Applied & Generative AI for Finance
Production AI for finance
We build AI systems for finance work that must survive an audit, a regulator, and a bad quarter.
- Document AIany format
- Generative copilotsgrounded in your data
- RAGanswers cite sources
- Governedmodel risk & audit
- Forecasting& anomaly detection
Why teams bring us in
Three situations account for most of the AI work that reaches us.
The pilot never left the lab
A team ships a demo that impresses a steering committee, then discovers nobody will put it in front of customers without logging, access control, and a rollback plan. We pick up that work. The model stays, the surrounding system gets built properly, and the launch date stops moving.
Manual work that grew with headcount
Analysts rekey statements, invoices, and KYC packets into core systems, and volume growth means more hires. Document AI moves the extraction, but the money is in the exception path, which is where accuracy thresholds, dual review, and clean audit trails matter. We design that path first, then the model around it.
Regulators asked for a paper trail
A scoring model sits in production with no record of which version made which decision, and internal audit wants answers this quarter. We reconstruct the lineage, wire versioning into the serving path, and document how each feature influences an outcome so credit and compliance reviewers can defend it.
Use cases and governance
Scope and control get decided together, before anyone opens a notebook.
Use case selection with a number attached
We start with a shortlist and kill most of it. Each surviving idea gets a baseline you can measure against, the touch time on a claim today, the hours spent closing a reconciliation, the share of policy questions that reach a human. Anything without a baseline waits, because there is no honest way to report on it later.
Data readiness before model work
Policy answers buried in scanned PDFs and a data warehouse with three definitions of customer will sink any assistant. Our engineers profile the sources, fix the identifiers, and build the chunking and metadata that retrieval depends on. Access rules get carried through, so a relationship manager in one region never retrieves a document they cannot open in the source system.
Model risk governance your reviewers recognize
Validation teams want documentation, challenger comparisons, and evidence that someone watches the model after launch. We produce the artifacts as part of delivery, including a model card, the evaluation set and its refresh cadence, thresholds for human routing, and a named owner for each. Generative components get the same treatment, with prompt versions tracked like code.
AI system engineering
Four build patterns cover most of what finance teams ask us for.
Document AI for statements, claims, and KYC files
Extraction runs against layout-aware models with per-field confidence, and low-confidence fields go to a review queue rather than straight into the ledger. Validation rules catch the arithmetic a language model gets wrong. Corrections from reviewers feed the evaluation set, which is how accuracy on your document mix improves instead of drifting on someone else's benchmark.
Assistants grounded in your own content
Answers get assembled from retrieved passages with citations back to the source document and section, so an operations lead can check the claim without leaving the screen. Out-of-scope questions get refused instead of improvised. Retrieval quality is measured with a held-out question set drawn from support tickets your team already answered, which keeps hallucination risk under control as the content changes.
Decision support for underwriting and reconciliation
Models produce a recommendation with reason codes, and the underwriter keeps the pen. Thresholds split the volume into automated approvals, referrals, and declines that require a second look, and every path writes a record of inputs, version, and outcome. Reconciliation works the same way, matching candidates ranked by the system and cleared by an accountant.
Forecasting and anomaly detection
Treasury forecasts living in spreadsheets break the moment the analyst who built them goes on leave. We move the logic into versioned pipelines with backtests you can rerun, then layer anomaly detection over transaction and cash flows. Alerts carry the contributing signals, and the investigator decides, because a flagged pattern is a question rather than a verdict.
Running it in production
The work after launch decides whether the first result holds.
Monitoring and drift detection
Input distributions shift when a product launches or a channel mix changes, and a model that scored well in April can quietly degrade by August. Dashboards track feature drift, confidence distribution, override rates, and latency, with alerts routed to the team that owns the decision. Silent failure is the expensive kind, so thresholds are set before go-live.
Evaluation sets and version control
Prompts, retrieval settings, and model choices all get treated as deployable artifacts with a diff and an owner. Each change runs against a graded evaluation set before release, and results are stored so a regression can be traced to the commit that caused it. Rollback takes minutes. Nobody edits a production prompt in a console.
Human review and exception routing
Confidence thresholds decide what a person sees, and the queue is designed as a workplace with keyboard-first review, side-by-side source, and a reason field on every override. Those reasons become training data and governance evidence at once. Volume through the queue gets reported weekly, which tells you when to raise a threshold or retrain the model.
How we work
What separates an AI demo from a system that stays in production.
Presales engineering
We start by asking what decision the model supports and what happens when it is wrong. If nothing changes either way, we say the project is not worth doing.
Retrieval before fine-tuning
Company knowledge belongs in a governed, updatable store rather than baked into weights nobody can audit.
Schema-validated output
Extraction is validated against a contract, so the systems downstream receive data rather than prose.
Ship narrow, widen on evidence
One process, measured against how it runs today, before the scope grows.
Your infrastructure, your keys, your retention policy
We deploy into your cloud account or data center, with model access, logs, and document stores under controls your security team already audits.
What you get
A working system in your environment
Code, infrastructure definitions, and pipelines land in your repositories, running on the account you nominate. No hosted black box in the middle. Your engineers can read every part of it, and the handover session covers how to change it without calling us.
The evaluation harness
A graded question or document set, the scoring scripts, and a record of every run. This is what lets your team judge the next model release on your own data instead of a vendor slide. It ships with the system and stays yours.
Governance documentation
Model cards, data lineage, threshold rationale, drift monitoring plan, and the human review path written down in language a validator accepts. Compliance reviews go faster when the answers exist before the question. We write these during the build rather than in a closing scramble.
A team that stays reachable
The engineers who built it handle the first months of production, then train your people or stay on a support arrangement you set. Escalation paths and response expectations are agreed in writing at kickoff, so an incident at month four does not turn into a procurement exercise.
Where this works best
The wrong AI project is expensive twice: once to build, once to explain.
Talk to the team who would build itGood fit
A document-heavy or judgement-heavy process with a measurable cost today, and data you cannot casually send to a third party.
Not a fit
A general assistant with no defined decision behind it. We say so before the budget is committed.
Questions
Answers we give in most first conversations.
How long before we see something working?
A scoped pilot on one document type or one assistant use case reaches a demonstrable state in a matter of weeks, and production hardening follows. The variable is data access, which is why we ask for sample documents and a security conversation in the first week. Any timeline quoted before someone has seen your data is a guess.
Can this run without sending data to a third-party model provider?
Yes. Open-weight models hosted in your own environment cover document extraction, classification, and most retrieval work, with quality close enough for these tasks. Hosted frontier models still win on open-ended drafting, so some clients split traffic by sensitivity. We build the routing layer either way and keep the provider swappable behind one interface. That decision belongs to your security team, and we document the tradeoff for them.
How do you handle hallucination in a customer-facing assistant?
It starts with grounding. The assistant answers only from retrieved passages, cites them, and declines when retrieval returns nothing relevant. Confidence and coverage checks sit in front of the response, and anything below threshold routes to a person. Regression runs against a graded question set catch drift after content updates. This controls hallucination risk, though no design removes it, which is why the review path and the audit log both exist.
Who owns the models and the code?
You do. Work is done under your IP terms, committed to your repositories, and deployed with your own cloud credentials. Prompts, fine-tuning datasets, evaluation sets, and infrastructure definitions transfer with everything else. We keep no runtime dependency on Oxagile-hosted services, which matters at renewal time and matters more if you decide to bring the work in-house.
What do you need from our side?
A business owner who can decide what good looks like, someone with authority over the data, and access to the people doing the work today. Engineering time from your side helps but is not required at the start. The first two weeks are mostly document samples, system access, and sitting with the analysts whose job the model touches.
Part of AI in Finance
AI put to work in finance operations, grounded, governed, and running in production.
Get AI into production,
not just a demo.
Have a workflow begging for AI, whether documents, copilots or forecasting? Tell us the use case and we'll take it to production, governed for risk.