Decision-support AI for banking
Decision-support AI for banking is the practice of using AI to score, rank, and recommend inside a regulated decisioning workflow with human-in-the-loop sign-off - inside the regulatory shape banking actually operates under. Banks already run inside the strictest model-risk regime in the world: SR 11-7 model risk management at US institutions, BCBS 239 on risk data aggregation, EBA loan origination guidelines, and DORA digital operational resilience. Every AI build inherits that audit machinery on day one. Every output Impetora ships in this category carries a citation back to the source it came from, so a reviewer can rebuild any decision in seconds.
Citation-grounded decision support, scoped to the regulatory shape banking actually operates under.
What does decision support in banking actually look like?
Decision-support AI scores, ranks, and recommends inside a regulated workflow without ever taking the final decision automatically. The architectural guarantee is that the human who signs the decision sees the reason codes, the evidence chain, and the model version that produced the recommendation, before they sign.
Banks already run inside the strictest model-risk regime in the world: SR 11-7 model risk management at US institutions, BCBS 239 on risk data aggregation, EBA loan origination guidelines, and DORA digital operational resilience. Every AI build inherits that audit machinery on day one.
The pipeline is the same shape across every Impetora decision support build: Case ingest -> Feature extraction -> Model scoring -> Evidence assembly -> Reason codes -> Human-in-the-loop sign-off -> Audit trail. Each stage is observable, each stage writes to the audit log, and each stage has a measurable failure mode the readiness sprint defines before any model is selected.
What regulations apply?
EU AI Act Annex III point 5(b) creditworthiness; SR 11-7 full model lifecycle; EBA Guidelines (EBA/GL/2020/06); BCBS 239; DORA; GDPR Article 22; FSB report on AI adoption in financial services. [1]
High-risk under Annex III point 5(b). SR 11-7 governance, EBA loan origination guidelines, BCBS 239 data lineage, DORA resilience, and GDPR Article 22 all converge. Full Annex IV technical documentation is the default.
Every system Impetora ships carries the AI register entry, the risk classification, and the underlying analysis with it. A regulator or an internal audit team sees the full chain on a single page.
What does TRACE require here?
Trust. EU data residency, EU AI Act risk classification documented, GDPR by default [8], sectoral regulator framing recorded inside the AI register.
Readiness. Banking workflows are sampled for at least 30 days before a model is selected. Baseline current handle time, current error rate, current escalation pattern. Document the workflow the AI sits inside.
Architecture. Versioned prompts, evaluation suites, shadow-mode rollout. Only what passes evaluation reaches production. ISO/IEC 42001-aligned governance scaffolding.
Citations. Every output - extracted field, drafted response, retrieved passage, decision recommendation - links back to the source it came from, the model version that produced it, and the timestamp. The audit trail rebuilds in seconds.
What can go wrong and how do we prevent it?
Each case lands with its raw features, the model scores it and produces reason codes, the evidence assembly step pulls the citations behind each contributing feature, the regulated person reviews the full package and signs off (or rejects), and the audit log captures the model version, prompt, retrieval source, reason codes, and signer at the moment the decision was taken.
The failure modes we engineer against on every banking build: hallucinated content surfaces (mitigated by grounded retrieval and a "no source, no answer" fallback), drift over time (mitigated by quarterly drift reports against the eval set), permission leakage (mitigated by ACL-aware retrieval), and silent regression after a model swap (mitigated by shadow-mode redeploys with eval delta sign-off).
What gets shipped in a Lighthouse build?
Phase one (weeks 1-2) is the readiness sprint: data sampling, baseline measurement, AI Act risk classification, scope sign-off. Phase two (weeks 3-4) is the build and shadow-mode rollout, where the system runs alongside the banking team with output logged but not actioned. Phase three (from week 5) extends to production, additional document categories or channels or knowledge domains, and the recurring drift and accuracy review that keeps the system honest.
Pilot engagements at this scope start at EUR 25,000 for a single, well-scoped category. Full production deployments typically land between EUR 60,000 and EUR 150,000 depending on integration complexity, evaluation-set breadth, and the regulatory documentation depth your team requires. Submit a project for a custom estimate.
How does this compare to off-the-shelf decision support tools?
Off-the-shelf platforms (UiPath, Salesforce Einstein, ServiceNow Now Assist, Glean, Microsoft Copilot for the banking variant) work well when your workflow is close to their reference customer. Where they break is when banking regulatory documentation has to be produced for the specific decision the system took, on the specific document or interaction it took it on, against the specific model version that was running at the time. The matrix combination of EU AI Act risk classification, sectoral regulator (EBA, BCBS, FSB, DORA), and your own internal control framework rarely fits a vendor template. Custom builds are how that fit is achieved.
What we don't build
We will not let the system take the final decision
The architecture is designed so the model cannot be the signer. The regulated person in your banking workflow signs every decision, and the audit log records who signed, on what evidence, and against which model version.
We will not ship a model we cannot explain
Reason codes for every recommendation are a hard requirement. If a model variant scores higher on the eval set but cannot produce reason codes the reviewer trusts, we ship the explainable variant.
We will not skip the shadow-mode rollout
The system runs alongside the human team for at least four weeks with output logged but not actioned. We measure the disagreement rate and the underlying reasons before any decision is automated.
Related reading
Banking AI
Decision-support AI
Document processing AI for banking
Customer support AI for banking
Decision support systems
Frequently asked questions
Is decision support for banking high-risk under the EU AI Act?
High-risk under Annex III point 5(b). SR 11-7 governance, EBA loan origination guidelines, BCBS 239 data lineage, DORA resilience, and GDPR Article 22 all converge. Full Annex IV technical documentation is the default.
Where is the data processed and stored?
By default, processing and storage runs in EU regions on infrastructure under EU jurisdiction. We support specific regional pinning when a regulator or contract requires it. Original documents and interaction logs land in immutable EU object storage with hashes recorded in the audit log. We do not train any model on your data unless you ask us to and the contract permits it.
How do you handle the regulator audit trail?
Every output the system produces - extracted field, drafted response, retrieved passage, decision recommendation - writes a structured event to a queryable, append-only audit log with the model version, prompt, retrieval source, confidence, and the human signer (where one exists) at the moment the action was taken. BCBS 239, SR 11-7, and the relevant sectoral guidance are accommodated by the same log shape. The trail rebuilds any decision in under 10 seconds.
Can it work with our existing systems?
Yes. The delivery layer sits in front of the system of record you already use - case management, claims platform, core banking, AML/KYC platform, ticketing, document repository, contract lifecycle - and writes back through documented APIs or queue-based bridges with idempotent writes. The audit log writes regardless of where the data lands.
What does this cost?
Pilot engagements at this scope start at EUR 25,000 for a single, well-scoped category. Full production deployments typically land between EUR 60,000 and EUR 150,000 depending on integration complexity, evaluation-set breadth, and the regulatory documentation depth your team requires. We quote against your specific scope before any code is written.
How long does a deployment take?
A first pilot reaches production-grade behaviour in 6 weeks: 1-2 weeks readiness, 1-2 weeks build, then a minimum 4-week shadow-mode period before any decision is automated. Subsequent decision categories add 2-3 weeks each.
Sources
- Regulation (EU) 2024/1689 - Artificial Intelligence Act, official text
- EU AI Act Annex III - high-risk AI systems list
- SR 11-7 - Federal Reserve Supervisory Guidance on Model Risk Management
- BCBS 239 - Principles for effective risk data aggregation
- EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06)
- DORA - Regulation (EU) 2022/2554 on digital operational resilience
- FSB report on AI adoption in financial services (2024)
- GDPR Article 22 - automated individual decision-making, including profiling
Book a discovery call
Submit a project for a custom estimate. We will quote against your specific banking decision support scope before any code is written.