How AI Interprets Invoice Payment Terms: Where Frontier Models Fail
Frontier models get invoice due dates right. They still miscalculate discounts, give different answers between runs and split on ambiguous terms, so AP needs a purpose-built payment-terms agent.

Abstract
Payment terms decide when an accounts-payable (AP) function settles an invoice, which early-payment discount it may claim and what late-payment interest it may owe. On most invoices they are free text, such as "2/10 Net 30", "Prox 10" or "45 jours fin de mois", so turning them into dates and amounts is an inference problem that document extraction does not solve.
We describe agentic payment-term inference, in which a language model reads the term and plans to call deterministic date tools, and a runtime executes the plan under an explicit policy. We set out the building blocks that payment terms are made of, propose five design principles and walk through real-world examples.
We then evaluate seven frontier model configurations from three vendors, on a test set of 210 cases with exact ground truth. Every model tested with extensive rules in the prompt was correct on all dates. When vanilla prompts without business rules were used, six model configurations still scored 63 or 64 of 64 on the hardest subset of the data. The failures lie in the domain understanding: 14 of 386 accounting answers (3.6%, from 9 of 46 cases) were wrong with no flags, some errors changed between runs, the models split on 7 of 8 ambiguous terms, and one model did not answer in its consumer app or the API.A custom-built payment-terms interpretation agent of the date layer resolved 58 of 61 held-out expressions (95.1%) and returned the same date across three repeat runs; on all test sets, the frontier models matched or exceeded that accuracy. Frontier models read payment terms well. The case for an agentic module rests on what a control needs beyond that: the same answer every time, an explicit policy, hallucination suppression when a date is missing, and evidence for attribution. Based on our experiments and observations, we recommend that a general model not be used as a payment-terms interpretation agent on its own; a purpose-built, domain-specific model wins.
Keywords: payment terms; accounts payable; temporal reasoning; tool calling; agentic AI; large language models; financial controls.
Key findings
Dates are largely solved. Every frontier model that answered got every date case right. Without complex rules, six configurations scored 63 or 64 of 64 on the hardest sets.
Accounting knowledge is unsolved. 14 of 386 monetary answers (3.6%) were wrong. Three models misread a partial-payment discount and overpaid by 450.00, and none flagged the error for review.
Answers drift between runs. ChatGPT and Opus gave different answers to the same input on different runs.
Models disagree on policy. Where a term allows more than one reading, they split on 7 of 8, whether or not we stated the conventions.
Accuracy alone does not make a control. Against a custom-built payment-terms model, frontier models matched the date accuracy. However, the guardrails and governance are missing.
Introduction: reading a term is not computing it
A payment term is a compact contract. "2/10 Net 30" grants 2% off for payment within ten days and otherwise requires the full amount within thirty. "45 jours fin de mois" admits two computations, both accepted in French practice. EDI standards offer coded fields for such terms [13, 14], but e-invoices and most PDF and scanned invoices still carry them as free text [15]. Any accounts payable automation system therefore must work out the invoice due date, the discount deadline and any late-payment interest from text.
Turning a term into an obligation takes four steps, and each can fail. The term must be read, its reference dates resolved, the choices it leaves open settled by the buyer's policy, and the dates and amounts computed exactly. Earlier benchmarks found language models weak at temporal and numerical reasoning [1–4], but most predate current reasoning models, and none tested payment terms. Our study of invoice arithmetic found that frontier models computed accurately yet failed on reasoning and policy [16]. This report asks the same question for payment terms and makes four contributions:
A formalisation of payment-term inference as date primitives, accounting relations and buyer policy, with explicit abstention (Section 4).
Design principles for an agentic, tool-calling module, with worked examples (Sections 5 and 6).
An empirical study of seven frontier configurations on 210 cases with exact ground truth (Section 7).
Results from a payment terms interpretation agent of the date layer on real-invoice terms, and a cross-evaluation of the frontier models on its held-out sets (Section 8).
Why invoice due dates matter to accounting: a critical financial axis
Payment-term inference is a financial metric. It sits between capture and the payment run, the last point at which a wrong date or amount is cheap to correct.
Error | Consequence | Example from this study |
Due date too late | Statutory interest and recovery charges | 108 days late on EUR 37,800 costs EUR 1,175.24 at the EU Directive's minimum rate |
Due date too early | Cash leaves before it is owed | A 30-day error on a 90-day term pays a month early |
Discount deadline wrong | Discount lost or wrongly taken | Forgoing 1.5/15 net 45 costs 18.53% a year |
Amount wrong | Over- or under-payment | $450 overpaid on a partial-payment discount |
Statutory limit ignored | Terms longer than the law allows | Net 90 to an Indian micro or small supplier exceeds the 45-day limit |
The EU, UK, France, Spain and India all set payment periods or late-payment costs in law [8–12]. Errors also repeat: a misread convention is applied to every invoice that carries the term. A module must therefore be judged on exactness, consistency, and auditability.
Related work
Temporal and calendar reasoning. Test of Time [1] uses synthetic tasks so that models cannot recall answers, and finds them error-prone on complex temporal logic. Our cases are synthetic for the same reason, with ground truth computed in code. Halder et al. [2] report sharp drops on lunar and lunisolar calendars, and Bhatia et al. [3] show that tokenizers fragment dates. The reasoning models we tested handled long horizons and era calendars well, but two placed a Hijri date one day late.
Numeracy and stability. NumericBench [4] found that even two-number arithmetic falls short of 100%, and Sato [5] argues for zero-error horizons over average accuracy. Atil et al. [7] found accuracy varying by up to 15% across identical runs at temperature 0. FinanceBench [6] tested analysis over filings; we isolate the calculation and policy step.
Standards for payment terms. EDIFACT (PAT) [13] and X12 (ITD) [14] code discount and net periods, but EN 16931 carries terms mainly as free text with an optional due date [15], so much of what an AP pipeline receives must be inferred.
Tool-using agents. Program-aided and tool-using models delegate computation to code or functions [18–20]. Our module plans in one pass and lets a runtime execute, which places it closer to program-aided generation [18] than to an iterative reason-and-act loop [19]. We earlier described a reader-and-compiler approach to payment terms [17].
Problem formulation: the term as a program over dates and amounts
Given a term τ, a context c (invoice date, reference dates such as delivery or acceptance, jurisdiction calendar and amounts) and a buyer policy π, inference returns net obligations {(dᵢ, sᵢ)}, each a due date and a share; discounts {(rⱼ, δⱼ)}, each a rate and a deadline; and, where asked, amounts to the cent. Any date that depends on an input absent from c is null.
Family | Primitive | Example term |
Offsets | Add N days, weeks or months | Net 45; Net 3 months |
Anchors | Month, quarter or period end; day N of a later month | Net 30 EOM; Prox 10 |
Business days | Working days under a holiday calendar | 10 Arbeitstage |
Cycles | Close on a cutoff, pay on a later fixed day | 20日締め翌月末払い |
Combinators | Later of, earlier of, cap, instalments | whichever is later |
Conversions | Locale formats and era calendars | 03/04 (DE vs US); 令和9年 |
Three accounting relations carried most errors: a dynamic discount D = A × ρ × k for k days paid early; the settling payment after a partial payment C under discount r, C₂ = T(1 − r) − C, because each payment settles C / (1 − r) of gross; and compound interest with monthly rests, I = P[(1 + a/12)ⁿ − 1].
The policy π settles what the term leaves open, such as the reading of "fin de mois" or whether a statutory cap overrides the contract. An answer is correct only if every date, share and amount matches exactly. Correct outputs satisfy four invariants: shares sum to 100%, discount deadlines fall on or before net dates, dates move off weekends only when stated, and the same input yields the same output.
Agentic payment-term inference
In an agentic module the model reads the term and plans calls to deterministic tools, and a runtime executes the plan under the buyer's policy. We build it on five principles.

Figure 1 · agentic payment-term inference: read, plan, compute, check, record
P1. Separate reading from computing. The model decides what the term means; deterministic tools do the date and decimal arithmetic. An answer a model computes cannot be replayed, while a plan can be checked [18–20].
P2. Make policy explicit. Open choices, such as "fin de mois", net days after an end-of-month discount and statutory caps, are set once per buyer and applied identically to every invoice.
P3. Abstain when a reference date is missing. A tool call that names an unavailable date, such as delivery or a bill of lading, is rejected, and the output is null with a reason.
P4. Validate both ways. Each answer is checked against the invariants and, where available, an independent reading; disagreements go to review.
P5. Leave evidence. Each posted date carries its plan, the policy applied and the trace.
The module has three parts: a compact fine-tuned reader that splits the term into obligations (net dates with shares, discounts, penalties, payment type), a compact fine-tuned planner that emits a function-calling plan over a closed catalogue of date tools, and a runtime that executes the plan, rejects calls with missing inputs and selects candidates by share. The plan is generated in a single pass.
Examples
Example A: an anchor followed by an offset
Term (invoice date) | Plan | Result |
Net 30 EOM (2026-01-15) | month_end(invoice_date) → add_days(result_1, 30) | 2026-01-31 → 2026-03-02 |
The date alone is weak evidence, since 2026-03-02 is also "invoice date plus 46 days", a reading that fails on any other invoice date. The plan shows how the term was read and can be replayed on any invoice (P1, P5).
Example B: a discount alongside a net obligation
"2/10 Net 30" on an invoice of 10,000.00 dated 2026-11-02 yields two obligations: a 2% discount with a deadline of 2026-11-12 (pay 9,800.00), and a 100% net obligation due on 2026-12-02 (pay 10,000.00). The due date is the date of the 100% net obligation; the discount deadline is a separate option the payment run may take. If a reader leaked the discount deadline into the net dates, a rule that picked the earliest date would pay about 20 days early, which is why the reader is trained to keep the two apart (P1).
Example C: a missing reference date
For "30 days after B/L" with no bill-of-lading date, the planner emits add_days(bl_date, 30); the runtime rejects the call and returns no date with the reason "bill-of-lading date not provided" (P3). The invoice goes to review instead of being scheduled from the invoice date.
Example D: a choice that belongs to policy
"45 jours fin de mois" on an invoice dated 2026-01-15 gives 2026-03-17 (end of month, then 45 days) or 2026-03-31 (45 days, then end of month). Both are lawful [10]. A model that picks a reading invoice by invoice is never wrong on any one of them, yet the same supplier ends up paid on different dates. The policy fixes the reading once (P2).
Example E: a partial payment under a discount
Item | Value |
Invoice total T; term 1.5/15 net 45 | 60,000.00 |
Cash already paid in the window, C | 30,000.00 |
Gross settled by C, that is C / (1 − r) | 30,456.85 |
Second payment, T(1 − r) − C | 29,100.00 |
Three models answered 29,550.00, treating cash paid as gross settled, an overpayment of 450.00. With the relation written into the prompt, every model solved all eight such cases. The models could do the arithmetic but they completely misread the contract rule. Once the rule is identified, a deterministic tool gets it exact (P1, P4).
Empirical study: ChatGPT, Claude and Gemini on payment terms
7.1 Setup
Six test sets hold 210 samples with ground truth tagged and validated by experts. The set includes the following distribution.
Set | Cases | What it stresses |
H: edge dates | 32 | Prox rollover, EOM, locale formats, holiday rolls, caps, splits |
K: hard dates | 32 | National holiday calendars, Japanese closing cycles, era calendars, ISO week 53 |
V: batch | 100 | 100 invoices in one prompt across 12 term families |
R: invoice text | 12 | Terms in invoice text with money: Skonto, TDS, OCR noise, partial payments |
S: accounting rules | 10 | Payment runs, credit notes, EU/UK/MSMED late interest |
X: stress | 24 | Day counts in accounting formulas, partial-payment discounts, compound interest |
Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 ran as subagents of an agent harness with tools unused, to approximate a bare API call. ChatGPT (Think) and Gemini 3.1 Pro and 3.6 Flash (High) ran in consumer apps. All prompts forbade tools. Sets R and S ran twice. Scoring is an exact match.
7.2 Results
Model | H | K | V | R, run 1 / 2 | S, run 1 / 2 | X | Accounting misses |
Claude Opus 5.5 | 32 | 32 | 100 | 12 / 12 | 9 / 10 | 24 | 1 of 68 |
Claude Fable 5.1 | 32 | 32 | 100 | 12 / 12 | 10 / 10 | 24 | 0 of 68 |
Claude Sonnet 5.5 | 32 | 32 | 100 | 11 / 10 | 10 / 10 | 21 | 6 of 68 |
Claude Haiku 5.5 | 32 | 32 | 100 | 11 / 10 | 10 / 10 | 22 | 5 of 68 |
ChatGPT (Think) | 32 | 32 | 100 | 12 / 10 | 10 / 10 | 24 | 2 of 68 |
Gemini 3.6 Flash | 32¹ | 32 | not run | 12 | hung, then 10 | 24 | 0 of 46 |
Gemini 3.1 Pro | — | no answer | not run | no answer × 2 | not run | no answer | — |
7.3 Failure examples
Model | What the case says | What the model said | Why it matters |
Sonnet (both runs) | R12: 0.04% per day on 50,000.00, paid 46 days early; discount 920.00 | 900.00: "Paid 45 days early" | One day lost; 20.00 under-discounted |
ChatGPT (run 2) | R06: tax 870.92, total 11,739.92 | Tax 871.67, total 11,740.67 | Reconciliation breaks |
Sonnet | X21: 840,000 at 16.5% with monthly rests for 5 months; interest 59,360.11 | 59,535.15 | 0.3% too high: a wrong formula |
Sonnet (run 2) | R05 due date 2026-03-19 | 2027-03-19 | Paid a year late |
Gemini 3.1 Pro | Sets K, R and X, and X with tools allowed | Error message or no completion | No output to act on |
ChatGPT scored 12 of 12 on R and then 10 of 12 on identical input; Opus scored 9 of 10 on S and then 10 of 10. Some misses recurred in both runs (Haiku on R05, Sonnet on R12) while others appeared once.
7.4 Where models disagree
On 7 of 8 terms with more than one accepted reading, the models split.
Term (invoice date) | Answers | Split |
"45 jours fin de mois" (2026-01-15) | 03-17 vs 03-31 | Opus, Haiku, ChatGPT, Gemini vs Sonnet, Fable |
"2% 10 EOM, Net 30" (2026-01-12) | 02-11 vs 03-02 | Sonnet, Haiku, Fable, ChatGPT vs Opus, Gemini |
"60 gg DFFM", Italy (2026-01-20) | 03-31 vs 04-01 | Opus, Fable vs the rest |
"1 Shawwal 1448 AH", Saudi Arabia | 2027-03-09 vs 03-10 | Sonnet and ChatGPT chose 03-10; the Umm al-Qura table gives 03-09 |
Net 90, Indian micro supplier (2026-04-01) | 05-16 vs no date | Cap applied vs abstained (Haiku, ChatGPT) |
Net 90, Spanish B2B (2026-03-01) | 04-30 vs no date | Cap applied vs abstained (ChatGPT) |
Milestones 33/33/34 with "Net 30" | 06-25 vs 07-25 | Haiku alone added 30 days |
"Net 30 ROG", no receipt date | no date | All abstained |
Figure 2 shows the first case on a single invoice.

Figure 2 · one term, two readings: 45 jours fin de mois
Most splits are policy choices that belong to the buyer [10]. The Indian and Spanish caps run from acceptance or receipt [11, 12], which we did not supply, so the abstaining models followed the prompt's rule. The Hijri split is a knowledge error, consistent with [2].
7.5 Without stated conventions
To test whether the stated conventions were carrying the date results, we re-ran our two hardest date sets, H and K, with every convention removed: 64 cases covering year rollovers, holiday calendars and observed holidays, Japanese closing cycles, era calendars and ISO week 53. Six configurations scored 63 or 64 of 64.
Model | H, no conventions (of 32) | K, no conventions (of 32) |
Claude Opus 5.5 | 32 | 31 |
Claude Fable 5.1 | 32 | 32 |
Claude Sonnet 5.5 | 31 | 32 |
Claude Haiku 5.5 | 31 | 32 |
ChatGPT (Think) | 32 | 32 |
Gemini 4 Hydrogen (High) | 32 | 32 |
K is scored on dates and discounts, because the instructions for its accounting amounts were part of the removed conventions.
Of the three misses, one was arithmetic, one a reading and one a formatting gap. Opus placed "Net 45" from 21 May 2027 on 30 June, with the note "Day 45 is Wednesday 2027-06-30"; the correct day 45 is 5 July, an observed US holiday, so the term required 6 July. Sonnet gave no date for "CIA", and Haiku left the shares of a 30/60/90 split blank. The eight ambiguous terms split as they did with conventions stated: the models disagreed on 7 of 8, and the Hijri date received four different answers (8, 9 and 10 March 2027, and no date).
7.6 Failure analysis
Dates are largely solved. The main prompts named conventions such as Prox and EOM explicitly, but removing them (Sections 7.6 and 8.1) changed little: six configurations still scored 63 or 64 of 64 on sets H and K, and 98 to 110 of 110 end to end on the prototype's set.
Accounting errors come from reading the contract rule. The arithmetic was within the models' ability; the failing step was mapping a clause to the right computation.
Policy, abstention and availability. Models resolved ambiguous terms differently, handled missing dates inconsistently, and in one case did not answer at all. Each gap maps to a principle: plans and tools for arithmetic (P1), configuration for policy (P2), runtime rejection for missing dates (P3), checks on every answer (P4) and a recorded trace (P5).
Payment terms interpretation agent results: an agentic, tool-calling date layer
We implemented the date layer as a payment terms agent: a compact fine-tuned reader and planner that run on a single GPU, a closed tool catalogue and a runtime. Test sets are held out and drawn from real-invoice terms.
Held-out test | Result |
Reader, 110 cases (79 distinct terms): parses / shares correct | 100% / 99.3% |
Reader: discount kept out of due dates; invented discounts | 39 of 39; 0 |
Planner: isolated date expressions | 61 of 61 (100%) |
End to end, term–date pairs¹ | 96 of 106 (90.6%) |
Same date across 3 repeat runs¹ | 106 of 106 |
¹ Earlier planner version, fixed reader output, near-greedy production decoding. Figures come from successive builds of a research prototype, not the production configuration. They were scored before four answer-key labels were corrected (Section 8.1).
Building the module taught us three things: run training data through the agent's own tools, compute ground truth independently of the code under test, and measure stability at production settings. An earlier build was only 82.5% stable until we found and fixed a data defect.
8.1 Cross-evaluation on the prototype's test sets
We ran the seven frontier configurations on the prototype's held-out sets: 110 end-to-end cases and 112 isolated date expressions, 61 of which are the set reported above. Each set ran in two conditions: with the answer key's twelve conventions stated in the prompt (scored against the conventions as stated), and with only an instruction to read each term as it is normally read in AP practice and to return no date where information is missing. Tools were forbidden and each condition ran once. Gemini's app runs with Gemini 4 Hydrogen in place of 3.6 Flash, and Gemini 3.1 Pro also answered.
Model | End to end, conventions stated | End to end, not stated | Expressions, stated | Expressions, not stated (61-case set) |
Claude Opus 5.5 | 110 | 109 | 112 | 112 (61) |
Claude Fable 5.1 | 109 | 109 | 112 | 112 (61) |
Claude Sonnet 5.5 | 110 | 98 | 112 | 107 (57) |
Claude Haiku 5.5 | 110 | 108 | 112 | 109 (58) |
ChatGPT (Think) | 108 | 106 | 112 | 105 (58) |
Gemini 4 Hydrogen (High) | 109 | 109 | 112 | 112 (61) |
Gemini 3.1 Pro | 107 | 110 | not run | not run |
Hyperbots Agentic Payment Terms module | 110 | 109 | 112 | (61) |
Correct due dates out of 110 end-to-end cases; correct dates out of 112 expressions, with the 61-case set in brackets.
With conventions stated, every configuration that ran the expressions resolved all of them, and the few end-to-end misses were readings: a date for one half of "50/50" (Fable, Gemini Hydrogen), the standard reading of "Net 15 EOM" in place of the stated rule (ChatGPT), and "Prox 25 2%" treated as having no net date (Gemini Pro). The remaining misses were readings rather than arithmetic: no date for prepayment terms such as COD, the net period of "2% 10 EOM, Net 30" counted from the invoice date, no date for business-day terms without a holiday calendar, and a date for "50/50".
Property | Frontier models, unaided | Agentic Payment Terms Interpretation module |
Held-out date expressions, conventions not stated | 57 to 61 of 61 | 61 of 61 |
Arithmetic | Sampled; 3.6% of accounting answers wrong | Dates by tools; accounting layer always accurate |
Same input, same output | Changed between runs | 106 of 106 across 3 runs |
Ambiguous terms | One reading each; 7 of 8 split | None |
Missing reference dates | Mixed | Abstained for all |
Availability | One model did not answer in its app | Runs on the buyer's infrastructure |
Evaluation methodology
We propose five measures:
Measure | Why it matters |
Exact-match accuracy, accounting to the cent | Partial credit hides the accounting errors that matter |
Abstention precision and recall | A guessed date is worse than a flagged gap |
Run-to-run stability | A single clean run predicts little |
Policy consistency across invoices and runs | A split policy pays the same supplier on different dates |
Plan-level accuracy | Most due dates are reachable by "invoice date + k days", so date-only scores can be met by accident |
In our cross-evaluation, stating the conventions raised one model's score by 12 of 110 and lowered another's by 3.
Conclusion
Frontier models no longer struggle with calendar arithmetic, even when no conventions are stated. They fall short where an AP control is judged: about one accounting answer in 28 was wrong with nothing to mark it, some errors moved between runs, ambiguous terms were resolved differently by different models, and one model did not answer. Agentic payment-term inference is designed to close these gaps: the model reads and plans, deterministic tools compute under explicit policy, missing dates lead to abstention, and every answer is checked and recorded. Extraction tells you what a term says. Inference under policy tells you what to pay, and when, and shows its working.
Frequently asked questions
Q1. What is payment-term inference?
Payment-term inference turns the free-text payment term on an invoice, such as "2/10 Net 30" or "Prox 10", into the dates and amounts an accounts payable team acts on: the invoice due date, any early-payment discount deadline and, where relevant, late-payment interest, all under the buyer's own policy.
Q2. Can ChatGPT, Claude or Gemini calculate invoice due dates correctly?
Yes, to some extent. In our November 2026 benchmark, four Claude models and ChatGPT (Think) answered all 164 date cases correctly, including holiday-dependent working days and a 100-invoice batch. Without the explicit rules in the prompts, the best models still resolved real-invoice date terms almost perfectly: three resolved all 112 date expressions and one all 110 end-to-end cases. On our hardest date sets with no conventions, six configurations scored 63 or 64 of 64.
Q3. Where do large language models make mistakes on payment terms?
In accounting and contract rules. 14 of 386 accounting answers (3.6%) were wrong, such as a misread partial-payment discount that overpaid $450, and nothing in the output marked them as wrong. Answers also changed between identical runs. On terms with more than one accepted reading, the models disagreed on 7 of 8.
Q4. What is an agentic, tool-calling payment-terms module?
A design in which a language model reads the term and plans calls to deterministic date tools, and the tools compute the date. Hyperbots' payment terms interpretation agent resolved 61 of 61 date expressions it had not seen in training and gave the same date on three repeat runs for all 106 end-to-end cases. On the same expressions, frontier models scored 57 to 61 of 61, so the module’s advantage is consistency, explicit policy and an auditable plan.
Q5. Should accounts-payable teams rely on a general LLM for payment terms?
If money is not a convern, use it to read the term, and keep the control in deterministic code and a check on every result before a date or amount reaches the ledger.
References
Fatemi, B., Kazemi, S. M., Tsitsulin, A., Malkan, K., Yim, J., Palowitch, J., Seo, S., Halcrow, J., & Perozzi, B. (2025). Test of Time: A benchmark for evaluating LLMs on temporal reasoning. ICLR 2025.
Halder, D., Pandya, A., & Dabre, R. (2026). Not all time is Gregorian: Evaluating LLMs on cultural calendar systems. ICLR 2026 Workshop.
Bhatia, G., Peyrard, M., & Zhao, W. (2025). Date fragments: A hidden bottleneck of tokenization for temporal reasoning. arXiv:2505.16088.
Li, H., Chen, X., Xu, Z., et al. (2025). Exposing numeracy gaps: A benchmark to evaluate fundamental numerical abilities in large language models. arXiv:2502.11075.
Sato, R. (2026). Even GPT-5.2 can't count to five: The case for zero-error horizons in trustworthy LLMs. arXiv:2601.15714.
Islam, P., et al. (2023). FinanceBench: A new benchmark for financial question answering. arXiv:2311.11944.
Atil, B., et al. (2025). Non-determinism of "deterministic" LLM settings. arXiv:2408.04667.
Directive 2011/7/EU on combating late payment in commercial transactions (2011).
Late Payment of Commercial Debts (Interest) Act 1998 (UK).
Code de commerce, art. L441-10 (France).
Ley 3/2004, as amended by Ley 15/2010 (Spain).
Micro, Small and Medium Enterprises Development Act 2006, ss. 2(b), 15, 16 (India).
UN/EDIFACT D.99B, PAT segment (payment terms basis).
ASC X12 004010, ITD segment (terms of sale).
CEN EN 16931-1:2017, Electronic invoicing: semantic data model (BT-9, BT-20).
Hyperbots Inc. (2026). Agentic arithmetic validation for financial transaction documents. Technical report.
Hyperbots Inc. (2026). Payment-term inference and augmentation for accounts payable. Article.
Gao, L., et al. (2023). PAL: Program-aided language models. ICML 2023.
Yao, S., et al. (2023). ReAct: Synergizing reasoning and acting in language models. ICLR 2023.
Schick, T., et al. (2023). Toolformer: Language models can teach themselves to use tools. NeurIPS 36.
