Summary
Developers are already showing financial firms what the future of AI looks like. By combining multi-agent workflows, deterministic execution layers, and domain-specific context, they are building more reliable AI systems. Asset managers that adopt these practices can reduce hallucination risk, improve automation, and create scalable AI workflows.
In its global survey, Grant Thornton found that 71% of asset management industry executives are planning to adopt GenAI within three years. Less than 10% are currently using agentic AI, while 18% plan to adopt over the next three years.i However, to hear software developers tell it, 100% of the asset management industry will be using agentic AI within three years.
The developer community has been working with AI tools longer than the rest of us. Arcesium’s developer and data engineering team, roughly 250 people, has a higher-resolution view of where agentic AI in financial services is headed than most. We design AI tools for hedge funds, asset managers, institutional investors, and banks, and their experience is the clearest signal we have that human-led clarity, model determinism, and domain-specific context determine AI effectiveness.
Developers (devs) have been riding the AI wave for years, with occasional wipeouts, working closely with these multi-engine workflows. They experienced the pain of hallucinations early and created strategies to mitigate them. Instead of just asking for code to solve a problem, they have code to solve the problem along with test code to prove the problem is solved. Instead of taking a first pass through generated code, they have an agent that critiques the output to provide first-level feedback before a human has to review. You can set these agents to work with each other to produce something that goes through cycles until it hits a certain quality threshold. The human dev does not have to babysit as many processes. Instead, they act more as orchestrators of agentic workflows.
Devs often try to “one-shot” a task. Instead of engaging with a large language model (LLM) in a series of prompts, they craft a single prompt that preemptively covers all the details in the ticket. The goal: the AI runs until all tests pass. It resembles explaining the task to another developer, and it may end up a page of typed notes to be fed into the AI model.
The key to the “one-shot” task is the human doing the hard work up front of knowing what a good solution looks like, giving instructions for tests, evaluation criteria, feedback loops, and permissions to modify outputs. The challenge is extracting the context in your head and embedding it into a machine process. The human feeds that comprehensive page of notes into the probabilistic LLM. When they hit enter, the agents go through all the cycles. An agent runs a test, sees a problem in the output, and makes a fix, on its own, then repeats. Then it runs another test, makes a fix, and so on. The end product is the deterministic execution layer.
In finance, we need our agents to be predictable and repeatable, guided by governance policies that prevent model drift and contain operational risk. Financial tasks require the same answer today as they did yesterday, but AI agents are probabilistic by nature.
“A key aspect of intelligence is to not only make predictions, but reason about the uncertainty in these predictions, and to consider this uncertainty when making decisions... Understanding the uncertainty in our model of the environment is crucial for planning. Moreover, modeling uncertainty in our model of the environment can be extremely useful in deciding where to explore. Learning a model can therefore help to dramatically reduce the sample complexity over model-free techniques. Often times this is crucial when developing agents for real-world use as in such settings, exploration is costly and potentially dangerous.” – Probabilistic Artificial Intelligenceii
AI systems are predictive models rather than super-intelligent beings; therefore, users must write down clear goals and standards.
LLMs hallucinate, and in finance, error tolerance is zero. LLMs are great at generating an answer that is plausible, looks right, and could be true. But they aren’t good at adding or counting. For example, investor reports are not things one leaves to probabilistic LLMs. The report has to be right every single time. A human and a probabilistic model do the workflow design and creation, but deterministic systems execute the tasks.
Devs use LLMs for reasoning and creating procedures. Then they allow the agent to generate code, run it, inspect results, modify as needed, and repeat. Calling the LLM is often the most expensive path, especially across multiple prompts. Therefore, it is advantageous to use deterministic software for what it’s good at: calculations, business rules, validation, matching records, and executing workflows.
People at firms will be automating their workflows, so agents can help them be more rigorous and be able to check more inputs than they could get done in a day. Enabling innate creativity like this is why Arcesium has built AI as a platform, not a feature layer. Firms are moving toward automating complex workflows, such as investigating P&L discrepancies or rebooked corporate actions, by embedding industry-specific context and skills into agents. In the near term, different firms’ results and use cases are going to be uneven and diverse. Some will race ahead and maybe their AI costs are 10 times higher than those of others, yet their productivity gains will be substantial.
As William Gibson has said, “The future is already here — it’s just not very evenly distributed.”iii This is certainly the case today in finance. The pace of AI innovation compresses that distribution. The future is arriving faster than Gibson imagined. Soon, firms will act as conductors, coordinating autonomous agents, LLMs, and specialized orchestrators.
Devs tend to onboard new colleagues with detailed written instructions rather than face-to-face interactions, and they approach AI agents the same way. As AI adoption grows, we’ll prompt agents for outcomes rather than steps. Instead of telling agents to add two and two, we'll ask them to highlight sector allocation changes or explain underperformance versus benchmark.
Success will belong to those who can effectively teach agents to exercise the judgment built from years of industry experience. That’s the standard we hold for ourselves. Firms operationalizing AI the way developers do will lead. Those treating it like tourists, observing rather than operating, will fall behind. The dividing line is understanding which parts of the work need a probabilistic model and which need a deterministic one.
Matt Katz
As Arcesium's Field CTO, Matt leads Arcesium's Forward Deployed Software Engineering and Client Success teams. His work to empower clients and simplify technical challenges stems from a 25-year career in financial technology working with clients and software. Outside of work, he enjoys books, bikes, and boards.
Sources:
i Grant Thornton, December 2, 2025. https://www.grantthornton.com/insights/articles/asset-management/2025/ai-is-transforming-asset-management
ii Andreas Krause, Jonas Hübotter, February 7, 2025. https://arxiv.org/pdf/2502.05244
iii National Public Radio: NPR: Talk of the Nation, The Science in Science Fiction. October, 2018. https://www.npr.org/2018/10/22/1067220/the-science-in-science-fiction
No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.
