What AI Maturity Level Have Financial Institutions Actually Achieved?

Over the past two years, large language models have moved quickly from experimental technology to a standard item on financial institutions' roadmaps, covering use cases such as knowledge Q&A, intelligent customer service, collection bots, research assistants, model development and various agents. However, the number of projects does not reflect capability depth; merely integrating a model does not mean the institution truly masters it.

The core of the article is a six‑layer framework (L1‑L6) that helps diagnose and plan AI capability. It is not a regulatory rating nor a mandatory upgrade path, but a way to measure how deeply a model is embedded in business and how well the institution understands the associated technologies.

Three key judgments are proposed:

AI capability should be judged by the layer of internalisation, not by the count of launched projects.

Higher layers do not automatically bring higher business value; technical depth and economic value must be evaluated separately.

The ultimate goal is not to turn everything into an agent, but to let AI reliably perform valuable tasks within auditable, controllable boundaries.

Layer descriptions :

L1 – Daily Support Tools / General Assistant : Uses a generic model directly via prompts and context for writing, summarising, retrieval, analysis and coding. No model parameters are changed. Success is measured by reduced task time, higher adoption rate, controlled rework/error rates, and stable integration into high‑frequency workflows.

L2 – Skill / Professional Assistant : Captures institutional expertise (methodology, steps, checklists, output schema) into repeatable AI skills. A Skill includes task description, professional steps, required context, tool interfaces, examples, output schema and quality checks. It moves knowledge from individual users to the organisation.

L3 – Complex Data Understanding / Semantic Interpreter : Turns hard‑to‑use financial data (credit reports, transaction logs, contracts, multimodal materials) into three asset types – Fact (verifiable fields), Profile (semantic portraits) and Embedding (high‑dimensional representations). Evaluation focuses on traceability to source evidence, consistency, and downstream model gains.

L4 – Professional Human‑Machine Interaction : AI conducts multi‑turn, goal‑driven interactions (e.g., customer service, collection) while handling state, business rules, tool calls and compliance constraints. The layer may start with prompts, knowledge bases and workflows, and only adopt SFT, preference optimisation or RL when baseline performance stalls and high‑quality data are available.

L5 – Business Model Paradigm Shift : Introduces a "pre‑training → post‑training" pipeline. Institutions first pre‑train on their own domain data to learn universal business representations, then fine‑tune for specific tasks such as risk, marketing or operations. Success requires cross‑task transferability, reduced sample cost and stable performance under drift.

L6 – Agent Decision System : AI becomes a governed decision orchestrator that can dynamically plan, select tools, gather evidence, compare alternatives and produce or execute actions while respecting risk, compliance and audit requirements. Human oversight remains the final authority, especially for high‑risk decisions.

The framework stresses that depth of model training and system autonomy are two orthogonal axes; progress does not follow a single linear path, and an institution may be mature at L2 while still exploring L5.

Value assessment distinguishes two types:

Efficiency‑type value (L1, L2, L4): AI reduces the unit cost of a task. The key metric is Cost per Accepted Task – the total cost (remaining labour, inference, platform, rework, governance) must be lower than the original human‑only cost.

Business‑gain value (L3, L5, L6): AI adds incremental business outcomes. Evaluation looks at risk‑adjusted incremental revenue, improved conversion, reduced loss, or superior model performance (AUC/KS) compared with strong baselines.

Both value streams require rigorous measurement: baseline comparison, A/B testing, phased rollout and continuous monitoring to ensure that the AI contribution is stable, positive and within risk limits.

In conclusion, the true watershed for AI in finance is not the number of models or tokens used, but the extent to which institutions have internalised AI across the six layers – organising knowledge, understanding data, training models, delivering services and ultimately shaping decisions.