Five AI Trends Finance Teams Need to Understand Right Now

Five AI Trends Shaping Banking and Finance in 2025

This post covers five AI trends that are already shaping real workflows in treasury, risk, and regulatory teams: agentic AI, retrieval augmented generation, AI governance and model risk, small language models, and multimodal AI. Understanding how they connect is more useful than tracking any one of them in isolation.

Why These Five

The filter is straightforward. Does it affect how finance teams handle data, decisions, controls, or regulatory obligations? If yes, it is in. If it is interesting to a technologist but largely irrelevant to a treasury or risk practitioner today, it is out.

Trend 1: Agentic AI, From Answering to Acting

Most people's mental model of AI is still a chatbot: you ask, it answers, you decide what to do next. Agentic AI breaks that model. An agent does not just answer a question. It takes a sequence of actions, calls external tools or systems, checks its own output, and completes a task with minimal human intervention at each step.

For treasury and payments, this matters immediately. Consider a cash management workflow where a system monitors intraday liquidity positions, identifies a projected shortfall by end of day, identifies the appropriate funding source according to rules set in advance, drafts the instruction, and routes it for approval. Each of those steps is something an AI agent can handle today, not in five years.

The risk side is equally real. An agent that can act is an agent that can make a mistake at speed. Controls that work well for human operators need to be rethought when the actor operates in milliseconds and across multiple systems in parallel.

If you want to go deeper on the specific mechanics and risk questions around AI agents in payments and treasury, the posts on agentic payments and what happens when AI can actually move money on this blog are worth reading alongside this one.

Trend 2: Retrieval Augmented Generation

Large language models have a limitation that is widely known: they are trained on data up to a cutoff point, and they do not know anything about your firm's internal policies, your specific loan book, or the version of a regulatory return you filed last quarter. Retrieval Augmented Generation, usually called RAG, is the practical answer to that problem.

RAG works by connecting a language model to a search layer over your own documents and data. When a user asks a question, the system retrieves the relevant passages or records first, then passes them to the model as context. The model answers using that retrieved material rather than relying solely on what it learned during training.

For a compliance team, this means you can build an AI assistant that answers questions about your actual policies, your internal procedures, and your current regulatory obligations, without retraining or fine tuning a model, and without sending your documents to a third party. A credit analyst could query the system across previous credit committee papers to identify how similar exposures were treated historically.

The important caveat is that RAG is only as good as the underlying data it retrieves. If your policy library is inconsistent, poorly structured, or out of date, the model will faithfully return answers based on those flaws. The data quality problem does not disappear; it moves upstream.

Trend 3: AI Governance and Model Risk Management

Regulators in the UK and across major jurisdictions are moving from general AI principles to specific expectations. The PRA and FCA have both made clear that AI used in material risk decisions, customer interactions, and regulatory reporting is likely to fall within existing model risk and operational risk frameworks, even where specific binding rules remain in development across different firm types.

Finance professionals do not need to become AI engineers to contribute meaningfully here. They do need to understand a few core concepts.

Model risk in the AI context asks the same questions it always has: What is the model doing? How was it validated? What are the failure modes? Who is responsible? The answers are harder with large language models than with traditional statistical models, because the reasoning inside them is less transparent.

Governance means being able to document what data trained or informs the model, what controls exist on its outputs, how drift or degradation is monitored, and what the escalation path is when something goes wrong.

For treasury and risk teams, the practical implication is that AI tools adopted informally at desk level may already require model risk treatment under your firm's internal frameworks. Awareness of that boundary is a meaningful contribution to the conversation, even if you are not the one setting the policy.

Trend 4: Small Language Models

The dominant story in AI over the past two years has been about scale: bigger models, more parameters, more compute. A quieter but equally important story is running in the opposite direction. Small language models, sometimes called SLMs, are compact enough to run within your own infrastructure, with the smaller variants capable of running on a single server, without sending data to an external cloud service.

For regulated firms, this changes the calculation significantly. Sending client data, trading positions, or internal credit files to a public cloud AI endpoint raises immediate data governance and confidentiality questions. A model that runs entirely within your own infrastructure does not have that problem.

The trade off is capability. A small model will not match a frontier model on general reasoning or tasks involving multiple steps. But for a specific, well defined task such as classifying transaction narratives, extracting structured data from a standard document type, or running a sentiment pass on earnings call transcripts, a small model tuned for the task can perform well enough, sometimes very well, within your own perimeter.

For treasury teams looking at AI for cash flow forecasting commentary, or risk teams wanting to parse regulatory documents automatically, small models are a realistic option to evaluate today rather than a future aspiration.

Trend 5: Multimodal AI

Until recently, AI tools in finance were mostly text in, text out. Multimodal AI processes multiple types of input together: text, tables, images, and documents in their original layout. This is more significant for finance than it might sound.

Consider a credit file. It contains narrative text, financial tables, scanned pages from company accounts, covenant schedules formatted in ways that vary by counterparty, and sometimes images of physical assets or property. A model that handles only plain text struggles with all of that. A multimodal model can read the table as a table, extract the numbers in context, and connect them to the narrative commentary in the same document.

Regulatory reporting workflows are another direct application. Annual reports, pillar 3 disclosures, and Basel template submissions contain mixed formats. A multimodal system can be pointed at source documents in their native form rather than requiring a preprocessing step to extract text first.

The applications in contract review, trade documentation, and KYC file assessment follow the same logic. The bottleneck in many of these workflows is not analysis but extraction: getting the relevant information out of unstructured or partially structured documents reliably enough to use. Multimodal AI addresses that bottleneck more directly than previous approaches.

What Finance Teams Should Do With This

These five trends are not independent. A useful way to see them together: agentic AI is the operational layer that acts, RAG is the knowledge layer that grounds it in your data, multimodal capability is the input layer that handles the documents your work actually involves, small models are the deployment option when your data cannot leave the building, and governance is the control layer that makes all of it defensible to your risk function and your regulator.

A finance team that understands all five is better placed to evaluate vendor claims, challenge technology proposals, and contribute to internal AI governance discussions than one that has only seen the agentic AI headlines or only understands the governance angle.

The practical starting point is not a large transformation programme. It is building enough fluency to ask the right questions: What data does this tool use? Where does it run? What happens when it is wrong? Who approved it? Those questions apply to every AI tool already in your workflow, not just the ones being proposed for next year.

The AI landscape will keep moving. These five trends will still be relevant in twelve months, even if the specific models and tools have changed. Understanding the underlying mechanics is more durable than tracking individual product releases.