AIGP / Sectors / Financial Services
Sector 04 — Financial services

Agents that understand the banking context.

For banks, insurers, and financial institutions. AI agents automating operations, fraud detection built on graph neural networks, decision intelligence, document authenticity verification. Built on open-source models — no vendor lock-in, aligned with digital sovereignty requirements.

Discuss a project → See the work
What hurts the sector

Three areas where AI shifts the economics of operations.

I spent more than twenty years in financial services. These three themes dominate board-level conversations at banks and insurers in 2026.

  1. Operating cost grows faster than revenue

    Compliance, KYC, claims handling, document analysis — all of it consumes the hours of people who could be redirected to work that actually creates value. AI agents built on open-source models now match commercial solutions at significantly lower operating cost.

  2. Fraud that learns faster than rule engines

    Classical rule-based systems cannot keep up with patterns that shift month to month. Graph neural networks surface patterns invisible to rules — and scale to volumes no analyst will ever review.

  3. Authenticity verification in a world where anything can be faked

    Voice deepfakes in call centres, synthetic photos of car damage, forged identity documents. AI-generated content detection has stopped being a theoretical challenge — it has become an operational requirement.

Partner engineering realisations

Technologies ready for banking adoption.

Projects whose architecture and metrics transfer directly to the financial sector.

AI agents · Automation AI agents for task automation

Open-source agents matching commercial alternatives.

AI agents built on open-source models (Deepseek, Qwen, GLM, Kimi K2) matching or outperforming commercial models at significantly lower cost. Agent pipeline: task intake, context analysis, solution planning (chain-of-thought + MCTS), implementation, verification, output ready for review.

In a banking context — applied to code review automation, test generation, legacy codebase refactoring (banks sit on decades of COBOL), developer onboarding, technical documentation.

Open-source
No vendor lock-in
AST-level
Whole-project code understanding
MCTS
Multi-path solution planning
Decision intelligence Risk analysis and forecasting system

Probabilistic answers to decision questions in <30s.

Central service accepting queries via API, processing them through a multi-stage pipeline. Questions fan out in parallel to several LLMs; a news module automatically retrieves and scores relevance of articles; final aggregation combines results with reasoning. Output: probability 0.0–1.0 with justification.

In banking — applied to scoring market events, regulatory risk assessment, decision support for corporate credit.

80%
Accuracy on validation data
<30s
Response time
~20/min
Query throughput
AI-generated detection · Compliance Detection of AI-generated content

Image and video authenticity verification at institutional scale.

Pipeline processing 25M+ media items. Models: ViT-L, VideoMAE, V-JEPA. ONNX export for low-latency deployment. In financial services — verification of damage photos, detection of synthetic ID documents, authenticity analysis for KYC materials.

~97%
Image classification accuracy
~93%
Video accuracy
ONNX
Production format
Graph neural networks · R&D Max-clique problem and graphs

GNNs in service of network-pattern detection.

Hybrid architecture combining classical algorithms with deep learning. GNN recognises structures in graphs from under 100 to more than 500 vertices, with sub-second resolution time. Financial-sector applications: fraud detection (transaction-network patterns), AML (capital-linkage networks), concentration-risk analysis.

<1s
Resolution for 500+ vertex graphs
GNN
Graph neural networks
Hybrid
Classical algorithms + deep learning
Commercial models

The financial sector rarely uses grants. That's fine.

Banks and insurers fund AI from IT / operational budgets. Grants can be leverage in consortium or R&D projects — but they are not the rule.

Primary model

Enterprise delivery

Institution's own budget. Deployment tempo measured in quarters. Commercial structure based on fixed-price or T&M with success fee.

IT / OpEx budget 12–24 weeks
R&D

FENG SMART Path

For genuinely R&D-flavoured projects (new algorithms, domain-specific models). Bank as consortium partner alongside a technology firm and research unit.

up to PLN 140M 16.06–11.08.2026
Partnership

Retainer model

Ongoing AIGP strategic advisory + priority access to the technology-partner network. For banks building long-term AI capability.

Quarterly Multi-year
Research consortium

Bielik.ai & partners

For institutions interested in working on sovereign language models — research consortia with AGH Cyfronet and the Bielik / SpeakLeash team.

Consortium R&D projects
How we work in financial services

A sector we know from the inside.

Over two decades in banking — including C-level roles — means we understand the difference between what a board says it wants and what it will actually buy.

01

Conversation at decision level

If a project is going to make sense, it needs an owner at board or divisional-director level. First meeting with that person — not with IT.

02

Regulatory assessment

DORA, AI Act, GDPR, KNF recommendations. Before the tech moves, we check how it fits the regulatory framework.

03

Partner selection — with emphasis on data security

The partner has to meet data-processing requirements. On-premise, private cloud, open-source models — usually essential.

04

Pilot in an isolated environment → rollout decision

The financial sector does not deploy "live". Pilot, validation, decision, production — in the organisation's rhythm.

Insights — financial sector

Reading before the conversation.

Banking · AI agents · 12 min

AI agents in banking: what's production-ready in 2026

From KYC and compliance to development automation — what works, what doesn't yet.

Read →
Fraud · GNN · 10 min

Fraud detection beyond rules: the graph-neural-network approach

Why classical rule engines lose to network patterns — and how banks are starting to change this.

Read →
Regulation · 9 min

DORA and the AI Act: practical implications for AI deployment in banks

What shifts in 2026 and 2027 regarding AI project compliance — and how to design so regulation isn't the problem.

Read →
Financial sector · conversation

Bank, insurer, investor? We speak the language.

First meeting — 45 minutes, one topic, no slides. We leave with a decision on whether this is a project worth continuing.

Schedule a call →