Insights
Notes from the build.
Writing on building AI for finance, and the parts that tend to go wrong once it hits production.
Field guide
AI architecture for finance
Retrieval, agents and evaluation, designed around the constraints finance imposes.
ReadField guide
The financial data layer
Point-in-time correctness, lineage and entity resolution, the layer most projects skip and then regret.
ReadField guide
AI governance & compliance in finance
The EU AI Act, model risk management and supervisory expectations, turned into engineering you can ship.
ReadField guide
AI financial modeling & automation
Forecasting, variance explanation and document extraction, wired into the workflow your team already runs.
ReadField guide
AI for fraud & financial crime
Real-time fraud scoring, AML alert triage and KYB/sanctions screening, built to survive a model examiner and a false-positive budget.
ReadField guide
AI finance-operations automation
Reconciliation, AP/AR, invoice processing and month-end close, automated with a human-in-the-loop review path finance can actually sign off.
ReadField guide
AI for credit & lending operations
Cash-flow underwriting, decision engines and CI/CD for credit policy, governed ML decisioning without a data-science army.
ReadField guide
AI for treasury & payments
Cash visibility, liquidity forecasting and payment operations, built on ISO 20022 data, bank APIs and instant rails.
ReadField guide
Private AI & self-hosted LLMs
Running language models on your own infrastructure: hardware, model selection, RAG, security and the cost math.
ReadAir-gapped AI deployments
What a real air gap requires: controlled model transfer, offline RAG, monitoring without vendor telemetry, and the hidden dependencies that break offline stacks.
ReadFine-tuning vs RAG for private data
A decision order for private enterprise data: when retrieval wins, when fine-tuning wins, what LoRA changes, and the maintenance cost each choice commits you to.
GPU procurement for private AI
Buy, lease, or rent: 2026 street prices and rental benchmarks, the utilisation breakpoint that decides it, and the power, lead-time, and contract risks behind the quote.
LLM evaluation and observability on private infrastructure
Self-hosted eval harnesses, prompt regression testing and OpenTelemetry-shaped tracing for teams whose prompts and completions cannot reach a SaaS backend.
Hardening a self-hosted LLM stack
Five layers that decide whether a private LLM deployment survives review: prompt isolation, weights provenance, container isolation, secrets handling, and audit logging.
Sizing on-prem LLM inference hardware
VRAM math, quantization trade-offs, GPU tiers from L40S to B200, and throughput estimation for on-premises inference, with the cases where CPU is enough.
Choosing open-weight models in 2026
Llama 4, Qwen 3, DeepSeek, Mistral and gpt-oss compared on licence terms, hardware footprint and task fit, with an evaluation method that survives procurement.
Private AI vs API: the actual cost crossover
Break-even for self-hosted LLMs sits far higher than most teams assume. The published 2026 thresholds, the hidden costs that move them, and when privacy decides instead.
Private AI in construction and industrial firms
Why bid pricing, RFIs and safety records push contractors toward private AI, which workflows repay the effort first, and what the 2026 ROI evidence actually shows.
Private LLM deployment architecture
On-premises, private VPC, and hybrid topologies compared: reference stack by layer, multi-tenancy options, and the criteria that decide which one you build.
RAG on private infrastructure
A working reference stack for self-hosted retrieval: vector store choice, private embeddings, chunking, hybrid search, permission-aware filtering, and evaluation.
What sovereign AI requires beyond data residency
Residency, sovereignty and operational sovereignty are three different guarantees. What GDPR, the AI Act and Gulf procurement actually demand, and how vendor claims fail.
MCP servers for financial data
Banks and fintechs are shipping MCP servers for banking, market and compliance data. The architecture, authorisation and governance that survive audit.
Spreadsheet agents in financial modelling
Excel's Agent Mode now builds and edits models inside the workbook. What it speeds up, how it fails, and the controls a finance team should require.
Controls for agent-initiated payments
AI agents are starting to pay for things. What AP2 mandates and verifiable credentials mean for merchant risk, issuer controls, and dispute evidence.
Detecting AI-generated documents before they clear
Forged bank statements, paystubs, and invoices now survive visual review. A layered defence: file forensics, consistency checks, external verification.
AI on the trading desk: ESMA raises the evidence bar
ESMA's February 2026 supervisory briefing converges scrutiny of algorithmic and AI-assisted trading. Firms must now evidence controls in live trading.
Climate risk in PD and LGD models
Supervisors now expect climate risk inside PD and LGD models, not beside them. The transmission channels, scenario mechanics and overlays that hold up.
AI vendor contracts after the liability directive
The AI Liability Directive is withdrawn. The revised PLD, the AI Act, and DORA now decide how AI risk is allocated in financial services vendor contracts.
PSD3 and the PSR: what to build before the rules apply
The EU payments package is agreed but not yet in force. What the 2028 application window means for payee verification, fraud reimbursement, and SCA.
US AI guidance for financial institutions, mid-2026
SR 26-2 replaced SR 11-7 and carved generative AI out of model risk. Treasury, the SEC, and Colorado fill the gap. How to map the layered US regime.
Transfer pricing documentation as a data pipeline
Master File, Local File, and CbCR now share inputs with Pillar Two. Why TP documentation automation is becoming compliance orchestration, not drafting.
The Digital Omnibus and the new EU AI Act timeline
The Digital Omnibus moves Annex III high-risk obligations to December 2027. What still applies now, what can wait, and what finance teams should keep building.
Lease accounting: extracting the data that drives it
IFRS 16 and ASC 842 run on terms buried in unstructured contracts. How lease abstraction, remeasurement, and evidence trails hold together in practice.
Agentic orchestration of the month-end close
Close agents are being sold as autonomous accountants. The credible version is an orchestrator that sequences tasks, drafts work, and stops at approval gates.
Flux analysis automation that survives audit
AI flux tools now detect variances, find drivers, and draft commentary. What separates audit-ready explanations from generated filler in the close package.
CSRD reporting automation after Omnibus I
Omnibus I narrowed CSRD scope and cut ESRS datapoints by more than 60%. What that means for data pipelines, AI drafting, and assurance-ready reporting.
Automating financial consolidation without losing control
Intercompany eliminations, ownership rollups, and currency translation still run on spreadsheets in many groups. Where automation holds up and where it breaks.
AI-assisted XBRL tagging that survives review
AI taggers can pre-tag an iXBRL filing in minutes, but SEC and ESEF filings still fail on semantics, extensions, and validation. The workflow that holds up.
E-invoicing mandates are an AP architecture change
Germany's B2B e-invoicing mandate turns invoice intake into structured data ingestion. What that does to AP capture, coding, and exception handling.
Running SEPA Instant as an operating model
SEPA Instant is a 24/7 liquidity and controls problem wearing a payments badge. Prefunding, 10-second budgets, and real-time risk for the always-on rail.
Payment orchestration and smart routing
Routing each transaction across acquirers by approval odds, cost, and FX turns payments into a decision problem. What the layer does and when it pays.
Stablecoin settlement for corporate treasury
GENIUS and MiCA turned stablecoins into regulated settlement rails. Where they fit in a treasury stack, and the accounting and custody traps that remain.
Bank connectivity: APIs versus files
Premium bank APIs have not retired SFTP, EBICS, or SWIFT FileAct. How to split treasury flows across channels by latency, control, and volume.
Verification of Payee: surviving the rollout
EU VoP is live for euro-area PSPs since October 2025. What the name-check actually demands from matching engines, corporate bulk files, and master data.
Virtual accounts and the in-house bank
Virtual accounts centralise cash while keeping entity-level attribution. Where they beat physical accounts, and where they add a layer of complexity.
Automating FX exposure management
Most FX automation fails at the exposure data, not the hedging. How to build the register, netting, and policy checks before letting AI near a trade ticket.
Intraday liquidity forecasting in practice
End-of-day cash forecasts miss the hours when funding problems actually happen. How to nowcast intraday positions and wire every forecast to an action.
Payment repair automation that clears the queue
Repair queues grow because exceptions outpace the people working them. How a rules-plus-ML repair layer classifies failures, enriches data, and raises STP.
ISO 20022 data after the migration
The CBPR+ cutover ended in November 2025. The returns now come from using structured payment data for reconciliation, exceptions, and analytics.
AI intake for commercial underwriting submissions
Broker submissions arrive as email threads and attachments. How intake AI builds structured submission objects, triage queues and appetite scores.
AI monitoring for private credit portfolios
Private credit passed $2 trillion while monitoring stayed in spreadsheets. Where AI helps with ingestion, covenants, marks and early-warning signals.
Compliance guardrails for AI in debt collection
Collections is where automation meets the tightest conduct rules. Here is how we build contact and treatment automation that stays inside them.
Dynamic credit limit management with ML
A static limit is wrong the day after you set it. Here is how we build limit increase and decrease models that stay inside affordability and conduct rules.
Generating explainable adverse-action reasons at scale
Every decline owes a specific, accurate reason. Here is how we generate ECOA-compliant adverse-action reasons from complex models, reviewably.
Risk-based pricing for lending with ML
Price for risk and you must prove the price is fair and explainable. Here is how we build risk-based pricing that survives a fair-lending review.
Fair-lending monitoring in production
Fair lending is not a launch test you pass once. Here is the production monitoring we build so drift into disparate impact is caught early.
Automating loan document processing
A loan file is a stack of documents that must agree with each other. Here is the extraction, validation and cross-check automation we build.
Early-warning signals for a loan portfolio
The best time to act on a deteriorating loan is before it is late. Here is how we build early-warning signals from behaviour and external data.
Reject inference done responsibly
You only see repayment for the loans you approved. Here is how we handle reject inference without fooling ourselves about the model's real performance.
Optimising collections strategy with ML
Not every delinquent account needs the same call. Here is how we prioritise and tailor collections while staying inside conduct rules.
Automating loan covenant monitoring
Covenants breach quietly between reporting dates. Here is how we extract, track and test covenants so a breach is caught when it happens.
Aggregating open-banking data you can build on
Aggregated bank data is inconsistent across providers and formats. Here is the normalisation and enrichment layer we build before any model sees it.
Automating financial spreading for SME lending
Spreading SME financials by hand is slow and inconsistent. Here is the extraction and normalisation pipeline we build to standardise the numbers.
Intelligent document processing for finance, end to end
IDP is more than OCR. Here is the classify, extract, validate and route pipeline we build so finance documents flow without a keying team.
Building a transaction categorization engine
Categorising transactions powers budgeting, underwriting and reconciliation alike. Here is the model and feedback loop we build for accurate categories.
Automating audit workpapers and evidence
Audit is evidence collection at scale. Here is how we automate sampling, tie-outs and workpaper assembly while keeping the trail a reviewer trusts.
Automating revenue recognition under ASC 606
Rev rec turns messy contracts into scheduled revenue under strict rules. Here is where AI extracts the terms and where a controller must still decide.
Automating subscription and usage billing
Usage billing breaks on proration, upgrades and disputes. Here is the billing and reconciliation automation we build so revenue ties out.
AI for treasury cash positioning and forecasting
A treasurer needs tomorrow's cash position, not last month's. Here is the categorisation and short-horizon forecasting we build for treasury.
Automating fixed-asset accounting
Fixed assets accrue depreciation, disposals and impairments nobody wants to track by hand. Here is how we automate the subledger with a review path.
Automating payment-operations exceptions
Payments succeed until they do not, and the exceptions eat the team. Here is how we triage returns, repairs and mismatches with a review path.
Spend controls and approval routing with AI
Approval chains are where spend policy meets reality. Here is how we automate routing and enforce controls without stalling the business.
Cleaning vendor and customer master data with AI
Duplicate and mismatched master records cause duplicate payments and broken matching. Here is the entity-resolution approach we use to clean them.
Automating purchase orders end to end
The PO is where spend gets controlled or lost. Here is the requisition, approval and matching automation we build so procurement scales without chaos.
Automating the tax provision with AI support
The tax provision is reconciliation and classification under a hard deadline. Where AI shortens the work, and where a preparer must still own every number.
Automating vendor onboarding and verification
Onboarding a vendor means verification, tax forms and bank details fraudsters target. Here is the automation and controls we build around it.
Behavioural biometrics for continuous authentication
A password proves who logged in, not who is typing now. Here is how we use behavioural biometrics for continuous, privacy-aware authentication.
Transaction monitoring for crypto and stablecoin flows
On-chain data is public and unfamiliar. Here is how we build monitoring that ties chain analytics to your customers and your reporting obligations.
First-party vs third-party fraud: why the split matters
The same loss looks different depending on who caused it. Here is why separating first-party from third-party fraud changes the model and the treatment.
Contract intelligence for procurement and finance
Obligations, renewals and price terms hide in PDFs nobody rereads. Here is the extraction and monitoring layer we build over a contract portfolio.
Adverse-media screening that cuts the noise
Name-match adverse-media screening buries analysts in irrelevant hits. Here is how entity resolution and relevance scoring make it usable.
Graph embeddings for AML and fraud
Hand-built graph features miss patterns embeddings can learn. Here is how we use graph representation learning for AML without losing explainability.
Perpetual KYC: automating ongoing customer due diligence
Periodic KYC reviews are a backlog machine. Here is how we move to event-driven, perpetual KYC with a human review path for real changes.
Automating chargeback and dispute handling
Disputes are a document and evidence problem at volume. Here is the extraction, evidence-assembly and outcome-prediction workflow we build for disputes.
Detecting friendly fraud and refund abuse
Friendly fraud looks like a real customer because it is one. Here is how we separate genuine disputes from abuse without punishing good customers.
Velocity rules and rate limiting for fraud, done right
Velocity rules are the oldest fraud control and the easiest to get wrong. Here is how we tune them against a false-positive budget alongside models.
Risk-scoring merchants at onboarding
A bad merchant approved today is chargebacks and laundering tomorrow. Here is the data and scoring we build for acquirer and PSP onboarding.
Detecting money-mule accounts
Mule accounts look normal until the money moves through them. Here are the behavioural and network signals we use to find them before the payout.
Detecting scams and authorised push-payment fraud
In APP fraud the customer authorises the payment. Here is how we score intent and coercion signals to intervene without blocking legitimate transfers.
Detecting transaction laundering in payments
Transaction laundering hides illegal sales inside a legitimate merchant account. Here are the signals and models we use to surface it.
Opex and headcount planning with AI
Headcount is most of opex and the hardest line to plan. Here is how we build driver-based headcount and opex plans that update with hiring reality.
Revenue and demand forecasting for operating companies
Revenue splits into hundreds of small series with their own drivers. Here is the hierarchical, driver-aware forecasting approach we build for operating finance teams.
Catching fraud rings with graph networks
Per-transaction scoring misses coordinated fraud because the signal lives in the connections between accounts. How we use graph features and GNNs to surface the rings.
Optimising working capital with ML
Cash trapped in receivables and inventory is expensive. Here is how we model DSO, DPO and inventory to free working capital.
Using LLMs to review spreadsheet models
Spreadsheet errors hide in plain sight. Here is how we use models to flag broken references, inconsistent logic and hard-coded overrides for a human to check.
Price and fee optimization with ML, within the rules
Pricing is a model and a compliance surface at once. Here is how we optimise fees and rates while staying inside fair-pricing constraints.
AI support for zero-based budgeting
Zero-based budgeting is thorough and exhausting. Here is where AI does the assembly and analysis while owners keep the decisions.
Anomaly detection in the financial close
A wrong number found after close is expensive. Here is the anomaly detection we wire into the close to catch it before sign-off.
Cohort and retention analytics with AI support
Averages hide what cohorts reveal. Here is how we build cohort and retention analysis that explains why a vintage behaves the way it does.
Automating narrative financial reporting
The numbers are done; the commentary takes the afternoon. Here is how we generate grounded, reviewable narrative reporting that cites its figures.
Churn prediction that a fintech can act on
A churn score nobody acts on is a dashboard. Here is how we build churn models tied to interventions and measured against real retention.
Customer lifetime value modeling for fintech
LTV drives acquisition spend and credit limits alike. Here is how we build LTV models that hold up when the cohort and the economy move.
Driver-based planning automation with AI
A plan is a set of drivers, not a spreadsheet of hard-codes. Here is how we automate driver-based planning so the forecast updates as reality does.
AI-assisted scenario and stress testing
Stress testing is judgment plus mechanics. Here is where AI accelerates scenario generation and analysis while a human owns the assumptions.
Running AI incident postmortems in finance
A model incident without a postmortem is an incident you will repeat. Here is the blameless postmortem process we run for finance AI failures.
Consumer Duty and AI in UK financial services
The FCA's Consumer Duty reshapes what an AI-driven decision must be able to show. Here is how we engineer for good outcomes and fair value.
Time-series foundation models for financial forecasting
Pretrained forecasters promise zero-shot accuracy. Here is where they beat classical models on finance series and where a tuned baseline still wins.
An AI governance operating model for a finance team
Governance fails when it is a committee with no wiring. Here is the operating model, roles and gates we help finance teams put around AI.
AI vendor due diligence: the questions that matter
Buying an AI vendor imports their model risk into your firm. Here is the due-diligence checklist we use before a finance team signs.
GDPR and automated decisions in finance AI
Article 22 and the right to an explanation shape what a finance model may decide alone. Here is how we engineer for it without stalling the workflow.
Designing human oversight for high-risk AI
The EU AI Act demands effective human oversight, not a rubber stamp. Here is how we design review points a person can actually exercise.
When the document is the attacker: securing RAG in finance
Retrieval pulls semi-trusted filings and emails straight into the prompt. Here is how we treat that content as untrusted and build a layered defence for finance AI.
A bias audit workflow for lending models
Disparate-impact testing, the less-discriminatory-alternative search, and production monitoring: the fairness workflow we run for credit models.
Writing EU AI Act technical documentation that holds up
High-risk classification means a documentation package, not a slide. Here is what Annex IV actually asks for and how we assemble it as engineering evidence.
Model cards that actually help finance governance
A model card is documentation a reviewer can act on, not marketing. Here is what we put on one for a finance model, and why each field earns its place.
Audit trails that make finance AI reproducible
When an examiner asks why the model said that, you need to reconstruct it exactly. Here is the logging and versioning that makes a decision reproducible.
Red-teaming AI systems in financial services
Adversaries probe finance AI for jailbreaks, data leaks and biased outputs. Here is how we red-team a system before an attacker or an examiner does.
Shadow AI: governing the tools staff already use
Your team is pasting data into chatbots whether you approved it or not. Here is the usage policy and controls we help finance firms put around it.
Continuous model monitoring that satisfies a regulator
Annual revalidation is not enough for a model that drifts weekly. Here is the monitoring, alerting and evidence trail we build for supervised finance models.
Data quality observability for finance AI
A broken feed shows up as a wrong answer three layers downstream. Here is the freshness, volume and schema monitoring we put on finance data.
DORA incident response when AI is the ICT risk
Under DORA, an AI outage is an ICT incident with reporting clocks. Here is the detection, classification and reporting workflow we build for AI systems.
Building a golden customer record for finance AI
Every system holds a partial view of the same customer. Here is the entity-resolution and survivorship approach we use to build one trusted record.
Standing up an AI model inventory and registry
You cannot govern what you cannot list. Here is the model inventory, registry and metadata we build so risk and audit can see every model in production.
Chunking strategies for financial documents
Naive chunking splits a table from its header and a covenant from its clause. Here is how we chunk filings, contracts and statements so retrieval stays coherent.
Versioning embeddings and reindexing without downtime
Change the embedding model and every stored vector is stale. Here is how we version embeddings and reindex a finance corpus without breaking retrieval.
Monitoring feature drift before it breaks the model
Models fail quietly when their inputs move. Here is the feature-level drift monitoring we wire in so you catch it before the output goes wrong.
BaFin now reads AI as ICT risk: what DORA asks of your models
BaFin's December 2025 guidance files AI inside DORA's ICT risk regime, not a separate ethics box. Here is what that means for registers, incidents and testing.
A data catalog so finance AI can find the right table
Models are only as good as the data teams can find and trust. Here is the catalog, metadata and discovery layer we build for finance AI.
Why AI in finance should be data-layer first
The model is the easy part. The reason finance AI projects stall is almost always the data beneath them. Here is how we approach it.
PII redaction and DLP for finance AI
Customer data flows toward the model whether you planned it or not. Here is the redaction, tokenisation and DLP layer we build so it does not leak.
Reference data management for finance AI
Currencies, calendars, instruments and codes are the quiet backbone of every model. Here is how we manage reference data so nothing silently drifts.
Change data capture for finance AI pipelines
Repolling source systems is slow and lossy. Here is how we use change data capture to keep features and retrieval fresh without hammering the ledger.
A semantic layer so AI speaks your finance metrics
Ask two systems for revenue and get two numbers. Here is how a governed semantic layer gives models and agents one definition of every metric.
Using synthetic data in finance AI, carefully
Synthetic data can fill gaps and leak reality at once. Here is where we use it for finance models, and the tests that keep it honest.
Batch or real-time? Inference patterns for finance AI
Not every score needs to be computed at request time. Here is how we split batch precompute from real-time inference to hit cost and latency targets.
Data contracts for finance AI pipelines
A silent schema change upstream breaks a model downstream. Here is how we use data contracts to make finance feeds safe to build AI on.
Making function calling reliable in finance workflows
A model that calls the wrong tool with the wrong arguments is a production incident. Here is how we make function calling dependable in finance systems.
Choosing a vector store for financial documents
The vector database is not where your project lives or dies, but the wrong choice adds latency and cost. Here is how we pick one for finance retrieval.
GraphRAG for financial entities and relationships
When relationships matter more than similarity, a vector store is the wrong tool. Here is how we use graph retrieval over issuers, counterparties and ownership.
An LLM gateway for access, cost and governance
Ungoverned model access is a cost and compliance leak. Here is the gateway we put in front of providers for keys, quotas, logging and policy.
Long context or retrieval? Choosing for finance documents
Bigger context windows do not retire retrieval. Here is how we decide between stuffing the context and retrieving, for filings, contracts and statements.
Eval-driven development for finance AI
Write the eval before the feature. Here is how we run finance AI like test-driven development, with a scored eval set gating every change.
Caching strategies for LLM systems in finance
The cheapest model call is the one you do not make. Here are the prompt, semantic and result caches we use to cut finance LLM cost and latency safely.
Building point-in-time feature pipelines
Leakage is the quiet killer of finance models. Here is how we build feature pipelines that only ever see what was known as of the event timestamp.
Where small language models beat frontier models in finance
Not every finance task needs a frontier model. Here is where a fine-tuned small model wins on cost, latency and control, and where it does not.
Governing agent tool use in finance workflows
An agent that can call tools can move money. Here is the authorization, rate-limiting and action-validation layer we require before an agent touches a finance system.
Multi-agent orchestration in finance, and when to avoid it
One capable agent beats a swarm more often than the demos suggest. Here is where we split work across agents in a finance system and where a single loop is safer.
On-prem, VPC or API: hosting AI in regulated finance
Where the model runs is a compliance decision as much as a technical one. Here is how we weigh API, VPC and self-hosted models for regulated finance data.
Reranking: the retrieval step most finance RAG skips
First-pass retrieval returns plausible passages; reranking returns the right one. Here is how we add a reranker to finance RAG without blowing the latency budget.
A reference architecture for finance AI systems
The same shape recurs across finance AI builds. Here is the reference architecture we start from: data layer, retrieval, model, guardrails, human review and audit.
When the model is someone else's: third-party AI and vendor model risk
A hosted model API is an ICT service, not a feature. Here is how we treat vendor model risk, concentration, and silent updates under DORA-style scrutiny.
Model routing and fallback for cost and reliability
One model for every call wastes money and breaks under load. Here is the routing, cascade and fallback design we use across a finance workload.
Output validation patterns for finance AI
A confident wrong answer is the failure mode that matters. Here are the schema, constraint and verification layers we put between a model and a finance workflow.
Where agentic AI belongs in finance, and where it does not
Agentic AI is spreading fast across finance teams. The useful question is not whether to adopt it, but which decisions you let an agent make alone and which a person signs.
LLM observability and tracing for finance workloads
You cannot debug what you cannot see. Here is the tracing, logging and evaluation-in-production setup we use to keep a finance LLM system diagnosable.
Building a RAG evaluation harness for finance
Vibes are not an eval. Here is the retrieval + generation test harness we stand up so a finance RAG system has a number that moves before it ships.
What the EU AI Act's August 2026 deadline asks you to build
Credit scoring and insurance-pricing systems are headed into the high-risk regime, and the date is now in flux. Here is what that means for the system you are building.
MiCA for stablecoins and crypto payments: what engineering teams must build
MiCA turns crypto-asset rules into concrete controls for issuers and payment firms. Here is the monitoring, reporting and data-lineage engineering it actually requires.
A finance-operations AI glossary: the terms that actually matter
Feature store, entity resolution, straight-through processing, champion/challenger, SAR, model drift, human-in-the-loop: defined for finance, each linked to the guide that goes deeper.
If a credit model can't explain a denial, it can't issue one
ECOA Regulation B requires specific reasons within 30 days of a credit denial. Here is what that demands from the model that made the decision, not the paperwork around it.
Measuring AI ROI in finance operations: the metrics CFOs accept
Hours saved is not a business case. Here are the payback, quality and risk metrics we use to prove a finance-ops AI system earns its keep in production.
Hybrid retrieval for filings, transcripts and analyst reports
Pure vector search loses the exact tickers and defined terms that financial answers hinge on. Here is how we build hybrid RAG that an analyst can trace back to the source line.
Build vs buy for finance-ops AI: a decision framework and TCO model
Point tool, platform or engineering-led build? Here is the framework we give finance and risk leaders to decide, including the total-cost and control trade-offs.
Choosing a model for finance workloads, starting from the task
A leaderboard cannot tell you which model fits your finance workload. The task and an eval set can. Here is how we pick, and what the tradeoffs actually cost.
Auto-decisioning safely: where to draw the human-in-the-loop line
Approving 80% of applications automatically is only safe if the thresholds are right. Here is how we set auto-decision bands, referral rules and override capture.
Champion/challenger and shadow deployment for underwriting models
You cannot A/B test lending decisions carelessly. Here is how we run challenger and shadow models in production to validate a new scorecard without harming customers.
Generating portfolio commentary that an analyst will sign
Automated portfolio commentary works only when every sentence is tied to an attribution number. Here is how we build it so an analyst can review and sign off.
CI/CD for credit policy: deploy decisioning changes in days, not months
Changing a credit policy should not need an engineering release cycle. Here is the versioning, testing and shadow-run pipeline we build so risk teams ship safely.
Decision engine vs rule engine: what changes when models enter lending
A decision engine does far more than run a rule tree with a model score attached. Here is how the two differ and how we migrate legacy credit rules into governed ML decisioning.
How to actually test a credit or pricing model for fairness
Dropping a protected attribute does not make a model fair. Here is how we test credit and pricing models for disparate impact and proxy discrimination, and keep the evidence.
Bank-statement parsing and income verification at production accuracy
Statement parsing demos fail on real PDFs and formats. Here is the extraction, validation and benchmarking method we use to trust an income figure.
Turning "it looks good" into a number before a finance AI system ships
A finance AI system needs a defensible accuracy number before it goes live. Here is the eval set we build, the baseline we hold it to, and why audit cares.
Cash-flow underwriting: turning bank transactions into affordability signals
Open-banking and bank-statement data is raw noise until you categorise, dedupe and derive income and affordability. Here is the pipeline we build for lenders.
Cash-flow forecasting with machine learning, without over-engineering it
A point forecast nobody can act on is worse than a range nobody argued with. How we build cash flow forecasting that treasury and FP&A actually use.
Insurance claims triage automation with false-positive control
Claims triage has to move fast without paying uncovered claims or dragging honest claimants into investigation. The extraction and scoring design we build.
Intercompany reconciliation: matching across entities and currencies
Intercompany breaks hide in timing, FX and mismatched references across ledgers. Here is the entity-resolution and matching approach we use to clear them before close.
Model risk management when the model is an LLM
SR 11-7 was written for deterministic models. Its 2026 replacement, SR 26-2, modernised model risk management but left generative AI out of scope. Here is how to extend it to an LLM.
Automating expense and invoice audit with AI
Expense audit is sampling because humans cannot check everything. Here is how we score every line for policy breaches and duplicate or fraudulent claims.
Controlling LLM cost when finance asks per-query
An LLM bill that looks fine in aggregate hides where the money goes. Here is how we keep per-query cost low enough to survive a CFO's questions.
AR automation: cash application and collections that learn
Applying a payment with no clean remittance is a matching problem, and collections is a prioritisation problem. Here is how we automate both with a review path.
Market data quality is a model problem
Gaps, stale ticks and vendor restatements don't crash an AI system. They make it confidently wrong. Here is how we catch quality issues before the model does.
Three-way matching AI: PO, receipt and invoice without the manual chase
Three-way match breaks on partial deliveries, unit mismatches and split invoices. Here is the matching logic and exception routing we build so AP does not stall.
Accelerating the month-end close with AI, without losing control
The close is a dependency graph of reconciliations, accruals and reviews. Here is where AI actually shortens it and where a human must still sign the number.
When the same company isn't: entity resolution across financial data feeds
Market, filings and alternative-data feeds disagree about what counts as one issuer. Here is how we build a resolution layer that holds up at quarter-end.
Bank reconciliation automation: deterministic matching first, LLM for the tail
Most reconciliation should never touch a model. Here is how we split exact matching from the fuzzy exceptions an LLM handles, with an audit trail a controller trusts.
Explainability methods that survive a model validation review
Most explainability work satisfies the data scientist who built the model and nobody else. Here is how we make it hold up for a validator, a regulator and the customer who got declined.
AP invoice automation: an architecture that survives the messy tail
Invoice capture demos look easy until real vendor formats arrive. Here is the extraction, validation and human-review architecture we use to hit high straight-through rates.
Account takeover detection: the signals that actually work
Account takeover hides inside legitimate sessions. Here are the behavioural and device features, and the model design, we use to catch it without locking out real users.
Making LLM output safe to feed a finance system
A free-text number from a model should never flow straight into a ledger. Here is the method we use to make LLM output reliable enough to act on.
Defending finance workflows against deepfakes and AI-agent abuse
Attackers now use generated voices, documents and their own agents against onboarding and payments. Here are the failure modes and the guardrails we build in.
Pulling structured ESG metrics out of unstructured reports
ESG numbers live in PDFs written by people who disagree on definitions. Here is how we extract, normalise and validate them without inventing precision that is not there.
Continuous forecasting and the engineering behind AI variance commentary
Continuous forecasting only works when the forecast is wired to live actuals and every variance explanation cites a driver and a number. Here is how we build it.
Consortium data vs first-party data for fraud models
Network signals catch fraud your own data cannot see, but they add governance and lineage debt. Here is how we decide what to source where and keep it auditable.
Drafting SAR narratives with LLMs, safely
A suspicious activity report narrative is a legal document, not a chatbot answer. Here is how we generate grounded, reviewable SAR drafts that cite their evidence.
Build the analyst a copilot, not a replacement
Finance models that AI drafts still need an analyst to own the assumptions. Here is where a copilot earns its keep and where the human signs off.
Automating KYB: business verification without the manual pack
KYB means resolving a business across registries, UBOs and documents nobody standardises. Here is the extraction, entity-resolution and risk-scoring pipeline we build for onboarding.
Building alternative-data pipelines that hold up
Web-extracted signals, card panels, satellite, job postings and scraped filings only pay off when the pipeline gets point-in-time capture and entity mapping right.
Detecting synthetic-identity fraud at application time
Synthetic identities pass traditional KYC because the pieces are real. Here are the data and model patterns we use to catch fabricated applicants before they book.
Sanctions screening that stops matching on spelling
Legacy sanctions screening fires on every fuzzy name match. Here is how entity resolution and context features cut hit volume while defending true-positive recall.
Shaping financial data so retrieval returns the right figure
Retrieval over financial data fails when storage ignores how the questions split. Here is how we structure a warehouse and a vector store so answers stay exact and traceable.
Automating AML alert triage with a human in the loop
Most AML alerts are noise, but you must review them anyway. Here is how we automate enrichment, prioritisation and disposition drafting while keeping the analyst accountable.
Reading earnings calls with LLMs without fooling yourself
Earnings-call transcripts are full of signal an LLM can surface. The hard part is making those signals traceable, point-in-time correct, and worth acting on.
Real-time payment fraud scoring: an architecture for sub-second decisions
Scoring a payment for fraud in under a second means the model is the easy part. Here is the streaming feature path, latency budget and fallback design we build around it.
When a financial team needs a feature store
A feature store earns its keep when point-in-time correctness and reuse start to hurt. Here is how we decide whether your team has hit that point yet.
Cutting false positives in transaction monitoring with ML
Rule-based AML monitoring drowns analysts in false alerts. Here is the feature engineering and model layering we use to cut false positives without missing real suspicious activity.
Extraction that holds up at quarter-end, not just in the demo
A clean PDF extracts in one call. Thousands of filings and scanned contracts with inconsistent layouts do not. Here is how we build extraction that an auditor can check.
Drafting the credit memo with AI, signing it without one
A first-draft credit memo can be assembled from financials, filings and internal data in minutes. The credit decision stays with the officer. Here is how we keep both true.
How a missed corporate action quietly corrupts your financial dataset
Splits, spin-offs and reused tickers break price history and entity matching long before anyone notices. Here is how we handle corporate actions so a backtest stays honest.