What DeepSeek really means for business

DeepSeek roiled tech stocks, opened a new era in China-U.S. competition and upended assumptions about AI models.

What does this so-called “Sputnik moment” mean for how businesses will use AI? It gives them more options and more flexibility at lower costs. In short, win-win-win.

As companies build out use cases, whether the vendors they sign up are open- or closed-source – Meta or now DeepSeek versus OpenAI or Anthropic, respectively – isn’t as important as whether they have flexibility to develop and iterate their applications. In terms of AI models, businesses are agnostics. They appreciate having options and don’t want to be locked into any single provider. Let the model arms race continue – there will be another breakthrough any day now (in fact, OpenAI claimed two with the release of its Operator and Deep Research products and Google just released Gemini 2.0 to everyone). Ultimately open-source models will drive their cost down to virtually nothing.

The advanced LLM and reasoning models do consume more energy and chips (or compute) than ever before, so the costs per query are going up, even as the costs of the model itself goes down. But that will simply encourage companies to make sure they customize models for select purposes, using smaller models for simple tasks (domain specific) and leaning on smarter, pricier models for the hard ones. As if when you lose your keys, you’re not searching your whole house but know to look only in the living room.

The next phase in AI development will see businesses constantly tweaking and iterating their own models, always pursuing both flexibility and control. It will take a lot of innovation to customize open-source models — or build model-agnostic applications as banks decide which underlying models they want to use. That’s already happening, as our Use Case Corner shows this week.  

The DeepSeek story gave open-source models a huge boost (you’re welcome, Mark Zuckerberg). In banking, open-source use cases are few and far between, but these three – from Goldman Sachs, Capital One, and AllyBank – show the expertise (and the infrastructure) that customizing open-source models requires…and hint at the benefits those models can offer as banks move more use cases closer to sensitive data.

#1 Goldman’s Go-Fers

Use Case: AI Assistant
Vendor: Meta, OpenAI, Google
Bank: Goldman Sachs

Why it’s interesting: Rolled out to 10,000 employees, the GS AI assistant can translate code between languages and summarize or proofread emails – all part of what CIO Marco Argenti calls phase one in creating an agent that will be able to “do things like a Goldman employee.” Depending on the task, the tool orchestrates between ChatGPT, Gemini and Llama, a balancing act between the customization and security capabilities of open source and the higher performance levels that come with flagship off-the-shelf models from vendors.

Potential ROI → Efficiency gains
Reported ROI →  N/A

#2 Agentic Car Sales

Use Case: Chat Concierge
Vendor: Custom Meta model
Bank: Capital One

Why it’s interesting: Capital One is the first bank we track to launch an agentic use case built on an open-source model, which allows the bank to control the increased inference costs incurred by querying AI agents while also hosting a deeply customized model. Chat Concierge helps customers buy cars, offering trade-in estimation and appointment scheduling. It builds on a strength for the bank, which already has more than 90 patents dedicated to car buying, including an ML car recommendation system or a vehicle depreciation analysis tool.

Potential ROI → Customer satisfaction
Reported ROI → N/A

#3 AI Anonymizer

Use Case: PII Masking Module
Vendor: LangChain
Bank: AllyBank

Why it’s interesting: Ally Bank used open-source code from the LangChain community to build elements of Ally.ai, the bank’s enterprise AI platform. In the open-source spirit, Ally Bank later contributed to LangChain, sharing an open-source solution that can mask someone’s personal identifiable information (PII) before sending data to AI models.

Potential ROI → Risk reduction
Reported ROI → N/A

Have feedback on or ideas about use cases? Let us know at [email protected].

"“[DeepSeek] came up with new ideas and built them on top of other people's work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source.”"

- Yann LeCun, Meta’s Chief AI Scientist on LinkedIn, Jan. 24

DeepSeek may have been the latest flare-up in the Sino-American AI Cold War, but in the banking space, our data shows China pulling ahead in the IP race since at least last year. China’s four largest banks filed a staggering 2,700 AI-specific patents in the most recent filing year (a 26% increase over already record levels).

Why should you care? Those four banks alone are outpacing the number of AI-specific patent filings across all 50 banks tracked in the Evident AI Index (by a 2x margin). Furthermore, the Industrial and Commercial Bank of China is pulling ahead of patent leaders Capital One and Bank of America in recent filing activity. Though Chinese patents are cited less, the sheer volume of this patent activity underscores the pace of innovation that enabled DeepSeek to crash the AI party.

UBS used its earnings report to tout its “AI-first” mindset, estimating that employees would collectively total 17.5 million prompts across the firm’s Gen AI tools in 2025.

• Decoded: Even saying you’re AI-first takes a big firmwide commitment, and UBS is following the global AI leaders by taking steps to back it up and put the other (already-lagging) European banks on notice.

Foteini Agrafioti, RBC’s SVP for data and AI, made a point of saying the bank’s partnership with Cohere will allow the bank to deploy Gen AI use cases using its own infrastructure and its own data.

• Decoded: Developing AI applications close to their most sensitive (and valuable) data is forcing some banks to swing the pendulum back towards on-premises infrastructure, where they have greater control and security. What's old is new again!

CommBank inked a new five-year deal with AWS, rolling out a new CommBiz Gen AI messenger (that took just six weeks to develop) to “tens of thousands” of business customers to mark the occasion.

• Decoded: Cloud migration (and cloud-native development) still helps banks move quickly when deploying AI, and partnerships like this won’t go out of style.

Amazon AI and machine learning vet Daniel Marcu moved to Goldman Sachs to become the bank’s Global Head of Artificial Intelligence Engineering and Science. Rahul Sharma stays on as Goldman’s Head of AI Platform Engineering and Bing Xiang stays as Head of AI Research, both working under Marcu.

Lloyds brought Magdalena Lis on as its new Head of Responsible AI. She’s one of more than 30 on Lloyds’ 200-person AI team that holds a Ph.D., and she’ll be tasked with expanding the AI Centre of Excellence, the bank’s AI chief Rohit Dhawan said on LinkedIn.

AI training could become increasingly important now that the first AI literacy provisions of the European Union’s AI Act are in effect.

• Pop quiz: Under the law, any person on staff using AI (not just those developing it) must have sufficient technical knowledge about what AI does, how it’ll be used and who it’ll affect.
• Report card: Evident data suggests European banks are at least partially ready. Fifteen of the 20 EU banks in the index offer at least some kind of AI training. Only 8 of those say they have a centralized, company-wide program to train their staff on AI or data – which is what we’d expect them to need to meet the new regulatory requirements.

Agentic AI startup Auquan is aiming to change what it means to be an analyst, and UBS (among other banks) is buying in.

What OpenAI’s newly-released “Deep Research” agent does for general reports, Auquan’s system is doing specifically for finance. It cleans messy data (like presentations and PDFs) and proposes a list of relevant questions to address – and then dispatches a swarm of agents to scour internal documents and the wider internet to answer them in a report.

In an interview, CEO Chandini Jain told Evident data scientist Alex Inch that clients are clamoring for solutions that address the inconsistency issues of LLMs, and that while AI agents can’t yet handle the workload of a graduate-level analyst, they soon will.

This Q&A has been edited for clarity and length. Read the full interview here.

INCH: What are agents enabling you to do that a normal chatbot can’t?

JAIN: For our tasks, off-the-shelf LLMs today don’t work. Let's say you have ChatGPT Enterprise. If you give it a few PDFs and ask simple questions you can get good answers, but once you’re dealing with hundreds of thousands of documents it fails. To get good results, you need to process all of the data and filter it down to useful, relevant information. Then, you need something that understands your task and can break it down for an LLM to operate on. That’s an agent orchestrator: it automatically figures out what needs to be done, and then spins up AI agents to do it. This orchestrator needs to be domain-specific. It needs to understand your workflow and your firm.

GenAI models are getting cheaper over time, but the top models are still expensive. How do you manage costs?

Cost is definitely a concern, but I think it is possible to keep costs in control. If you use the best model for everything you would theoretically get the best performance, but that’s using a big hammer for a small nail. We keep costs low by using the most cost-optimized model for each task while preserving performance. If you can do it with an internal, hosted Llama, use that.

Do you have any thoughts on the effects of GenAI on work? Is Auquan aiming to eliminate the role of analyst?

It is eliminating a lot of the work that analysts do today, but I don't think it is eliminating the analyst role in a bank. Twenty years ago being an analyst in a bank meant that if you gave a presentation you had to print it out, cut it to shape and bind it together. No one does that today, but that doesn't mean you don't need analysts. People will continue to focus on the creative bits, but you can't escape from technology and choose not to become AI-aware. Eventually, just as you manage analysts, you will have to learn to manage AI agents.