By Connor Heaton, Director of AI, SRM
At the start of 2025, DeepSeek, a Chinese artificial intelligence (AI) startup, sent shockwaves through the technological landscape with the release of its R1 model. More recently, Manus, another Chinese AI creator and component of Beijing Butterfly Effect Technology Ltd. Co, is grabbing headlines with its highly performant general AI agent, taking on extended tasks like analyzing e-commerce data, supplier sourcing, trip planning, and stock analysis.
According to benchmarks, DeepSeek’s latest models are on par with U.S. counterparts like Anthropic, OpenAI, and Google, but at a lower price point and are open source.
From a credit union and financial services standpoint, DeepSeek’s impact remains incremental rather than a significant game-changer. Most of the buzz surrounding DeepSeek focuses on the details of its training and cost rather than its capabilities, which don’t drastically differ from U.S. models.
Being an open-source model means that the efficiency innovations introduced by DeepSeek are already being integrated into the rest of the industry’s offerings. U.S. open-source models like Meta’s Llama (already one of the most popular among businesses as a basis for their models) are expected to match DeepSeek’s performance and cost in short order.
Some cybersecurity experts have expressed concerns over DeepSeek and, more broadly, the potential for open-source models to hide malicious intentions, such as enabling the exfiltration of data on which the model is fine-tuned or attempting to introduce backdoors.
Manus is still in limited beta and not available for comprehensive benchmarking, but it has similarities to DeepSeek. Its capabilities aren’t strictly new from a technical perspective, but it’s notable in how good it is for how little advance warning the mainstream market had.
Anthropic’s existing Computer Use and OpenAI’s Operator, as well as the Deep Research offerings released by most of the major labs, offer the same promise of agents that can interact with applications and websites over longer task horizons and are already finding a place in the business ecosystem.
The main difference with Manus is that it may be a step ahead of US companies in reliability for a broad range of tasks, at least in terms of what’s been publicly released (OpenAI has teased specialized agent products at a price of up to $20,000 a month). As a rule, the more reliably AI agents can complete lengthy cognitive tasks with due adjustment for work context, the more disruptive and valuable it will be for financial institutions and businesses in general.
DeepSeek and Manus are part of a larger pattern indicating that AI progress and adoption are rapidly accelerating. For credit unions to maintain their competitiveness in today’s market, this is a technology they should keep on their radars and build into their strategies.
The Growing Potential for Self-Hosted Models
The falling cost of running generative AI makes it more economically feasible for credit unions to leverage fully secure, self-hosted models. Credit unions could feasibly own their infrastructure, running their chosen open-source model locally.
Self-hosting is typically very expensive, especially for leading models, with cost estimates ranging from $100K to $125K for hardware. However, this could quickly change, given trending cost reductions of 1.5x to 4x year over year.
Most financial institutions (FIs) have been content to use enterprise offerings from frontier tech companies, which offer privacy contractually but not physically. Nevertheless, these institutions tend to be very conservative about their use cases and data. The ability to self-host provides an option for even the most risk-averse institutions to adopt AI.
Generative AI Offers Tremendous Value, and It’s Improving at a Blistering Pace
There is no denying the transformative impact generative AI is having on banking. As a tool, AI offers tremendous value and is quickly leading to several industry success stories.
For example, in 2023, Florida-based South State Bank launched Tate, a generative AI-powered knowledge management bot for employee use. The bot is based on OpenAI’s ChatGPT and built in Microsoft’s Azure cloud. Leveraging Tate, the bank cut search times from an average of seven minutes to under 30 seconds. Additionally, the Commonwealth Bank of Australia uses a generative AI system to scan more than 20 million transactions daily and flag suspicious activity. As a result, the bank claims fraud is down 30%, and customer wait times are down 40%.
Standing still isn’t an option for credit unions. Generative AI tools and assistants are now part of the business environment. AI is the new Excel – capabilities like sentiment analysis are now table stakes, and fintechs and payments companies are making aggressive plays for the retail banking market.
Picking the Right AI Investments is Critical but Complex
AI is both overhyped in some respects and legitimately transformative. Deciding where and how to engage with AI isn’t easy for credit union leaders. There is a very legitimate case for taking a “wait and see” approach for certain types of solutions and use cases for small institutions. Still, there are other areas where it’s clear credit unions should be moving immediately (e.g., strategy, training, development, marketing).
The pace of AI development means that the market enjoys rapidly improving capabilities effectively for free, without larger institutions having exclusive access to better fundamental tech. This also means that creating specialized solutions based on today’s models and capabilities carries the risk of swift obsolescence.
For example, Bloomberg invested early in training its own custom large language model (LLM) from scratch to help with finance tasks and write queries in its proprietary version of SQL. Less than a year later, a fully unspecialized commercial model, the original GPT-4, outperformed Bloomberg’s on nearly all finance tasks.
Deciding what AI capabilities to integrate into your workflows and how to handle complicated questions requires an understanding of some of the technical dynamics and trends.
Recognizing Internal and External Risks
The main internal risks of AI adoption for FIs arise from misuse of AI tools – misplaced trust in generative AI systems, both from the perspective of content accuracy and bias and the use of unsecured tools for sensitive work. Publicly available AI assistants aren’t safe for use with proprietary data.
Moreover, employees are using AI assistants to automate parts of their work and be more efficient, with or without corporate approval. There are benefits for FIs that have a strategy and policy to foster this adoption safely and with appropriate oversight – and serious risks for those that don’t.
External risks center around malicious actors gaining similar efficiency advantages, which leads to increased phishing, fraud, cyberattack volume, personalization and sophistication, which will impact FIs whether they adopt AI internally or not.
Preparing for Future Change
All businesses will reshape their processes and structures to make the best use of AI assistants, agents, and tools—playing to their strengths and mitigating their weaknesses. While there are some domains where FIs should wait and see, much can be done today to adopt valuable AI solutions and gain a competitive edge when the next big advancement in AI arrives.
Connor Heaton is the Director of Artificial Intelligence at SRM.