“I’ll be honest” is a phrase that usually signals the opposite. So let me state upfront, I work in AI for a (mostly) FinTech vendor, and yes, we’re all talking about agentic AI. Vendors are selling it, top tier consultants are excitedly selling that they
can integrate it, and customers seem to be either cautiously curious or waiting for the hype to settle. My company has customers taking tentative steps with agents, while I was at a streaming analytics for financial services event where one CDAO was vehement
that his bank didn’t have the skills to adopt them. I’m frankly therefore making up my mind if I’m part of the early stage of a “game changing” zeitgeist or merely an echo chamber for shiny suits.
Earlier this year, I mentioned agentic AI in a 2026 predictions blog. Since then, the noise has only grown louder. The analysts tend to
love it. The leading analyst firm predicts that by 2028, 33% of enterprise software will include Agentic AI, with half of day-to-day decisions made autonomously.
Capgemini pegs the market at $47 billion by 2030. Big numbers. Big change. But also, a lot of fluff. One experienced analyst at a private dinner said don’t confuse your already confused audiences with agentic jargon – they’re not ready for it. That’s even
with his firm selling agentic AI research. Even the leading analyst firm cited earlier have rolled back:
40% of agentic AI projects will be scrapped by 2027. Folks like Stephen Klein vociferiously argue contrarian opinions against the Eduardo Ordax-like “LinkedInfluencers.”
Let’s cut through the noise, baseline the nomenclature at least if not real world implementations or consultant and vendor echo chambers. My opinion on which it is? No one really cares, but for those that do, I believe agentic AI is part of an emerging zeitgeist.
What Is Agentic AI—Really?
Agentic AI isn’t just a rebranded chatbot or a glorified workflow. It’s about systems that are proactive, not reactive. These agents don’t just respond—they plan, decide, and adapt. They can break down complex tasks, choose tools, recover from errors, and
adjust strategies based on feedback. They’re not just executing instructions—they’re reasoning through them.
That’s a far cry from early LLMs like ChatGPT, which simply answered questions. Today’s agents can retrieve information, remember context, and orchestrate multi-step processes. But they’re still not infallible. Even the most advanced frameworks make mistakes.
That’s why human oversight remains essential—at least for now.
Workflows vs. Agents: Know the Difference
Workflows are deterministic. They follow predefined paths and deliver consistent outcomes. They’re great for structured, repeatable tasks—think document processing or data extraction. They can even include LLMs, but they don’t adapt or learn.
Agents and agentic systems, on the other hand, are dynamic. They decide what to do, when to do it, and how. They’re ideal for messy, unpredictable problems, like assessing complex customer interactions or navigating multi-system processes. They’re expensive,
yes, but powerful where flexibility and autonomy are needed.
Systems of Agents: The Real Deal
A single agent is useful. A system of agents is transformative, at least on paper. Imagine an orchestrator agent coordinating a team: one chats with users, another generates visualizations, another monitors alerts. They share memory, follow rules, and adapt
in real time. It’s not human-level intelligence, but it’s a step toward intelligent systems design.
Think of it like managing 100 interns on a summer project. Each agent has a role, but the orchestrator keeps the big picture in view. It’s not just tech—it’s organizational design and mangement. I was at an
AI For The Rest of US MeetUp recently in Shoreditch. There, a speaker who is implementing this stuff (in security and defense more so than financial services) went beyond the interns on a summer project analogy. He phrased his “systems of agents” (not agentic
AI systems – he never used that phrase) as like managing and orchestrating teams of specialists.
Enter MCP Servers
To make all this work, agents need to talk to tools and the rest of us. That’s where MCP (Multi-Agent Communication Protocol) comes in. Developed by Anthropic, the organization that gave us Claude, MCP gives agents a standard way to connect with tools, services,
and data. It’s not perfect—security and complexity concerns have been raised—but as a standard, it’s a leap forward from hand-coded integrations, closer to a common denominator than a center of excellence. Other better standards that raise the bar may follow.
But at the very least, it’s a standard.
Should they catch on, you may see MCP servers everywhere: vendors hosting them, customers orchestrating them, and agents using them to collaborate not just within but across organizations.
When Should You Use Agentic AI?
Use agents when:
- The task is open-ended or unpredictable.
- You need flexibility and adaptability.
- Multiple tools must be orchestrated dynamically.
- Human oversight is still required, but you want to scale.
Avoid agents when:
- The task is simple, structured, and repeatable.
- Speed and cost efficiency are priorities.
- A traditional workflow or rule-based system will do.
Financial Services Implications & Final Thoughts
Agentic AI isn’t magic. It’s not AGI. But it’s a meaningful evolution in how we build intelligent systems. The key is knowing when to use it—and when not to. Workflows matter! So too do people and decision-makers and complexity.
As with any emerging tech, nuance matters more than noise. If it works, agents are a prime fit for key tasks.
- In Quant and capital markets, imagine time-series agents, backtesting agents, pricing agents, equity research agents, VaR calculation agents etc.
- For the middle office and compliance, agents will read reports (e.g.SAR or STR), alert, monitor, score, build graphs, assess counterparties, write reports, etc.
When bundled together and orchestrated as systems of agents, or Agentic AI systems if you prefer that phrase, they’ll augment, service and heavy lift fully fledged processes, like orchestrating equities or FX trading systems, derivatives risk management
frameworks, compliance and validation improvements, and software development.
But video didn’t quite kill the radio star. It brought a ton more entertainment opportunities. Just look at this weekend’s sparkling Glastonbury line-up and excitement! Nor did electronification destroy the financial services industry. Far from it, we can
trade anything, anywhere, anytime (mostly), not just the FTSE or NYSE. The internet hasn’t quite killed off daily newspapers, bringing us conspiracies and fake news as well as a gazillion and one perspectives.
With agentic AI, we’ll see.
