Contents
  1. Financial work and its development
    1. Four kinds of financial work
    2. Scoring, rules, and financial language
  2. Financial records and their use
    1. Preserve what a financial fact means
    2. Separate reporting time from known time
    3. Check permission for the actual use
  3. Applying AI to financial workflows
    1. Build checked financial research
      1. Calculation and interpretation
    2. Reconcile records and resolve exceptions
    3. Resolve the customer's financial problem
    4. Keep scores separate from decisions
    5. Identify the exposure and scenario
  4. Authority and authoritative effects
    1. Authorize the actual financial action
    2. Connect terms, entitlement, and receipt
    3. Resolve uncertain execution
  5. Evidence of useful performance
    1. Measure the financial workflow
    2. Respect the historical information boundary
    3. Keep missing and delayed outcomes visible
  6. Institutional responsibility and intervention
    1. Assign ownership and independent challenge
    2. Make review capable of changing the outcome
  7. Deployment and continued control
    1. Choose a bounded operating mode
    2. Reconstruct, monitor, and reassess
  8. Check understanding
  9. Open questions
  10. Selected talks
  11. References
  12. Talk library
← All topics

AI in finance

Learn how teams apply AI to financial research, document extraction, reconciliation, and customer assistance. This guide connects lessons from AI Engineer talks with the controls needed to evaluate and operate these systems.

From previous AI Engineer events

For an in-person event on this theme, explore AI Engineer New York’s financial-services conference (October 12–14, 2026).

Financial work and its development

Four kinds of financial work

Financial institutions and finance teams work with rights and obligations. A financial instrument is a contract that creates a financial asset for one party and a financial liability or equity instrument for another. A loan, for example, connects a lender's right to receive payments with a borrower's obligation to make them. This contractual meaning matters: describing an obligation, recording it, and fulfilling it are different activities. AASB 132 supplies the formal definition.

Financial operations maintain financial records and carry transactions through their required stages. Decision support prepares information for a consequential choice. Neither automatically grants decision authority: permission and responsibility to approve or execute that choice. Research, operations, customer service, and decision support can share document-reading capabilities while requiring different completion checks.

Treat these as workflow families, not mutually exclusive departments.
WorkInputs and useful outputResponsibility and possible effect
ResearchDisclosures and records → checked analysisAn analyst reviews the conclusion; a separate decision may commit capital.
OperationsStatements and internal records → resolved discrepanciesOperations staff control corrections and completion; errors can misstate books or misdirect payments.
Customer serviceCustomer request, policy, and account facts → resolution or escalationService staff act within account permissions; answers can affect access to help and financial choices.
Decision supportApplicant or transaction information → assessment and proposed dispositionAuthorized decision-makers determine credit or payment treatment; affected parties include borrowers and legitimate customers.

This breadth is visible in the Bank of England and FCA's 2024 survey of 118 firms. Respondents described operational, fraud, risk, and other applications; the survey separately classified automation and materiality. Adoption therefore did not mean autonomous authority. The reported benefits were firms' perceptions, not controlled measurements of productivity.

Scoring, rules, and financial language

Financial AI draws on several continuing traditions. Credit scoring estimates credit-related outcomes from applicant or account information. An expert system applies explicitly represented domain rules. Financial language processing interprets text whose meaning depends on accounting and business context. Their histories explain why a language assistant can extend an existing workflow without replacing its numerical models or controls.

Contributions that still coexist

  1. 1941Credit-Rating FormulaeCombines borrower attributes statistically to support credit investigation.Sources & context

    Contributors: David Durand; 1941 NBER chapter.

    What changed: Examined statistically derived weights as an alternative to intuitive weighting. Durand treated the formulae as supplements to judgment and warned that previously approved loans did not represent the full applicant population.

  2. 1988 deploymentAuthorizer’s AssistantCombines account records with credit and fraud policies for authorization.Sources & context

    Contributors: American Express and Inference Corporation; reported in 1989 by Dzierzanowski, Chrisman, MacKinnon, and Klahr.

    What changed: Handled selected U.S. Personal and Gold card transactions automatically and advised human authorizers on others. Integration, acceptance testing, and transfer of business control to operations were part of deployment.

  3. 1992FalconApplies neural-network modeling to payment-fraud detection.Sources & context

    Contributors: FICO; date attributed to its retrospective corporate history.

    What changed: Addresses suspicious payment patterns, distinct from estimating repayment risk or processing credit applications. FICO’s retrospective dates Falcon to 1992.

  4. February 2011Loughran–McDonaldMeasures financial text with domain-appropriate word categories.Sources & context

    Contributors: Tim Loughran and Bill McDonald; When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks.

    What changed: General negative-word lists mischaracterized financial filings. The contribution was better measurement of financial meaning, not simply a larger vocabulary.

  5. March 2023BloombergGPTUses mixed financial and general text for broad financial-language tasks.Sources & context

    Contributors: Shijie Wu, Ozan Irsoy, and colleagues; March 2023 introduction.

    What changed: The 50-billion-parameter model supports tasks including entity recognition, entity linking, and question answering. Its benchmark findings concern language capabilities, not autonomous investment decisions or completed transactions.

Statistical scoring, explicit policies, fraud models, and financial-language processing support different responsibilities rather than successively replacing one another. Spacing is not to scale.

These developments did not eliminate one another. A financial application can use language processing to interpret documents, a predictive model to prioritize cases, explicit rules to enforce policy, and people to investigate exceptions. The useful architectural question is which responsibility each component can support.

Financial records and their use

Preserve what a financial fact means

An issuer creates an instrument; an account records a financial relationship and activity; a holding, or position, gives the instrument quantity held there. A transaction records activity; a balance gives an amount at a time. Keep these identities distinct when joining records: an issuer does not identify an instrument, holding, or payment.

Different records answer different questions. Reference data supplies identifying and contractual attributes; a security master is an institution's maintained collection of such instrument records. Transaction data records activity. Financial statements summarize position or performance. A balance sheet reports assets, liabilities, and equity at an instant: resources, obligations, and the owners' residual interest. An income statement reports revenue and expenses over a period; a cash-flow statement reports cash movements. Profit therefore differs from cash received. Before comparing extracted amounts, preserve the entity, measure, period, currency, scale, and accounting context. The U.S. Securities and Exchange Commission's financial-statements guide explains these distinctions.

Standards preserve different parts of this meaning. A Legal Entity Identifier (LEI) identifies a legal entity; an International Securities Identification Number (ISIN) identifies an instrument. Their issuer mapping does not identify a customer's holding or payment, and missing coverage is not proof of nonexistence. eXtensible Business Reporting Language (XBRL) represents reporting facts with context rather than bare values.

The xBRL-JSON tutorial provides an illustrative assets fact, not an actual company's report.
FieldMeaning to preserve
Entity and conceptWhose assets the fact describes, and which reporting concept it uses.
Reporting instant2020-01-01 at midnight represents the end of December 31, 2019.
Value and unitEUR 1,230,000; numerical strings avoid representation-related precision loss.
Precision and qualificationsDeclared precision, applicable dimensions, and footnotes qualify interpretation; they do not certify truth.

Extraction must retain these relationships even when they appear in a table heading or footnote rather than beside the number. See Recover table meaning. A standardized record is easier to process, but still needs semantic checking against its source.

Separate reporting time from known time

Provenance records information's origin and history. Distinguish reporting period, publication, application receipt, and correction. As-of information is eligible at a cutoff. A restatement revises earlier reported amounts; the IAS 8 overview describes retrospective correction of material prior-period errors, subject to its conditions.

FRED's real-time periods retrieve historical information; its current view includes revisions. Date-level vintages establish neither intraday publication nor ingestion. Application replay must respect later receipt.

Aggregation can obscure this boundary. EDGAR frames select the last-filed fact fitting a calendar period; that historical period is not a historical availability guarantee. Its aggregate XBRL APIs also exclude some custom and segment disclosures. Preserve the selected source version rather than assuming an API's convenient view is the required information set.

In a hypothetical example, both versions describe the same company’s quarterly revenue in USD millions. The application receives the original value of 100 before its decision cutoff. A correction to 96 is published before the cutoff but reaches the application afterward. Historical replay retains 100; the later corrected view uses 96. See Origins, times and permitted use for provenance and Applicable sources and unresolved conflicts for alignment before synthesis.

Published does not mean received

Both versions describe Company C’s quarter Q revenue, in USD millions. Event order runs left to right; spacing does not measure elapsed time.

Publication and receipt define different boundariesOriginal 100 is published and received before cutoff. Revision 96 is published before cutoff but received after it. Historical application replay uses 100; a later corrected view uses 96.Earlier eventsLater eventsDecision cutoffOriginal: 100PublishedReceivedRevision: 96PublishedReceivedReplay uses 100Corrected view uses 96Thin horizontal lines align version events. Arrowed branches identify selected versions, not transfers of money.

At cutoff: USD 100 million
The original was received; the published correction had not reached the application.

After revision receipt: USD 96 million
Corrected reporting uses the revision. It does not rewrite the historical input.

The reporting period, entity, measure, currency and scale stay fixed. Historical replay uses the version actually received by its cutoff; the later corrected view uses the revision.

Check permission for the actual use

Governance determines and enforces appropriate information use. Access is only one part of permission: a source may be readable internally but restricted from external processing or redistribution. Source entitlements specify permitted access and uses; redistribution passes information to another recipient. The general distinction among legal, contractual, organizational, and technical permission is developed in Establish permitted uses.

Public visibility does not remove contractual restrictions. CME's inspected website market-data terms prohibit specified AI uses of covered data and metadata, including training and generating outputs. That example does not describe every exchange or a separately negotiated license. It shows why a programmer must check the actual agreement before putting a visible price into an AI pipeline.

Some information requires restrictions because of its significance to investors, independently of a data license. In U.S. securities terminology, material nonpublic information is information important to a reasonable investor that has not been disseminated so investors generally can access it. The SEC's adopting release explains these concepts; confidentiality alone does not establish materiality. Information barriers restrict flows between functions, including flows to investment decision-makers. Putting restricted records into a broadly searchable internal index can defeat that separation even if the records never leave the institution.

Review each proposed path separately: internal analysis, processing by a model provider, and disclosure to a customer. Name the purpose, recipient, fields, permitted retention, and output use. Combining eligible inputs does not settle whether the combined output may be disclosed. Provider approval must cover the actual service path, not merely a familiar company name; see Approve actual service paths.

Applying AI to financial workflows

Build checked financial research

Financial research turns eligible information into an analysis someone can inspect. A disclosure communicates financial or business information, such as a company filing. An earnings call is a discussion of reported results and business conditions, often including management commentary and questions. A research workflow begins with a specific question, locates relevant sources, aligns comparable facts, performs necessary calculations, and presents a conclusion for analyst review.

Keep the kinds of statement visible. Reported revenue is a source fact; management's expectation is an attributed statement; a ratio is a derived quantity; next year's estimate is a forecast; an assessment of business quality is a judgment. Grounding connects a claim to applicable supporting material. It does not make a forecast true or prove an analyst's interpretation. Check support, not just references explains the general checking task.

Financial vocabulary also changes what counts as support. Loughran and McDonald's 2011 study found that 73.8% of general-dictionary negative-word occurrences in its 10-K corpus came from words typically not negative in financial usage. A liability can be an ordinary accounting category rather than evidence of deteriorating conditions. The lesson is to validate the meaning being measured, not just recognize financial-looking words.

Calculation and interpretation

For numerical analysis, separate deciding what to compute from executing the calculation. The model can identify the needed facts and propose operations; tools retrieve the values and perform the arithmetic. Preserve the selected inputs and operations so the result can be checked. Exact execution still cannot establish that the chosen inputs or formula answer the question. FinQA, introduced in 2021, made this separation explicit by pairing financial questions with executable reasoning programs over report text and tables. Kepler's financial-services approach assigns planning to the model and extraction and arithmetic to tools, retaining each number's derivation.

For example, suppose compatible records report quarterly net income of USD 8 million and revenue of USD 100 million. Under the selected definition, net margin is 8 ÷ 100 = 8%. The calculation service can reproduce that result. The analyst must still establish that both figures concern the same entity and period and that the selected definition answers the question. Concluding that this margin is sustainable introduces a further judgment the division cannot verify.

Comparisons need equal care. Non-GAAP measures use adjustments outside the relevant generally accepted accounting principles. The SEC's interpretations describe how changed adjustments and misleading labels undermine comparison. Similarly named measures need not be equivalent across firms. Preserve definitions and adjustment histories alongside values. A useful research assistant reduces total inspection and correction effort while leaving these decisions visible.

Reconcile records and resolve exceptions

Financial operations must distinguish recording an obligation from fulfilling it. A ledger is an accounting record; posting enters an accounting transaction into it. Settlement completes the agreed transfer of cash or assets. In double-entry bookkeeping, an entry has debit and credit sides whose totals must match. A balanced entry need not move cash: providing a service on credit records revenue and a receivable, an amount the customer owes. Collecting payment later increases cash and reduces the receivable without recording revenue again. The entries balance in both cases, but only collection brings in cash. Balance alone does not prove correct accounts, amounts, or dates.

Reconciliation compares independently maintained records and investigates differences. An exception is an item requiring handling outside the routine path: an unmatched payment, an ambiguous correspondence, or a discrepancy needing explanation. AI can help interpret payment descriptions and assemble likely supporting records. The proposed match must still satisfy the financial correspondence, not merely resemble another description.

Consider a statement line for USD 500 and two internal records for USD 500. Amount equality does not establish which record corresponds to the payment. A transaction reference, counterparty, date, or further investigation must resolve the ambiguity. Dynamics 365's matching documentation makes this concrete: default rules can take the first qualifying transaction, while configuration can require manual matching when multiple documents match on amount. A successful rule execution is not evidence that its selection was correct.

Keep completion states equally precise. In Dynamics' reconciliation workflow, unmatched items can carry forward after a statement is marked reconciled, and completion can trigger correction postings. Separate a proposed match, an approved correction, a confirmed posting, and an unresolved item. Otherwise, an automation-rate dashboard can conceal both open discrepancies and unintended accounting changes.

Resolve the customer's financial problem

Account servicing helps customers understand and manage an existing financial relationship. Agent assist supplies information or drafts to a human service representative. DBS's July 2024 announcement described an assistant combining call transcription, knowledge-base search, summaries, and prefilled service-request fields. Its pilots began in October 2023. This is a concrete assistance workflow; the announcement's projected handling-time savings were not demonstrated end-to-end results.

General product information and account-specific information require different access. Authentication establishes control of an account identity; authorization determines permitted operations. Neither a recognized customer nor a retrieved policy automatically authorizes an account change. See Distinguish people and actors. The service design should identify which record supplies the balance, transaction status, applicable fee, and effective policy rather than letting the model blend them.

A pending transaction has not finished processing or posting. Its amount can change, as with a restaurant tip; it is not interchangeable with a posted entry. Product and channel also matter. Chase's published dispute process requires a credit-card charge to post before opening a dispute. It allows a pending debit-card transaction to be disputed by telephone, while online debit disputes require posting. Applying one rule to every pending transaction gives incorrect instructions.

Dispute handling addresses a contested transaction, not merely a request to explain its status. For covered U.S. electronic fund transfers, the Regulation E interpretation distinguishes checking whether a transfer occurred from alleging an error. Reporting a lost access device together with possible unauthorized use changes the servicing path. These are scoped examples, not universal rules for every payment product.

The assistant should say what is established, avoid promising an unconfirmed refund or completion date, and transfer unresolved facts or out-of-authority requests to a reachable owner. The CFPB's chatbot report describes generic answers and difficult human offramps leaving financial problems unresolved. Count verified resolution, repeat contact, and successful escalation—not just conversations that ended without a person.

Keep scores separate from decisions

Credit underwriting assesses a proposed loan and its terms. Default means failure to meet the relevant contractual obligation. Fraud detection identifies activity that may be deceptive or unauthorized. Both can use a predictive model—behavior fitted from examples, as introduced in Machine Learning Fundamentals—but neither task reduces to producing a score.

An additive credit scorecard assigns points to applicable borrower attributes and sums them. In the SAS implementation, grouped attributes feed a fitted statistical model that is scaled into points. Grouping, fitting, scaling, and the subsequent approval policy are separate choices. A convenient score direction or point total does not itself establish calibration or an appropriate lending threshold.

A decision policy maps information to actions under objectives and constraints. Ranking cases orders attention; a calibrated probability estimates how often a defined outcome occurs among similarly scored cases. Even a meaningful probability does not choose the action. Changing the cost of blocking a legitimate payment relative to missing fraud can change a threshold without changing the prediction. From probabilities to actions develops this distinction.

The Worldline-informed fraud study separates authorization checks from later investigation using rules, learned scores, records, and cardholder contact. An alert is a reason to investigate, not a finding of fraud. In credit review, document synthesis can prepare an assessment, while policy and authorized judgment determine the disposition. Missed losses and unjustified denial or blocking must remain separate consequences in both workflows.

Preserve the actual decision path when explaining an outcome. For covered adverse-action statements, Regulation B requires specific principal reasons reflecting factors actually considered or scored. A plausible explanation generated afterward cannot substitute for those reasons.

Identify the exposure and scenario

Risk exposure is a position or obligation through which an adverse event can cause loss. A counterparty is another party to the financial arrangement. Before choosing a model or explanation, identify the loss mechanism and the time horizon. Several mechanisms can affect the same transaction.

RiskMechanismRelevant evidence
CreditA borrower or counterparty fails to meet obligations.Obligation, due time, outstanding exposure, and recoverable value.
MarketMarket-price movements reduce a position's value.Positions, price sensitivities, and adverse market assumptions.
LiquidityCash or collateral is unavailable when needed, or selling a position materially moves its price.Cash timing, funding access, market depth, and exit horizon.
OperationalA failed process, person, system, or external event causes loss.Routing, approvals, execution records, and failed controls.

Expected credit loss depends on the likelihood of default, the amount exposed at default, and the portion lost after recovery. A ranking of borrowers addresses only part of that assessment. Funding liquidity concerns meeting cash and collateral needs; market liquidity concerns trading without unacceptable price effects. An institution can have assets exceeding liabilities yet lack cash at the required time.

A scenario asks what follows if specified conditions hold. It might assume delayed receipts, reduced recoveries, or difficulty selling assets. Its output is conditional analysis, not a guaranteed forecast. Expected loss also differs from extreme-loss severity: a typical outcome or percentile cutoff can conceal worse losses beyond it. Keep scenario assumptions and the exposure they act on visible so reviewers can judge whether the analysis covers the decision.

Authority and authoritative effects

Authorize the actual financial action

Reading a record, drafting a recommendation, approving it, and submitting an instruction are different permissions. Segregation of duties separates responsibilities that should not be exercised unchecked by one actor. Maker-checker control separates preparation from independent approval; calling a second model does not create that institutional independence. Basel Principle 26 addresses delegation, approval limits, reconciliation, and separation among committing the bank, paying funds, and accounting for assets and liabilities. These supervisory standards are not automatically local law.

Delegated authority should identify who may act for whom, on which accounts or instruments, for what amounts and recipients, and for how long. Enforce these restrictions outside generated content. The model may propose an action; a protected operation must determine whether that action is permitted. Authorization belongs at the protected operation explains the general interface boundary.

Approval must remain attached to the actual proposal through delays and retries. A useful financial precedent is Article 5 of EU Regulation 2018/389: within its strong-customer-authentication scope, the authentication code binds to the agreed amount and payee, and changing either invalidates it. As an engineering rule, changed material arguments, expired approval, or invalidated source conditions should stop submission and require the appropriate renewed decision.

Limits must also cover accumulated exposure. The SEC staff's market-access FAQ describes aggregate credit or capital thresholds and checks before orders enter the market for covered broker-dealers. A per-order limit alone cannot establish an aggregate bound. Test concurrent submissions against shared exposure, including relevant pending commitments; the required property is that individually acceptable requests cannot jointly bypass the limit.

Approval and shared capacity are separate conditions

Fictional instructions share a USD 100,000 limit, with USD 70,000 already committed. Each has its own approval for exactly USD 20,000 to its named recipient. Supporting facts and other permissions are unchanged.

Instruction A after approval
First to reach the shared gate

Instruction A

Approval binds
USD 20,000 · Recipient A
Current instruction
USD 20,000 · Recipient A
Approval lifetime
Current

Instruction B

Approval binds
USD 20,000 · Recipient B
Current instruction
USD 20,000 · Recipient B
Approval lifetime
Current

One atomic check and reservation per arrival. Each instruction sees the exposure left by the previous accepted instruction. Changing a choice recomputes this same two-instruction example.

ArrivalBeforeIf addedAfter gateDisposition
Instruction AUSD 70,000USD 90,000USD 90,000Eligible for submission: pending exposure reserved.
Instruction BUSD 90,000USD 110,000USD 90,000Shared limit exceeded: held without reservation.
1 of 2 instructions eligible for submission. Reserved exposure: USD 90,000 of USD 100,000. Held instructions reserve nothing; accepted pending commitments remain counted.

A lower amount is still a different action. Renewed review is necessary for changed or expired approvals; it would not remove the independent shared-limit check. This example ends at submission eligibility, with no payment or settlement result.

This proposed engineering control uses fictional values, not a universal legal threshold or a claim about a provider implementation. Exact-action approval and current aggregate capacity are independent requirements.

Customer communications deserve explicit authority too. Morgan Stanley's June 2024 Debrief announcement describes drafting an email that an advisor may edit and send at their discretion. It separately describes saving a note into Salesforce. A review boundary for one output does not establish the boundary for every write the application performs.

Connect terms, entitlement, and receipt

A system of record is the designated authority for a particular business fact. One source may own instrument terms, another accounting entries, another external transaction status, and another investigation status. An assistant's summary is a derived view of these records. Name the authoritative records explains the general ownership principle; a financial integration must also preserve identifiers, exact monetary values, units, relevant dates, and evidence of completed effects.

Consider bond-interest reconciliation as a teaching example. A bond is debt issued to investors. Principal, also called face value, is the amount repayable under its terms; interest is the contractual payment for borrowing. These are promised payments, not guaranteed receipts, and market value can differ from principal. The SEC's bond guide introduces this vocabulary.

A custodian holds or administers clients' assets. Entitlement is the payment due under event rules. Review identified, dated terms, eligible holdings, calculations, and external and internal records. Today's holding times a displayed rate is insufficient: eligibility and contractual conventions matter. Keep licensed terms and restricted holdings on permitted processing paths.

The Depository Trust Company's (DTC) Distributions Service Guide separates announcements, entitlements, and allocations of received funds. Payable dates do not prove receipt; position or rate corrections can generate credits or debits. Participant allocation does not prove bank receipt by the beneficial owner, the investor whose assets are held. Reconcile corresponding external allocations and internal postings, retaining unresolved differences.

This schematic separates expected entitlement from independently owned allocation and posting records. Reconciliation maps account identities for the same instrument and event; agreement does not establish source truth or confirm beneficial-owner cash receipt. Adjustments remain linked to original records.

Resolve uncertain execution

A lost response changes what the application knows, not necessarily what the provider did. Keep unknown outcome separate from confirmed failure without an effect and confirmed completion. Preserve the logical operation identity while consulting authoritative records. A fresh payment request can duplicate an effect that already occurred.

Idempotency is a receiver-defined contract for repeated requests. Stripe's contract reuses the first executed request's stored status and body when the same key and parameters recur, including stored failures. Parameter changes are rejected. Keys can be pruned once at least 24 hours old; reuse after pruning creates a new request. Validation failures and concurrent execution conflicts do not create saved results. This is bounded duplicate protection, not permanent exactly-once settlement.

Stripe's error guidance recommends retrying network failures with the same key and parameters using backoff. A server error can remain indeterminate despite a repeatable cached response. Provider reconciliation and subsequent events may reveal resulting objects; a local operation identifier in metadata helps correlate them. Preserve pending state until authoritative evidence resolves it rather than changing the key to escape uncertainty.

The effect and knowledge of it can diverge

One possible run of logical operation P. Time runs downward; spacing does not encode duration. The request arrives and creates an effect, but its response is lost.

Provider effect versus application knowledgeSubmit P with key K and unchanged parameters. Provider creates an effect; its response is lost. Application remains unknown. Same-key retry can return the stored response under the provider contract. A separately correlated authoritative record establishes the specific effect, not settlement.ApplicationProviderSubmit P · key K · unchanged parametersEffect occurred for PResponse lostOutcome unknownEffect is not undoneRetry P · same K · same parametersStored response under the provider contractA repeated response may still leave the outcome unresolved.Correlate authoritative record / event with PSpecific provider effect confirmedNo settlement claimAuthoritative record linked to P

Other possible resolutions: confirmed effect, confirmed no effect, or still unknown. These are alternatives, not three later events in this run.

The lost response travels back from the provider; it does not undo the effect. Retrying retains operation P, key K and parameters. A correlated authoritative record establishes only the specific effect it records.

Recovery can create new financial work. A compensating entry records a correction instead of erasing the original accounting event; Dynamics documents reversal through a new transaction rather than editing an already reconciled statement. Track the original effect and the remedy separately, including whether the remedy completed. See Recover from the effects that occurred for the wider recovery pattern.

Evidence of useful performance

Measure the financial workflow

An evaluation assesses specified behavior against an intended use. A baseline is the credible alternative: existing rules, software, models, and human work—not necessarily the newest competing model. Define the assessment unit before choosing a metric. A correct answer, a correct match, and a resolved case describe different units and support different claims. What an evaluation establishes and Define worthwhile improvement provide the general framework.

Use these as assessment contracts to specify locally, not as interchangeable quality scores.
Workflow and unitChecksComplete-work comparison
Research: claim and analyst taskSource support, numerical correctness, omitted material, and justified interpretation.Coverage and total analyst time, including checking and correction.
Operations: match and reconciliation caseCorrect correspondence, confirmed postings, unresolved amounts, and deadlines.Complete handling effort, remaining exceptions, and correction burden.
Service: customer problemApplicable information, authorized actions, resolution, and usable escalation.Repeat contact, waiting, access to help, and customer consequences.
Decision support: decision and affected populationPolicy adherence, decision-relevant errors, and outcomes after sufficient follow-up.Losses, unjustified restrictions, group differences, and investigation effort.

FinanceBench, introduced by Pranab Islam and colleagues in November 2023, paired financial questions with answers and supporting passages. Its experiments assessed 2,400 outputs across 16 configurations on 150 cases. Providing the correct evidence pages improved results but did not eliminate reasoning errors; qualitative review also found valid answers differing from references. It is useful evidence about the separation of retrieval, reasoning, and assessment, not certification for a new institution's workflow.

For an investigation queue, precision is the fraction of flagged cases confirmed to have the specified condition. The condition matters: an anomaly is not necessarily fraud. In her cross-document compliance presentation, Varsha Shah reported approximately 91% precision in an evaluation involving roughly three million records over five years and four jurisdictions. Confirmation concerned genuine anomalies, not uniformly established fraud. The report did not supply the labeling process or evaluation split details. This result concerns the quality of flagged cases; it does not establish how much fraud went unflagged, how long investigations took, or whether the workflow produced a net financial benefit.

Respect the historical information boundary

Backtesting assesses behavior on historical situations. Look-ahead bias occurs when later information enters a decision represented as historical. A chronological split helps, but only if source selection, preprocessing, and model selection respect each cutoff. Rolling-origin evaluation repeats this process as the decision date advances and scores the forecast horizon actually needed. See Validation without leakage.

Different historical distortions require different repairs.
DistortionWhat it changesEvidence to retain
Later source revisionsLater values enter historical decisions.Versions and availability cutoffs.
Later knowledge in the modelHistorical source restrictions do not restrict everything learned during training.Model lineage and tests for prohibited later information.
Survivorship biasToday's survivors replace the population eligible historically.Historical membership, identifier intervals, distributions, and exits.
Selection overfittingThe reported winner was chosen from many historical experiments.Alternatives tried and independent assessment of the selected configuration.

The model itself can cross the cutoff. Lookahead Bias in Pretrained Language Models found later pandemic information in risk descriptions generated from pre-pandemic earnings calls. Instructions to ignore later facts did not eliminate the demonstrated leakage. Restricting retrieved documents therefore supports only the retrieval boundary, not a complete historical-information guarantee.

Survivorship bias conditions inclusion on a later outcome. Removing securities that subsequently disappeared can omit investments available at the time, including failures and their losses; not every delisting is a failure, since mergers also cause exits. Separately, repeatedly selecting the strongest historical configuration can overfit the selection process. An attractive winning run does not disclose how many alternatives lost.

Keep representative later-period assessment separate from stress testing, which deliberately examines severe but plausible conditions. Stress scenarios reveal vulnerabilities; their frequency in a test set is not an estimate of how often they occur in ordinary work. Both are useful, but they answer different deployment questions.

Keep missing and delayed outcomes visible

A label is the specified outcome or judgment used for assessment. Selective labels arise when earlier decisions determine whose outcomes become observable. Repayment on an institution's loans is observed only after it grants those loans; rejected applicants do not become successful repayment examples. Evaluating solely on approved borrowers can misrepresent a replacement policy serving a different population. The selective-labels research formalized this problem in another decision domain; lending has the same observation structure.

This limitation predates modern AI. Durand's 1941 credit study explicitly noted that its records contained previously approved loans and omitted important information. More sophisticated modeling does not manufacture the missing outcomes. Account for incomplete feedback explains the general problem.

Selection and delay are separate. An outcome horizon is the follow-up period required by a claim. For default within twelve months, an account observed without default for only three months remains unresolved. It is right-censored: observation ended before the eventual event time became known. Treating it as a twelve-month negative silently changes the target.

Approval selects whose repayment is observed

Example

Historical rejection and insufficient follow-up create different gaps in outcome evidence.

Historical applicationsThe population encountered by the prior decision policy.
Granted loansCredit was extended and repayment can subsequently be observed.
Rejected applicationsNo loan was extended by this institution.
Loan outcome unobservedNot a repayment success or failure label for the proposed loan.
Outcome follow-upCompare observed events and observation duration with the specified target.
Target outcome assessableThe target event occurred, or sufficient event-free follow-up completed.
Outcome still unresolvedNo event observed, but follow-up ends before the target horizon.
For default within twelve months on loans granted by this institution, rejected applications lack the proposed loan’s outcome. An observed default within the horizon is assessable immediately; a negative label needs twelve event-free months. Three event-free months remain unresolved. Branch widths do not encode prevalence.
Read the diagram as text
  • Historical applications. The population encountered by the prior decision policy.
  • Granted loans. Credit was extended and repayment can subsequently be observed.
  • Rejected applications. No loan was extended by this institution.
  • Loan outcome unobserved. Not a repayment success or failure label for the proposed loan.
  • Outcome follow-up. Compare observed events and observation duration with the specified target.
  • Target outcome assessable. The target event occurred, or sufficient event-free follow-up completed.
  • Outcome still unresolved. No event observed, but follow-up ends before the target horizon.
  • Historical applications → Granted loans: historical policy: grant.
  • Historical applications → Rejected applications: historical policy: reject.
  • Rejected applications → Loan outcome unobserved: no corresponding loan outcome.
  • Granted loans → Outcome follow-up: observe subsequent records.
  • Outcome follow-up → Target outcome assessable: event observed or horizon completed.
  • Outcome follow-up → Outcome still unresolved: event-free observation ends early.

Payment outcomes require equally explicit definitions. Stripe's dispute lifecycle distinguishes payment time, later notification, evidence review, and changing case status. No observed dispute at a cutoff is not confirmed non-fraud. Its fraudulent dispute category records an allegation of unauthorized payment, which can include a legitimate charge the customer did not recognize. A reported fraud dispute, a won or lost case, and independently investigated fraud are different labels.

Investigation capacity also selects feedback. The Worldline-informed study assessed precision among the limited number of cards investigators could inspect, rather than treating multiple transactions on one card as independent investigations. Retain both the review-selection rule and outcome maturity. Better results on the selected queue do not establish performance on everything outside it.

Institutional responsibility and intervention

Assign ownership and independent challenge

Model risk concerns harm from an incorrect or misused model. Its intended use defines the decisions, population, and conditions for which it is accepted. Validation assesses suitability for that use; effective challenge requires expertise, objectivity, and enough influence to change the decision. Documentation without the power to require remediation is not equivalent to challenge.

The IIA's Three Lines Model distinguishes responsibilities, not mandatory department names.
ResponsibilityRole in deployment
Business managementOwn day-to-day risks and operate controls.
Risk and compliance specialistsSupport, challenge, and monitor the business's handling of risk.
Internal auditProvide independent assurance on controls, with functional reporting to the board.

Independent assurance assesses whether controls are appropriate and effective; it is not another execution stage inside the model's workflow. A payment team checking duplicates operates a control. Internal audit examining that control performs a different responsibility. Supplier involvement does not dissolve the institution's own decisions about use, monitoring, or response.

Scope each obligation before translating it into a release gate. The Federal Reserve's April 2026 SR 26-2 model-risk guidance is nonbinding, risk-based guidance primarily relevant to banks above its stated asset threshold, with exceptions. Its scope explicitly excludes generative and agentic AI. Applying its validation and monitoring practices to those systems is an engineering analogy, not a GenAI mandate; historical SR 11-7 should not be presented as the current guidance.

Other sources have different authority. FINRA Notice 24-09 reminds member firms that existing technology-neutral obligations apply to generative AI, including third-party tools; it creates no new requirements. Covered Regulation B notices must give actual principal decision reasons. Data licenses create contractual restrictions, while an institution can impose stricter internal approval policies. Record which category supports each requirement, its applicability, owner, and enforcement evidence. See Verify and revisit decisions.

Make review capable of changing the outcome

Human participation matters only if the person can make the needed intervention. Factual review checks evidence and calculations. Policy review determines whether the proposed treatment fits applicable rules. Authorization grants permission for an effect. Takeover assigns someone responsibility for unresolved work. A reviewer may be qualified for one role but not another.

Provide a review packet suited to that decision: applicable source versions, calculation inputs and definition, unresolved discrepancies, proposed effects, and approval limits. For bond reconciliation, distinguish the expected entitlement, external allocation evidence, and internal posting. A summary should help the reviewer reach those records, not replace them. Review before consequential commitment develops the interface choices.

Automation bias is inappropriate reliance on automated output. In a Duolingo false-alert experiment, skilled reviewers accepted half of fabricated cheating alerts inserted into legitimate historical sessions. Customers were not affected by the experiment. This is cross-domain evidence that approval is not a correctness measure, not an estimate of financial-review error. Test local reviewers with independently assessed correct and erroneous proposals, measuring whether they detect and correct consequential errors before commitment.

Review capacity constrains the permitted workload. Track required handling effort, qualified staffing, aging, and the time remaining before financial or customer deadlines. When demand exceeds capacity, invoking human review does not supply an answer: admission, prioritization, escalation, or the operating mode must change. Urgent work still needs an accountable owner. Operate the exception workload explains queue management.

Correction also needs a route after the initial decision. NIST's voluntary AI Risk Management Framework supports feedback, appeal, and override mechanisms. A practical design links a complaint to the decision and its versions, assigns a reviewer empowered to change the disposition, records the rationale, and tracks downstream correction and communication. A trace explains recorded events; it does not itself give an affected person a remedy.

Deployment and continued control

Choose a bounded operating mode

Deployment is a choice about permitted effects, not a race toward autonomy. Integration tests establish interface behavior. Historical assessment examines past cases under a defined information boundary. Shadow operation processes current inputs without applying candidate effects. Assisted use introduces the candidate into human work. Limited live action exposes a bounded set of real operations. Each answers a different remaining question.

Modes can remain separate products or operating choices; they need not be successive stages.
ModeEvidence it can addWhat remains unestablished
Historical assessmentBehavior on reconstructed past information and cases.Performance under today's work and consequences of applying outputs.
ShadowCurrent-input compatibility, proposed outputs, and resource demand.Effects on customers, reviewers, and financial outcomes when proposals are used.
Assisted useReviewer handling, corrected results, and actual use of suggestions.Safety or usefulness without that human intervention.
Bounded live actionAuthorized effects and recovery within the exposed scope.Suitability for broader populations, higher limits, or new activities.

Shadow execution must technically deny or isolate writes, notifications, and financial actions; discarding the final answer does not undo tool effects. The live-experiment chapter explains this boundary. Define eligible work, allowed sources and recipients, permitted effects, aggregate exposure, intervention capacity, acceptance conditions, and stopping rules before exposure. Read-only assistance can remain the appropriate final mode.

Compare the complete process with its incumbent: source licensing, integration, validation, model calls, review, correction, support, and waiting. Faster case preparation can be valuable while failing to reduce total handling effort. Price accepted outcomes develops that accounting. The historical Authorizer's Assistant report treated integration, acceptance testing, and transfer of business control to operations as deployment work—not as consequences automatically supplied by a working model.

Reconstruct, monitor, and reassess

An audit trail is attributable evidence connecting information, decisions, approvals, and observed effects. For financial AI, preserve the relevant source and instrument versions, as-of conditions, model and policy versions, checked calculations, reviewers, authority, execution attempts, and resulting external identifiers. Generated reasoning is not a substitute for records of what was read, approved, and done. Record meaning-changing boundaries explains the diagnostic design.

The SEC's 2022 electronic-recordkeeping amendments provide a scoped example: an audit-trail alternative preserves modifications, deletions, timestamps, applicable identities, and information needed to recreate original and intermediate records. It is not a universal requirement to retain every prompt. The NIST Privacy Framework also calls for minimizing audit data. Preserve necessary evidence with controlled access and defined lifetimes; do not indiscriminately copy customer payloads into logs. See Set artifact lifetimes.

Monitoring should identify which accepted deployment claim needs reconsideration.
Observed changeReassessment
Stale sources, corrections, or changed rightsWhether inputs remain applicable and permitted, and which outputs need review.
Growing unresolved amounts or older casesWhether reconciliation and intervention capacity still meet the operating contract.
Complaints, overrides, or changing outcomesWhether service and decision quality remain acceptable in the affected population.
New provider, product, policy, or authorityWhether previous tests and approvals cover the changed system.
Failed limits or shutdown activationWhat further work stopped, what already executed, and who must reconcile it.

Containment and recovery are separate. A runtime kill switch can stop future decisions or tool calls when active work next checks it; a flag read only at session creation misses ongoing work. Stopping does not reverse payments. Identify completed and uncertain effects, reconcile authoritative records, authorize corrections, test recovery, and obtain approval to restore the affected scope. Record who changed controls and whether mitigation actually worked.

Finally, examine dependencies beyond one application. The FSB's 2024 assessment identifies provider concentration, correlated behavior, cyber risk, and model risk as possible amplification channels. Local usefulness does not remove a shared failure dependency. Continued permission to operate rests on maintained evidence, workable intervention, and tested recovery—not the age of the launch approval.

Open questions

  1. End-to-end financial productivity remains difficult to establish because faster preparation can shift work into verification, corrections, and support. Useful progress would compare complete analyst or reconciliation tasks against the incumbent process, preserving quality and unresolved-work measures alongside total effort.

  2. Effective financial oversight needs evidence that qualified reviewers detect consequential errors under realistic workload and deadline pressure. Approval rates cannot supply that evidence. Progress would demonstrate correct intervention before commitment, including overload and unavailable-reviewer conditions.

  3. Historical language-model evaluation faces an information boundary that document filtering cannot fully enforce: later facts may already be encoded in the model. Progress would establish auditable historical-knowledge constraints or evaluation designs that do not depend on pretending those facts are absent.

  4. Outcome learning must handle decisions that hide alternatives and labels that mature late. Progress would make observation rules, unresolved cases, and deployment-population coverage explicit, rather than treating increasingly large operational datasets as automatically representative.

  5. Local controls do not resolve shared infrastructure risk. Institutions need ways to test substitution and recovery without silently changing data permissions or financial behavior. Progress would demonstrate continuity across actual dependencies, not merely configure an alternate provider name.

Follow the curated reading path through the speakers and demonstrations behind this entry.

23 min

AI Engineer World's Fair 2026 · 2026

How Kepler Built Verifiable AI for Financial Services

Vinoo Ganesh

Cited in this entry

Develops the separation between model-directed analysis, deterministic numerical execution, and replayable derivation histories. Read it for financial calculation architecture, not as a substitute for validating inputs and measure definitions.

Watch talk
18 min

AI Engineer World's Fair 2026 · 2026

Your Agent Didn’t Fail. Your Harness Did.

Vinoth Govindarajan

Cited in this entry

Connects scoped approvals, durable state ownership, execution attempts, and external completion evidence. These distinctions are especially useful when financial operations continue through retries or delayed callbacks.

Watch talk

Explore more talks

The rest of the library, beyond the curated path. Cited talks support this entry; reviewed transcripts were processed in full. Metadata candidates have not been reviewed as sources or verified as topic members.

125 matching talks

TalkSpeakerEventYear
Varsha ShahAI Engineer World's Fair 20262026
Waseem AlshikhAI Engineer Summit 20252025
Leo PekelisAI Engineer World's Fair 20242024
Shreya Rajpal, Aman GuptaAI Engineer World's Fair 20262026
Ritvik PandyaAI Engineer World's Fair 20262026
Devendra Chaplot, Devendra Singh ChaplotAI Engineer World's Fair 20242024
Ahmed MenshawyAI Engineer World's Fair 20242024
Cedric ClyburnAI Engineer World's Fair 20262026
Roy DerksAI Engineer Summit 20252025
Daniel WhitenackAI Engineer World's Fair 20242024
Lawrence JonesAI Engineer Europe 20262026
Lance MartinAI Engineer World's Fair 20242024
Udi MenkesAI Engineer World's Fair 20262026
Jesse HuAI Engineer Code 20252025
Parth AsawaAI Engineer World's Fair 20262026
Sandipan BhaumikAI Engineer Europe 20262026
Agents Need Feature Flags

Cited in this entry

Sachin GuptaAI Engineer World's Fair 20262026
Christopher Lovejoy, Saul HowardAI Engineer World's Fair 20262026
Sumaiya ShrabonyAI Engineer World's Fair 20262026
AI’s Jurassic Park Period

Transcript reviewed

Aaron StanleyAI Engineer World's Fair 20262026
Angus J. McLeanAI Engineer Europe 20262026
Sarthak AggarwalAI Engineer World's Fair 20262026
Chaitanya AsawaAI Engineer World's Fair 20262026
Ofer MendelevitchAI Engineer World's Fair 20252025
Vaibhav Page, Infant VasanthAI Engineer World's Fair 20252025
State of Data

Transcript reviewed

Sean CaiAI Engineer World's Fair 20262026
Dan MasonAI Engineer World's Fair 20252025
Jeremy Silva, Chris HernandezAI Engineer World's Fair 20252025
Stephen ChinAI Engineer Europe 20262026
Nishant GuptaAI Engineer World's Fair 20262026
Sachin GuptaAI Engineer World's Fair 20262026
Giran Moodley, Mayan Soni, Oussama Hafferssas, Mayank SoniAI Engineer Europe 20262026
Frank CoyleAI Engineer World's Fair 20262026
Corey CooperAI Engineer World's Fair 20252025
Divakar KumarAI Engineer World's Fair 20262026
Christopher LovejoyAI Engineer Summit 20252025
Anant ShankhdharAI Engineer World's Fair 20262026
Martin Harrysson, Natasha ManiarAI Engineer Code 20252025
Balázs HorváthAI Engineer World's Fair 20262026
Ramana Siddanth EmaniAI Engineer World's Fair 20262026
Ayush BhardwajAI Engineer World's Fair 20262026
Vibhor KumarAI Engineer World's Fair 20242024
Nathan WanAI Engineer World's Fair 20252025
Vasuman MozaAI Engineer World's Fair 20262026
Shawn ChanAI Engineer World's Fair 20262026
Anju KambadurAI Engineer Summit 20252025
Rachna SrivastavaAI Engineer World's Fair 20252025
Sahil Yadav, Hariharan GanesanAI Engineer World's Fair 20252025
Sheila Gulati, Nischal NadhamuniAI Engineer World's Fair 20242024
Nina Lopatina, Rajiv ShahAI Engineer World's Fair 20252025
Fuzzing in the GenAI Era

Metadata candidate

Leonard TangAI Engineer World's Fair 20252025
Mustafa Ali, Kyle CorbittAI Engineer Summit 20252025
Mitesh PatelAI Engineer World's Fair 20252025
Identity for AI Agents

Metadata candidate

AI Engineer Code 20252025
Juan Herreros ElorzaAI Engineer Europe 20262026
Yuval Belfer, Niv GranotAI Engineer World's Fair 20252025
Kshitij GroverAI Engineer World's Fair 20252025
Scaffold Wisely

Metadata candidate

Rahul SengottuveluAI Engineer Summit 20252025
Laurie VossAI Engineer Europe 20262026
Yogendra MirajeAI Engineer World's Fair 20262026
Kobie CrawfordAI Engineer Europe 20262026
Stop Using RAG as Memory

Metadata candidate

Daniel ChalefAI Engineer World's Fair 20252025
Jim BennettAI Engineer World's Fair 20252025
Nuno CamposAI Engineer Europe 20262026
Sandipan BhaumikAI Engineer Europe 20262026
Emil EifremAI Engineer World's Fair 20262026
Lucas PalmaAI Engineer World's Fair 20262026
Peter GostevAI Engineer Europe 20262026
Soumith ChintalaAI Engineer Summit 20252025
Rustin BanksAI Engineer World's Fair 20252025
Shawn "swyx" WangAI Engineer World's Fair 20262026
Armanas PovilionisAI Engineer World's Fair 20262026
Apoorva JoshiAI Engineer World's Fair 20262026
Brendan RappazzoAI Engineer World's Fair 20262026
Steve KaliskiAI Engineer Europe 20262026
Justin McCartyAI Engineer World's Fair 20242024
Jess Grogan-Avignon, Jack WangAI Engineer Europe 20262026
Jeronim MorinaAI Engineer World's Fair 20242024
Elizabeth Fuentes LeoneAI Engineer World's Fair 20262026
Jan CurnAI Engineer World's Fair 20252025
Nik CaryotakisAI Engineer Summit 20252025
Jesús BarrasaAI Engineer World's Fair 20252025
Lincoln MurrEvent unknownUnknown
Anekha SokhalEvent unknownUnknown
Harneet KaurAI Engineer New York 20262026
Jay Sahnan, Arik BirdEvent unknownUnknown
Hari KumarEvent unknownUnknown
Dan Barrett, Raymond Weitekamp, Jose Montes de OcaEvent unknownUnknown
Noyan TokgozogluEvent unknownUnknown
Howie Liu, Shawn "swyx" Wang, Joshua XuAI Engineer World's Fair 20262026
Soukaïna AbelhadAI Engineer NYC 2026 Online Talks2026
Jordi SprangerAI Engineer NYC 2026 Online Talks2026
Gareth YoderEvent unknownUnknown
George HeEvent unknownUnknown
Michael Manapat, Jack McDonaldEvent unknownUnknown
Freddy LecueEvent unknownUnknown
Laurie VossAI Engineer World's Fair 20262026
Michele GrantAI Engineer NYC 2026 Online Talks2026
Ryan AhernEvent unknownUnknown
Ankur DuggalAI Engineer World's Fair 20262026
Shy RuparelAI Engineer New York 20262026
Janie Gu, Suyeol YunEvent unknownUnknown
Jack WotherspoonEvent unknownUnknown
Jake BroekhuizenAI Engineer World's Fair 20262026
Priyanka Phatak, Gabe CemajAI Engineer World's Fair 20262026
Somya Rai, MD ShariqueAI Engineer NYC 2026 Online Talks2026
Nina Žižakić, Eleonora DrykinaAI Engineer NYC 2026 Online Talks2026
Joseph KimEvent unknownUnknown
Corey GallonAI Engineer NYC 2026 Online Talks2026
Al ChenEvent unknownUnknown
Karthic RaoAI Engineer NYC 2026 Online Talks2026
Ryan AhernEvent unknownUnknown
Vinoo GaneshEvent unknownUnknown
Nikhil Benesch, Aaron LinskyEvent unknownUnknown
Annabell Schäfer, Doneyli De JesusEvent unknownUnknown
Jesse HallEvent unknownUnknown
Charles DickensAI Engineer World's Fair 20262026
Dor SassonAI Engineer World's Fair 20262026
Paden GayleAI Engineer NYC 2026 Online Talks2026
Sam Green, Aaron LinskyEvent unknownUnknown
Laurie VossEvent unknownUnknown
Aman SinghEvent unknownUnknown
Sam ParsonsAI Engineer World's Fair 20262026
Nacho MartinezEvent unknownUnknown
Matthew RagoneAI Engineer NYC 2026 Online Talks2026

References

Coverage and source review
Processed transcripts
44 processed in full · 5 in the curated path
Automated source review
Passed
Metadata candidates
86 unreviewed; not verified topic membership
Corpus version
1bd8e407b26a07b33815594e1b2db5f41827119a2b3cb6fbf240f9fc571fc767

Automated review checks source support; it is not publication approval.

A synthesis of selected conference talks and technical references. Citations link to the source material; they do not imply that every talk on this subject is included.

  1. AASB 132 Financial Instruments: Presentation

    https://standards.aasb.gov.au/aasb-132-dec-2022

  2. Basel Core Principles: Principle 26, Internal control and audit

    https://www.bis.org/committees/bcbs/basel-framework/standard/bcp/40/inforce/2024-04-25/published/2024-04-25

  3. Artificial Intelligence in UK Financial Services — 2024

    https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024

  4. FINRA Regulatory Notice 24-09

    https://www.finra.org/rules-guidance/notices/24-09

  5. Microsoft Dynamics 365: Reconcile bank statements using advanced bank reconciliation

    https://learn.microsoft.com/en-us/dynamics365/finance/cash-bank-management/reconcile-bank-statements-advanced-bank-reconciliation

  6. CFPB: Chatbots in consumer finance

    https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/

  7. The Authorizer’s Assistant: A Knowledge-Based Credit Authorization System for American Express

    https://cdn.aaai.org/IAAI/1989/IAAI89-028.pdf

  8. Credit Card Fraud Detection: A Realistic Modeling and a Novel Learning Strategy

    https://dalpozz.github.io/static/pdf/TNNLS_2017.pdf

  9. David Durand: Credit-Rating Formulae

    https://www.scribd.com/document/974920274/c9265

  10. When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks

    https://www.uts.edu.au/globalassets/sites/default/files/adg_cons2015_loughran-mcdonald-je-2011.pdf

  11. GLEIF and ANNA: ISIN Issuer to LEI Mapping Table Factsheet

    https://www.gleif.org/lei-data/lei-mapping/download-isin-to-lei-relationship-files/2021-06-08_isin-to-lei_factsheet_v1.pdf

  12. DTC Distributions Service Guide

    https://www.dtcc.com/-/media/Files/Downloads/legal/service-guides/Service-Guide-Distributions.pdf

  13. SEC Beginners’ Guide to Financial Statements

    https://www.sec.gov/about/reports-publications/investorpubsbegfinstmtguide

  14. XBRL International: xBRL-JSON tutorial and examples

    https://www.xbrl.org/guidance/xbrl-json-tutorial/

  15. FinQA: A Dataset of Numerical Reasoning over Financial Data

    https://arxiv.org/html/2109.00122

  16. FRED real-time periods and historical information sets

    https://fred.stlouisfed.org/docs/api/fred/realtime_period.html

  17. IFRS Foundation: IAS 8 Basis of Preparation of Financial Statements

    https://www.ifrs.org/issued-standards/list-of-standards/ias-8-basis-of-preparation-of-financial-statements/

  18. SEC: EDGAR Application Programming Interfaces

    https://www.sec.gov/search-filings/edgar-application-programming-interfaces

  19. CME Group: Market Data Explanation and Disclaimer

    https://www.cmegroup.com/trading/market-data-explanation-disclaimer.html

  20. NIST Privacy Framework 1.0: lifecycle and minimized audit evidence

    https://nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.01162020.pdf

  21. SEC: Selective Disclosure and Insider Trading

    https://www.sec.gov/rules-regulations/2000/08/selective-disclosure-insider-trading

  22. How to Build Planning Agents Without Losing Control - Yogendra Miraje, FactSet

    https://www.youtube.com/watch?v=sl3icG-IjHo&t=595

  23. Trust, but Verify: High-Fidelity Reasoning in Agentic Workflows

    https://www.youtube.com/watch?v=MWTJIAwAAnk&t=551

  24. How Kepler Built Verifiable AI for Financial Services

    https://www.youtube.com/watch?v=Tt2kX2sgQio&t=760

  25. How Kepler Built Verifiable AI for Financial Services

    https://www.youtube.com/watch?v=Tt2kX2sgQio&t=858

  26. SEC: Non-GAAP Financial Measures Compliance and Disclosure Interpretations

    https://www.sec.gov/rules-regulations/staff-guidance/corporation-finance-interpretations/non-gaap-financial-measures

  27. Trust, but Verify: High-Fidelity Reasoning in Agentic Workflows

    https://www.youtube.com/watch?v=MWTJIAwAAnk&t=618

  28. Double-entry bookkeeping and cash movement

    https://www.accountingcoach.com/debits-and-credits/explanation

  29. Basel Committee: credit risk and contractual obligations

    https://www.bis.org/publ/bcbs75.pdf

  30. Microsoft Dynamics 365: Set up bank reconciliation matching rules

    https://learn.microsoft.com/en-us/dynamics365/finance/cash-bank-management/set-up-bank-reconciliation-matching-rules

  31. AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection

    https://www.youtube.com/watch?v=Iwe_RY-fYgI&t=953

  32. DBS Empowers Its Customer Service Officers with a Gen AI-Powered Virtual Assistant

    https://www.dbs.com/NewsPrinter.page?locale=en&newsId=lyq3tqw8

  33. Your Agent Didn’t Fail. Your Harness Did.

    https://www.youtube.com/watch?v=BInpv7lGp1o&t=58

  34. Chase: What are pending transactions and how long do they take?

    https://www.chase.com/personal/credit-cards/education/basics/pending-transactions

  35. Chase: Disputing a Charge

    https://www.chase.com/personal/credit-cards/dispute

  36. CFPB Regulation E: § 1005.11 Procedures for resolving errors

    https://www.consumerfinance.gov/rules-policy/regulations/1005/11/

  37. SAS credit scorecards: grouped attributes and additive points

    https://support.sas.com/documentation/cdl/en/emcsgs/66024/HTML/default/p0vkqpxn5cceoqn1m5697l4rw7lj.htm

  38. Conditional estimates versus decision thresholds

    https://scikit-learn.org/stable/modules/classification_threshold.html

  39. Calibration, reliability diagrams and discrimination

    https://scikit-learn.org/stable/modules/calibration.html

  40. CFPB Regulation B: § 1002.9 Notifications

    https://www.consumerfinance.gov/rules-policy/regulations/1002/9/

  41. Basel Committee: funding and market liquidity

    https://www.bis.org/publ/bcbs144.pdf

  42. Basel Committee: sound management of operational risk

    https://www.bis.org/bcbs/publ/d515.pdf

  43. Basel Committee: Explanatory Note on Minimum Capital Requirements for Market Risk

    https://www.bis.org/bcbs/publ/d457_note.pdf

  44. Federal Reserve: Credit loss-estimation methodologies

    https://www.federalreserve.gov/bankinforeg/stress-tests/ccar/august-2013-estimation-methodologies-for-losses-revenues-and-expenses.htm

  45. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    https://www.youtube.com/watch?v=q-WOjZhOMCA&t=755

  46. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    https://www.youtube.com/watch?v=q-WOjZhOMCA&t=842

  47. Commission Delegated Regulation (EU) 2018/389, Article 5

    https://eur-lex.europa.eu/legal-content/en/ALL/?uri=CELEX:32018R0389

  48. Your Agent Didn’t Fail. Your Harness Did.

    https://www.youtube.com/watch?v=BInpv7lGp1o&t=717

  49. SEC Staff FAQs Concerning Risk Management Controls for Brokers or Dealers with Market Access

    https://www.sec.gov/rules-regulations/staff-guidance/trading-markets-frequently-asked-questions/divisionsmarketregfaq-0

  50. Morgan Stanley: Launch of AI @ Morgan Stanley Debrief

    https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch

  51. Your Agent Didn’t Fail. Your Harness Did.

    https://www.youtube.com/watch?v=BInpv7lGp1o&t=882

  52. Investor.gov: Bonds — FAQs

    https://www.investor.gov/introduction-investing/investing-basics/investment-products/bonds-or-fixed-income-products/bonds

  53. Your Agent Didn’t Fail. Your Harness Did.

    https://www.youtube.com/watch?v=BInpv7lGp1o&t=793

  54. Resolving an ambiguous payment request

    https://docs.stripe.com/error-low-level

  55. Stripe retry keys and retention boundaries

    https://docs.stripe.com/api/idempotent_requests

  56. Production Evals For Agentic AI Systems

    https://www.youtube.com/watch?v=vljxQZfJ9wY&t=78

  57. Production Evals For Agentic AI Systems

    https://www.youtube.com/watch?v=vljxQZfJ9wY&t=163

  58. FinanceBench: A New Benchmark for Financial Question Answering

    https://arxiv.org/html/2311.11944v1

  59. The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables

    https://cs.stanford.edu/people/jure/pubs/contraction-kdd17.pdf

  60. AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection

    https://www.youtube.com/watch?v=Iwe_RY-fYgI&t=588

  61. Rolling-origin forecast evaluation

    https://otexts.com/fpp3/tscv.html

  62. Lookahead Bias in Pretrained Language Models

    https://openreview.net/pdf/dc9c551b0ea15f5d9ca91ae6c4109b5cd9fe6f13.pdf

  63. CRSP10 historical security, distribution and delisting records

    https://www.crsp.org/wp-content/uploads/2023/08/CRSP10-User-Guide.pdf

  64. The Probability of Backtest Overfitting

    https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf

  65. Basel Committee: Principles for Operational Resilience

    https://www.bis.org/bcbs/publ/d516.pdf

  66. Right censoring and unresolved event outcomes

    https://scikit-survival.readthedocs.io/en/stable/user_guide/00-introduction.html

  67. Stripe card-dispute timing and lifecycle

    https://docs.stripe.com/disputes/how-disputes-work

  68. Stripe dispute categories and contested fraud claims

    https://docs.stripe.com/disputes/categories

  69. Federal Reserve SR 26-2: Revised Guidance on Model Risk Management

    https://www.federalreserve.gov/supervisionreg/srletters/SR2602a1.pdf

  70. The IIA’s Refreshed Three Lines Model

    https://www.theiia.org/en/content/podcast/getting-started-with/2026/ep-0040/

  71. The Build-Operate Divide: Bridging Product Vision and AI Operational Reality

    https://www.youtube.com/watch?v=1__V4KTv_Gw&t=251

  72. Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo

    https://www.youtube.com/watch?v=CDqzWpwkSls&t=359

  73. Mission-Critical Evals at Scale: Learnings from 100,000 Medical Decisions

    https://www.youtube.com/watch?v=cZ5ZJy19KMo&t=482

  74. NIST AI RMF 1.0: accountability, appeals and override

    https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf

  75. Production Evals For Agentic AI Systems

    https://www.youtube.com/watch?v=vljxQZfJ9wY&t=225

  76. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    https://www.youtube.com/watch?v=q-WOjZhOMCA&t=105

  77. IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork

    https://www.youtube.com/watch?v=q-WOjZhOMCA&t=676

  78. Moving away from Agile: What's Next?

    https://www.youtube.com/watch?v=SZStlIhyTCY&t=244

  79. The Build-Operate Divide: Bridging Product Vision and AI Operational Reality

    https://www.youtube.com/watch?v=1__V4KTv_Gw&t=129

  80. SEC: Electronic Recordkeeping Requirements, Release 34-96034

    https://www.sec.gov/files/rules/final/2022/34-96034.pdf

  81. Agents Need Feature Flags

    https://www.youtube.com/watch?v=zU4EagB311U&t=1097

  82. Agents Need Feature Flags

    https://www.youtube.com/watch?v=zU4EagB311U&t=579

  83. FSB: Financial stability implications of artificial intelligence

    https://www.fsb.org/2024/11/fsb-assesses-the-financial-stability-implications-of-artificial-intelligence/

  84. Event time and actual availability require separate handling

    https://github.com/feast-dev/feast/issues/6615

  85. How to Build Planning Agents Without Losing Control - Yogendra Miraje, FactSet

    https://www.youtube.com/watch?v=sl3icG-IjHo&t=480

  86. FICO Celebrates 70 Years of Innovation and Says “Hello, Future” to Applied Intelligence

    https://investors.fico.com/news-releases/news-release-details/fico-celebrates-70-years-innovation-and-says-hello-future/

  87. BloombergGPT: A Large Language Model for Finance

    https://arxiv.org/html/2303.17564