the note AI meets deterministic systems: rules, workflow, and compute.

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

This paper contains no proprietary or confidential information from any previous or current employer or working relationship. I invest in Tracfox.

Some problems require both deterministic systems and AI

For all its flexibility and power, there are critical problem areas where AI fails to deliver reliable, explainable, repeatable results. Broadly, these are rules, workflow, and computation, often in combination. It's worth looking at how the strengths and failings of AI differ in each category, and how those problems show up in different business domains. That helps show which approaches can work for which problems and domains, and which vendors may be able to help.

Rules. Without encoded rules, a model misreads contract terms and gives explanations that don't match how it reached the answer. Here what AI lacks is curated structure: ontologies, and rule and contract content.

Workflow. Without a fixed procedure, a workflow that was right in month one drifts by month six. Here the gap is consistency with the approved procedure, which the customer organization needs to own.

Compute. Without a calculation engine, numbers go wrong and errors spread through a financial model. The missing pieces are the calculation engine and the representation that an agent works in. Here incumbents have most of the calculation engines.

And of course real decisions cross types. A covenant test needs the definition of EBITDA from the credit agreement, a quarterly procedure, and then the arithmetic. The risk collects at the handoffs, and few vendors cover more than one type.

Three kinds of deterministic results

Business and regulators need accurate, explainable, testable, repeatable results. They care more about whether the answer is right and traceable than what we call the process. But the challenges of rules, workflow, and computation expose different strengths and weaknesses in AI, and call for different solutions.

Rules Workflow Compute
What is fixed What the text or contract requires The sequence of steps Arithmetic over a financial model
What the AI does Turns text into structure, or explains a decision Reads inputs; handles exceptions; picks the procedure Builds or edits the financial model; reads the request; explains the result
Where AI fails Misread terms; explanations that don't match how the answer was reached Drift over long horizons; invented steps; a path an auditor can't follow Wrong numbers, wrong cells, errors spreading through dependencies
What is hard to build Ontologies, curated rule and contract content, people who know both Integrations, and an execution harness that leaves an audit trail A calculation engine, and a representation of the financial model an agent can work in cheaply
Who does it today Computational law and contract-data providers; content owners; standards bodies RPA vendors and AI-native entrants; the customer owns the procedure itself Planning vendors; in banking, ALM, pricing, and credit loss vendors

A fourth sense cuts across all three: repeatable model output, meaning the same answer to the same input. It's an engineering choice. 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). A smaller study of financial tasks found the same pattern by task type: outputs from SQL generation stayed stable, while retrieval-based answers drifted (Khatchadourian and Franco, Nov 2025). Small sample.

Repeatable output is worth having, and a model can be repeatably wrong. The three types address whether the answer is right.

Rules

Some decisions are set by written rules or contract terms, and language models are still weak at reasoning that has to follow them. The scarce part is structure: ontologies, curated rule and contract content, and people who know both the domain and the technology.

Across several familiar segments, some decisions share four 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.
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

We see similar patterns in insurance claims against policy wording, in tax treatment, and in trade documents checked against a letter of credit.

Challenges for LLMs

Language models are catching up on some traits and remain weak on others.

  • 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 (see the inference discussion above).
  • 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 own account of its reasoning is not a reliable record.

General approaches each cover some of the traits. A language model alone reads the documents and doesn't guarantee a checkable path. Classic machine learning needs labeled outcomes, which are rare here, and returns a score rather than a reason. A rules engine is exact and repeatable, and someone has to write every rule from the text and rewrite it when the text changes.

Where a combination gives advantage

LLMs are increasingly and widely available. What's scarce is the symbolic content: domain ontologies that are slow to build, curated content kept current as it changes, and people who know both the domain and the technology.

Because this is hard work, 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).

Another way to narrow the advantage is shared standards: 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 belongs to any one vendor.

In syndicated lending, Nammu21 positions its NEL Protocol as "the data protocol for credit markets." Versioning is native, so "when a facility amends, every projection updates," and its F-Codes are "a common reference for loan terms, as CUSIPs are for securities" (Nammu21). It's a vendor protocol rather than one run by an industry body, which makes it a test of whether a vendor product can become a shared standard.

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; Nammu21's NEL Protocol and F-Codes (vendor-run) 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 (Versana). Nammu21 creates persistent information objects from the agreement itself, which is arguably the more durable and extensible approach.

In private credit, standards are still forming. 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, Jan 2026). ILPA publishes reporting templates and a model LPA (ILPA). How closely LPAs follow the model is not measured.

Who does it today

Pairing learning (reading, recognizing patterns) with symbolic reasoning (explicit knowledge and logic that can be checked) is what researchers call neuro-symbolic AI. Production systems in finance pair strong components, and differ in how much of the work the AI does.

Vendor What it does Role of AI Evidence
Droit (FIS since 9 Mar 2026) Encodes regulatory obligations as executable logic; decides in real time whether a trade or product complies with each jurisdiction's rules Light: a generative Decision Decoder explains decisions after the fact A-Team Insight; Droit
Nammu21 Turns credit agreement provisions into structured, computable data (NEL Protocol); in Finastra's Loan IQ Nexus Build API since 5 May 2026 Not published Finastra; Nasdaq Ventures, State Street, and UBS invested in March 2025 (PRNewswire)
Tracfox AI agents that take regulatory change through to audit evidence: horizon scanning, a regulatory repository, obligations and controls management, and evidence validation, with "every output traceable to its source" Agents work within defined tasks; "compliance professionals remain responsible for review, judgement and approval" Tracfox. Vendor
CUBE Regulatory content and change tracking, now with agents (Kodex AI, 4CRisk) Agents on curated content CUBE releases (see above)
Unlikely AI Building toward one coherent system that combines learning with symbolic reasoning The target state TechCrunch, Jul 2024; Amadeus Capital

Droit is the clearest production reference for deterministic rules, with AI on the explanation.

Distribution is part of the picture. Tracfox reaches banks as a third-party agent on Fiserv's agentOS, which launched on 14 May 2026 with 4 Fiserv agents and 9 from third parties (Tracfox; Fiserv, May 2026).

Open questions and missing data

  • As models improve, does the AI's role move from explaining decisions toward encoding the rules?
  • Should contract-data providers such as Nammu21 be read as rules, or as data? The advantage differs: a curated ontology, or a persistent data standard.
  • No benchmark tests language models on covenant tests or side-letter obligations.

Workflow

Financial workflows are business critical, so the sequence and numbers need to be correct every time. But there's evidence that using agents can lead to problems. The first is that errors compound across sequential steps (SpreadsheetBench 2, Jun 2026; AccountingBench, Jul 2025). The second is that a workflow that's correct on the first run drifts over time (AccountingBench, Jul 2025). When that workflow is closing the books, drift is unacceptable. Finance leaders sense this. Deloitte Finance Trends found only 14% of finance leaders support full autonomy in critical decisions (Deloitte, Sep 2026). In response, a vendor like Kognitos says its agents compile each procedure once, run it the same way every time, and add no steps outside the approved procedure (Kognitos, Oct 2026).

Challenges

Compounding over long horizons. In a test of models closing the books for a real software business, models were within 1% of a CPA in month one, then drifted. None completed the year, and some invented transactions to make the books balance. Vendor-run, older models. Scores on multi-step finance tasks are rising: the top score on Vals' Finance Agent benchmark went from about 52% in May 2026 to 67.8% by October (Vals Index). Private test set. Better task scores don't yet show that long-horizon drift is solved.

Repeatability. An auditor expects the same case to take the same steps. An agent that decides each step as it goes can take different paths through identical cases.

Exceptions. A fixed procedure handles the expected case. What happens to the unexpected one, and who sees it, is where workflow products differ.

Who does it today

Approach Vendors What they offer
Compiled procedures Kognitos Each procedure compiled once and run the same way every time, with no steps outside the approved procedure; it calls the approach neurosymbolic (Kognitos, Oct 2026)
RPA incumbents adding agents UiPath; Automation Anywhere UiPath: deterministic automation as the backbone of regulated, high-volume functions, with agents for exceptions; Maestro coordinates month-end close at one customer (Pulse, citing Diginomica, May 2026). Automation Anywhere: 55+ finance agents with governance controls and human review of exceptions (MarTech Series). Secondary sources
Agents that decide each step as they go AI-native accounting agents; general agents Flexible, and harder to repeat

The customer owns the approved procedure, so a vendor needs to provide integrations, an execution harness, and audit evidence. RPA incumbents already sell to the same buyer and are adding agents.

Open questions and missing data

  • Does a compiled procedure handle exceptions better, or route them to a person? No exception-handling outcomes are published.
  • Is "neurosymbolic" a fair label for workflow determinism, or does it blur the rules type?

Compute

Language models are getting better at arithmetic through tools, so failure has moved to building and navigating the financial model: the workbook or planning model that holds the numbers.

A calculation engine is software that holds a financial model as structured data (dimensions, line items, and formulas) and recalculates it itself. The agent asks the engine for a number rather than working it out. Anaplan's Hyperblock is one; the rate-risk engine inside an ALM system is another.

What works better, and what less so

Results improve as more of the math moves out of the language model.

  • Model alone. On ORCA, a benchmark of 500 everyday calculation prompts with deterministic answers, the best of five frontier models answered 63% correctly, and 76.7% of the 43 finance and economics prompts. Binary accuracy, models tested without tools (ORCA, Nov 2025).
  • Model writing code for an interpreter. Having the model write a program and letting an interpreter run it raised accuracy on GSM8K from 65.6% to 72.0%, and on large-number problems from 23.1% to 61.2% (PAL, 2022). The models are dated; the direction has held.
  • Model inside a spreadsheet. The arithmetic goes right more often than the financial model around it. In one finance modeling benchmark, a top model passed 87% of checks on formula structure and 61% on numbers (Vals, via Let's Data Science). Private test set. On 321 end-to-end workbook tasks, the best model passed 34.9%, and 12% of debugging tasks. Formulas were often textbook-correct and pointed at the wrong cells (SpreadsheetBench 2, Jun 2026).
  • Model on a dimensional planning engine. Claimed rather than measured. Anaplan says pairing language models with its "deterministic platform" delivers "trusted, auditable answers" and avoids the token cost of having a model calculate (Anaplan, Jun 2026). No public benchmarks.

Challenges

Building and navigating the financial model. The failure is in choosing the right cells, line items, and dependencies across a large model. Dimensional engines, and the entrants that use them, argue (persuasively) that a model small enough to fit in an agent's context, expressed as formulas over named dimensions, removes much of that problem.

Trust. Buyers don't (yet) trust autonomous AI. In Deloitte's Finance Trends 2027 survey, among the 42% whose agentic ambitions exceed current capability, 41% name employee trust in the quality of AI outputs as a leading barrier (Deloitte, Sep 2026). Vendors are also advising caution. Microsoft warned against using its Excel COPILOT function for numerical calculations (Aug 2025) and withdrew it from 14 September 2026 (Microsoft; Slashdot, Aug 2025).

No measured comparison across engines. Entrants benchmark against spreadsheets, which is easy and dramatic. Nobody has published a like-for-like test of an agent working on one dimensional engine against another.

A low floor and a crowded top. General models such as ChatGPT, Claude, and Gemini can write and run code for arithmetic at no extra cost to the user. That's the baseline any engine has to beat on accuracy, cost, or auditability. Above it, the established engines belong to incumbents who can add agents themselves.

Who does it today

Function Incumbent engines Agent activity
Planning and FP&A Anaplan (Hyperblock), Pigment, IBM Planning Analytics (TM1), Oracle EPM (Essbase), Board Anaplan announced agents for the office of the CFO in June 2026, due by October; shipment unconfirmed. Pigment launched a Modeler Agent in March 2026
Bank ALM FIS, SAS, Moody's, QRM, SS&C, Empyrean, Mirai, Abrigo SAS Copilot for ALM; Mirai's AI modeling and agent module; no outcomes published
Loan and deposit pricing; credit loss Established products with established owners Not reviewed

Sources: Anaplan, Jun 2026; Pigment, Mar 2026; Chartis 2025 coverage and vendor pages. Vendor-reported.

AI-native entrants build calculation engines for agents. cfo.ai (formerly Runway) launched its AI CFO, Ari, in September 2026 on Reambase, which it describes as a "multidimensional calculation engine built specifically for AI agents." Values get business addresses built from named variables and dimensions rather than cell coordinates, so "meaning travels with the value" while the engine handles formulas, dependencies, and recalculation (cfo.ai via PRNewswire, Sep 2026; SiliconANGLE, Sep 2026). The launch material publishes no benchmarks. Entrants like this compete on how cheaply an agent can work in the representation, on the data around it, and on distribution. How durable any of those is remains open.

Open questions and missing data

  • Is an agent measurably more accurate, or cheaper, on one dimensional engine than another? No like-for-like test is public.
  • Did Anaplan's agents for the office of the CFO ship in October? No release confirms it.
  • Should calculation be treated as two problems, arithmetic and financial model construction, with different answers for each?

Business domains

The type that matters most differs by domain. In most of financial services the hard one is rules, and calculation is the hard one in fewer places than the attention it gets suggests.

Domain Rules Workflow Compute
Compliance Main Some
Syndicated lending Main Some Some
Private credit and funds Main Some Some
Capital markets: trading Main Some
Insurance Main Main
Office of the CFO Some Main Some

Compliance

The decisions are which rules apply, which control meets each one, and what evidence proves it. The scarce asset is the mapped chain from rule to obligation to control to evidence, kept current as rules change. A buyer can ask a vendor for that lineage rather than for answers, because an examiner will walk it backward.

The cost falls hardest on small institutions. Community banks in the smallest quartile spent 11% to 15.5% of payroll on compliance, against 6% to 10% for the largest (CSBS, Nov 2025). Reference points: Droit, CUBE, and Tracfox (see Rules). Distribution runs partly through core platforms such as Fiserv's agentOS.

Syndicated lending

The decisions are covenant tests, amendments, consents, and waivers, all set by the credit agreement. Whoever structures the agreement data is in a position to own the workflow that runs on it.

Amendments are the main risk. The definition changes, and every test and calculation that depends on it has to follow. Nammu21 says of its NEL Protocol that "when a facility amends, every projection updates" (Nammu21). Versana takes the other route, publishing structured facility data from agent banks (Versana).

Private credit and funds

The decisions are which side-letter terms a fund owes each investor, which loans count toward a borrowing base, and what goes into investor reporting. Each document is negotiated one deal at a time, and little shared structure exists. Firms build it deal by deal or go without.

Regulation is adding hard limits. 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 (Dechert, Jan 2026). ILPA's reporting templates and model LPA are partial standards; how closely LPAs follow the model is not measured.

Capital markets: trading

Firms must check compliance on every trade to see whether it complies with each jurisdiction's rules, and how it's processed after execution. Droit's Adept encodes those obligations as executable logic (see Rules).

Shared standards are furthest along here. ISDA's Common Domain Model, run by FINOS and in production at version 7, gives the industry a machine-readable model of trades and legal agreements (FINOS). Where the structure is shared, the advantage shifts from owning the structure to running it well.

Insurance

Here it's critical to make sound decisions about claims against policy wording, underwriting against guidelines, and administration of policies written decades ago. Rules and workflow both matter, and both run through several generations of systems.

75% of life and annuity insurers run 4 or more admin platforms for closed blocks, and 58% have yet to cut their mainframe footprint (EY Closed Blocks Survey, 2023). In one 2026 survey, 52% of policy administration processes still needed manual intervention (RSM for INTX, Mar 2026). Vendor-sponsored. The rules often live in code, sometimes in more than one generation of it, so mapping them out of the code comes before automating them.

Office of the CFO

Some key decisions for the CFO are the close, consolidation, forecasts, and board reporting. The close is a procedure problem more than a math problem: the same steps, in the same order, every month, with evidence an auditor accepts (see Workflow).

Calculation matters most in planning, where the engines are already held by incumbents adding agents (see Compute). Buyers are cautious: 14% of finance leaders support full autonomy in critical decisions (Deloitte, Sep 2026).

Where the types meet

Handoffs between types add risk, and few vendors cover more than one.

Decision Rules Workflow Compute
Regulatory change The new obligation, from the rule text Updating and retesting the control Little
Insurance claim Coverage, from the policy wording The claims procedure The payout
Covenant test EBITDA definition from the credit agreement Quarterly test cycle, waivers The ratio
Month-end close Accounting policy The close procedure Reconciliation
Borrowing base Eligibility terms in the facility agreement Reporting cycle The base
Cash forecast Little Data collection The forecast model

Challenges at the handoffs

Definitions drift between rules and compute. The definition of EBITDA lives in the credit agreement. The calculation lives in a financial model. When an amendment changes the definition, nothing forces the model to change with it.

A procedure that calls an engine can fail in either category. The steps can run in the right order but still produce a wrong number, or produce the right number but skip a step.

Explanations stop at each tool's boundary. A rules system can trace its answer to a clause. An engine can trace a number to its inputs. An examiner or auditor asks for the whole path, and each tool's lineage ends where the next tool starts.

Each type has a different owner. Legal or credit owns the terms, operations owns the procedure, and finance owns the model. Vendors tend to sell to one of them.

What helps

  • Structure the terms once, and let the calculation read them. Where agreement terms exist as structured data, a covenant test can read the current definition rather than a copy of it. Nammu21's computable agreement data in Loan IQ points this way.
  • Keep arithmetic out of the model and the procedure. The agent and the workflow call an engine for numbers and log the inputs and outputs of each call.
  • Carry lineage end to end. Each result traces from the clause, through the procedure, to the calculation, in one record an outsider can follow.
  • Use shared standards where they exist. ISDA CDM, LSTA and LMA forms, and ILPA templates reduce how much structure each firm builds alone.
  • Test the whole chain. Run the same case twice through every type, and check that both the steps and the numbers repeat.

Questions when evaluating approaches and solutions

  1. Rules: Can each answer be traced to the clause or rule behind it? Who maintains the content, and how fast after a change? What happens when the text is ambiguous?
  2. Workflow: Given the same case twice, does it take the same steps? What happens on an exception? What does an auditor see?
  3. Compute: Does the agent do any arithmetic itself? How does it find the right cells or line items in a large financial model? What is the evidence against an incumbent engine, beyond a spreadsheet?
  4. Inference: Is model output repeatable for identical input, and at what cost?
  5. Across types: Which types does the decision need? Where does lineage break between them? What does the buyer still build or maintain?

Areas to probe or discuss

  • How much is calculation really solved? With tools, arithmetic works (PAL), but building and navigating the financial model still fails (SpreadsheetBench 2, Vals).
  • How fast are agents getting better at workflow? The Finance Agent top score rose from about 52% to 67.8% in five months, but long-horizon work still compounds errors (AccountingBench).
  • Where to find explanation of results? A structured decision path is an explanation, but where decisions are too complex to understand, use a model afterward, like Droit's Decision Decoder.
  • Who has the advantage in compute? Incumbents have calculation engines and are adding agents, but smart entrants are making engines cheaper for agents to use (for example, Reambase). Where a decision is mostly calculation, is the engine an advantage for anyone, or only the data and distribution around it?
  • Who's building the rules? Shared standards and vendors are each adding structure, and sometimes collaborating: CDM in derivatives, Versana and Nammu21 in syndicated lending, and ILPA in private funds. In each field, does the structure end up as a vendor product or a shared standard?
  • Where is regulation creating time pressure? The EU AI Act treats credit scoring as high-risk, with explanation and documentation duties, but the AI Omnibus, in force since 27 July 2026, moved those duties to 2 December 2027 (European Commission).
  • One system or paired components? When is it sufficient to put systems side by side (some workflows?), and when is there real advantage from a fully coherent system (perhaps financial modeling and planning?)
  • Which decisions in your work need which type, and which need more than one?
  • As models improve, does each deterministic layer shrink to a thin audit trail, or does the need for outside checking keep it thick?
  • Does any vendor credibly span two types today?

I work on these questions with lenders, investors, and fintech and regtech companies. The advisory page describes how.

Appendix

Sources

Claim Source
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
Output consistency by task type in financial workflows Khatchadourian and Franco, "LLM Output Drift: Cross-Provider Validation and Mitigation for Financial Workflows," arXiv 2511.07585, November 2025. 480 runs, 16 per condition. Preprint
Best of five frontier models 63% accurate on 500 calculation prompts; 76.7% on 43 finance and economics prompts Herambourg et al., "The ORCA Benchmark: Evaluating Real-World Calculation Accuracy in Large Language Models," arXiv 2511.02589, November 2025. Models tested October 2025, without tools; binary accuracy. Authors include the Omni Calculator research team
Program-aided models: 72.0% vs 65.6% on GSM8K; 61.2% vs 23.1% on large-number problems Gao et al., "PAL: Program-aided Language Models," arXiv 2211.10435, Carnegie Mellon University, 2022. Solve rate (accuracy) on GSM8K and GSM-Hard
87% of formula-structure checks and 61% of number checks Vals AI Excel Modeling Benchmark, via Let's Data Science, 2026. Private test set; secondary coverage
Best model 34.9% on 321 workbook tasks; 12% on debugging; correct formulas with wrong references "SpreadsheetBench 2," arXiv 2606.29955, June 2026. Pass rate on end-to-end tasks. Listed as a NeurIPS 2026 poster
Anaplan pairs language models with a deterministic platform; agents for the office of the CFO due by October 2026 Anaplan, "Anaplan Introduces the Agentic Enterprise," 30 June 2026. Vendor announcement
cfo.ai launch of Ari on Reambase; Reambase described as a multidimensional calculation engine for AI agents cfo.ai press release via PRNewswire, 10 September 2026; SiliconANGLE, 10 September 2026. Company claims; no benchmarks published
Pigment Modeler Agent Pigment press release via PRNewswire, 4 March 2026. Vendor announcement
14% support full autonomy in critical decisions; 42% say ambitions exceed capability; 41% of those name trust in AI output quality Deloitte, "Finance Trends 2027," press release, 9 September 2026. 1,434 finance leaders in 26 countries, surveyed spring 2026
COPILOT function warning and withdrawal Microsoft Support, COPILOT function page, viewed November 2026; Slashdot, 20 August 2025
Bank ALM vendors and AI features Chartis Research, "ALM Solutions, 2025: Quadrant Update," 12 December 2025 (category names); SAS Chartis vendor analysis, 2025, hosted by SAS; Mirai release, 22 September 2026. Vendor-reported
Models within 1% of a CPA in month one, then drifting; none completed the year Penrose, AccountingBench, July 2025, via the author's thread. Vendor-run; not peer reviewed
Finance Agent top score about 52% (May 2026) and 67.8% (October 2026) Vals AI, Vals Index, update of 19 October 2026; May figure via Pebblous, 2026. Private test set
Kognitos compiled procedures Kognitos news release, 21 October 2026. Vendor claim
UiPath CEO on deterministic automation; Maestro for month-end close Pulse, citing Diginomica, May 2026. Secondary
Automation Anywhere finance agents MarTech Series, undated. Trade press on a vendor launch
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. CUAD dataset; 4 proprietary and 15 open-source models
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
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
NEL Protocol as the data protocol for credit markets; native versioning; F-Codes Nammu21, Solutions page, viewed November 2026. Vendor
Tracfox capabilities, traceability, and human review; third-party agent on Fiserv agentOS Tracfox website, tracfox.com, viewed November 2026. Vendor
Fiserv agentOS launch: 4 Fiserv agents and 9 from third parties Fiserv press release, "Fiserv Launches agentOS," 14 May 2026
Unlikely AI: founded 2019 by William Tunstall-Pedoe; combines machine learning with symbolic computing TechCrunch, 9 July 2024; Amadeus Capital, portfolio page
CUBE acquisitions 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"
Versana's shared hub for agent-sourced facility data Versana news release, 30 April 2026
Community bank compliance share of payroll by asset-size quartile CSBS, Siems and Ross, "Too Small to Scale: What 10 Years of Data Say About Community Bank Compliance Costs," 13 November 2025. Annual surveys 2015 to 2024
Life and annuity closed blocks: 4+ admin platforms; mainframe footprint EY US Life and Annuity Closed Blocks Survey, 2023. Respondent count not stated
52% of policy administration processes need manual intervention RSM for INTX, via Business Wire, 11 March 2026. 250+ insurance respondents. Vendor-sponsored
AIFMD II loan origination rules 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

Not sourced

Judgment and framing:

  • The three types of deterministic, and repeatable output as a fourth, cross-cutting sense.
  • What fails without each type, what is hard to build, and who does it today.
  • That failure in compute has moved from arithmetic to building and navigating the financial model.
  • The ordering of approaches from model alone to dimensional engine.
  • That a small dimensional representation removes much of the navigation problem, and that the argument is persuasive.
  • That general models' code execution sets the floor for engine vendors.
  • What entrants in compute compete on.
  • That errors compound across sequential steps and workflows drift over time, beyond the AccountingBench result.
  • That the customer owns the approved procedure, and what a workflow vendor needs to provide.
  • The four traits of the rules pattern, the examples, and who asks how.
  • The assessment of where each general approach falls short.
  • That Droit's AI element is light, and Droit as the clearest reference for deterministic rules.
  • That the symbolic side is the advantage, what it takes to build, and that shared standards narrow it.
  • The three-case table, and that Nammu21's approach is more durable and extensible than a data hub.
  • The domains-by-type grid, and which type is hardest in each domain.
  • That rules, more than calculation, is the hard type in most of financial services.
  • The domain so-whats: lineage as the compliance buyer's ask; advantage shifting from owning to running structure in capital markets; agreement data and the workflow on it in lending; deal-by-deal structure in private credit; mapping rules out of code before automating in insurance; the close as a procedure problem.
  • That risk collects at the handoffs; the handoff challenges; and what helps.
  • The decisions-by-type table.
  • That Nammu21's NEL Protocol is a test of whether a vendor product can become a shared standard.
  • The evaluation questions, and the areas to probe or discuss, except where a source is cited.

Exists, but not used or not found:

  • A like-for-like comparison of agent accuracy, token cost, and tool calls across dimensional engines.
  • Confirmation that Anaplan's agents for the office of the CFO shipped in October 2026.
  • Benchmarks of language models with and without a calculation engine on finance tasks such as forecasts or interest rate risk.
  • Benchmarks of language models on covenant tests or side-letter obligations.
  • Exception-handling outcomes for compiled-procedure or RPA agent products.
  • Finastra's claim that 70% of syndicated loans are serviced on Loan IQ: a vendor marketing figure with no published method.
  • CUBE's customer count, for the same reason.
  • Unlikely AI's accuracy claims, which are the company's own and not independently tested.
  • How closely fund LPAs follow the ILPA model.
  • Measures of AI accuracy by business domain on these decisions.
  • Vals' methodology page, which was not read.
  • Agent activity among loan pricing and credit loss vendors.
  • Published benchmarks for Reambase or other engines built for agents.

Questions or a different read? Email me.