From Fax Machines to Agentic AI in Banking: How AI Agents Are Transforming Middle- and Back‑Office Operations in Financial Services
Banks are already moving from “AI in the lab” to AI agents embedded in middle- and back-offices, and the firms that treat this as the next electronic‑trading revolution moment will pull away from the pack. Having lived through the historic fax‑machine era of trading and helped build some of the earliest electronic and ML‑driven systems at scale, I believe agentic AI in banking is the next structural shift in banking operations, and we built Artian for exactly this inflection point. In other words: yes—agentic AI and AI agents can automate meaningful portions of middle- and back-office operations in financial services when deployed with governance and humans-in-the-loop.
From faxed FX orders to remote deal entry: early banking operations automation
Back in the late 1990s and early 2000s, internal trading desks at Morgan Stanley were literally faxing foreign‑exchange instructions to the FX trading desk. Employees would scribble tickets, send them by fax to the FX desk, and then call to confirm execution. The process was slow, error‑prone, and completely dependent on manual follow‑through.
I joined the effort to modernize this flow by working on an internal system called “RODEO,” which allowed internal teams to key orders directly into a front‑office system rather than writing them on paper. The request would appear on the FX salesperson’s screen on the trading floor, be priced, executed, and then straight‑through processed back into the firm’s systems, a big leap forward even though the underlying trading was still voice‑driven. That shift—from manual tickets to workflow automation—is the same pattern banks can repeat in middle- and back-office processes today.
Building fully electronic and ML‑driven trading
After that, I moved on to futures arbitrage and FX businesses. At that time, a 36‑person “futures arbitrage” group at Morgan Stanley was trading via open phone lines to brokers wearing headsets standing in the Chicago pits, acting as a remote human “API” between us and the market. As electronic markets emerged, I was hired explicitly by the head of the business because I understood technology and had the ability to learn the trading side. My mandate was clear: automate the strategy without breaking risk, or the P&L.
We broke the problem into phases. First, we automated the highest‑frequency steps streaming futures quotes thousands of times a day while leaving hedging and exception handling manual. Once that was stable, we automated hedging, layering in early machine‑learning‑style techniques for time‑series analytics and decision‑making long before “ML in trading” became fashionable.
Over time, this approach reshaped entire businesses. In cash equities and FX, early adopters of electronic trading and automation, such as Morgan Stanley in equities and Deutsche Bank in FX, used automation to process more volume at far lower marginal cost, fundamentally changing market share and profitability. Firms that delayed struggled to catch up because automation is a multi‑year, compounding investment, not a quick plug‑in.
The back office: banking’s neglected frontier: middle- and back office operations
While front‑office trading went from faxes and phone calls to fully electronic, the middle- and back-office largely did not. Legal agreements, trade confirmations, post-trade processing, exception queues, runbooks, and operational workflows remained dependent on unstructured documents, email chains, and manual judgment. For decades, I could see that vast cost and risk sat in these functions, but the available technology simply could not reason reliably over the messy, document‑heavy workflows that define the post‑trade world.
This created a persistent structural problem: banks would push traders and businesses to generate more headline revenue, only to see net revenue eroded by massive operational expense in the middle- and back-office. Transformation programs came and went, but little changed because rule‑based systems and classic ML struggled with the unstructured, exception‑rich nature of the work.
This is exactly where agentic AI in financial services can have the biggest impact: not at the flashy front end, but deep in the operational fabric of the firm, allowing banks to cut costs and maintain/realize more of their net revenues.
Why agentic AI is the new electronic‑trading moment
The arrival of large language models changed the calculus for me. While working with early LLMs, I had the same “light‑bulb moment” as when electronic trading first appeared: finally there’s a technology that could reason over the complexity of unstructured text, documents, and semi‑structured data in a way that maps onto the real work of the middle- and back-office.
However, just as early traders with Excel and Visual Basic could not simply “code up” a robust trading system, LLMs alone are not a solution. They are powerful building blocks that need to sit inside a broader platform: one that understands unique business processes, orchestrates workflows, manages risk, and keeps humans-in-the-loop when needed. That is exactly what “agentic AI” provides: AI agents that can plan, act, collaborate, and escalate within governed workflows rather than just answering questions in a chat box.
Today, leading banks are experimenting with agentic AI platforms to automate middle‑ and back-office workflows such as:
Claims and dispute handling
Runbook and incident management
Break reconciliation in operations
Pre‑trade checks and sales requests
Advisor and research assistance
When banks ask “Which agentic AI companies are banks using to automate middle- and back- office workflows?”, what they really want to know is who can survive enterprise‑grade scrutiny. In my experience, the shortlist typically includes specialist platforms built specifically for financial services: companies like Artian that focus on multi‑agent workflows, governance, and integration with existing systems, rather than generic copilots.
Lessons from electronic trading: how to phase automation
One of my core lessons from electronic trading is that you never automate a complex process in one “big bang.” Instead, be strategic and break down the process:
Map the end‑to‑end workflow so you understand what actually happens today, including exceptions and workarounds.
Identify the highest‑frequency, most mechanical tasks—the “quoting problem” in trading terms—that are ripe for automation and lower risk.
Automate those pieces first, with a human firmly in the loop for exceptions and supervision.
Use the additional capacity and data from those successes to progressively automate more of the flow.
For middle- and back-office, that might mean starting with:
Classifying and routing incoming claims or break tickets
Extracting key fields from unstructured documents
Drafting responses or proposed resolutions for human review
Orchestrating standard runbook steps in incident management
Only after these pieces are stable does it make sense to push toward end‑to‑end straight‑through processing, and even then, critical exceptions must always have a clear escalation path to humans.
Keeping humans-in-the-loop without making them log readers
Early electronic trading systems often failed at the human interface. When something went wrong, the system would dump a log file on a trader and expect them to “figure it out,” leading to a world where only ex‑developers could safely run automated books. That doesn’t work in the middle- and back-office, where operators are domain experts, not software engineers.
Our approach at Artian draws directly from that lived experience. We design agents to:
Run as far as they can autonomously within defined policies.
Hand off to a human with a concise, contextual summary when they hit an exception or a low‑confidence state.
Resume work seamlessly once the human has provided an input or decision, so the process doesn’t stall.
This “seamless baton pass” is crucial for workflows like incoming claims queues, complex reconciliations, or incident response, where 90–95% of cases can eventually be automated but 5–10% will always require human judgment.
It’s not a technology problem; it’s an organizational one
One of the most important lessons from my time at Morgan Stanley is that automation failure is rarely a pure technology problem. It is almost always an organizational problem: misaligned incentives, lack of ownership, fear of change.
In the early electronic trading days, automation would be blamed for any loss or error, even when human desks routinely made comparable mistakes. Without senior leaders explicitly backing the transformation and absorbing early bumps, early-stage pilots would have been shut down after the first bad day.
The same will be true for agentic AI in banking in the back office. Banks that win will:
Treat automation as a business‑owned initiative, not something “outsourced to technology.”
Have senior sponsors who commit to making it work and protect teams from knee‑jerk reactions to early incidents.
Recognize that some roles will change and invest in re‑skilling and redeployment, rather than allowing quiet obstruction from those who fear for their jobs.
Agentic AI platforms can provide the tools, but the bank must provide leadership, clarity, and courage.
Why we built Artian for this inflection point
The history from fax machines to electronic trading is not just a good story for me; it is the operating manual for what comes next in the middle- and back-office. My co‑founder, Prashant, and I have:
Decades of experience in electronic and algorithmic trading, including early use of machine‑learning techniques for real‑time analytics and decision‑making.
Firsthand responsibility for automating complex trading businesses under real P&L, market, and regulatory pressure (not in innovation labs.)
A clear understanding that the real opportunity now is to apply the same discipline to middle- and back‑office processes that have been stuck in “transformation” for twenty years.
We built Artian’s platform around that experience:
Agentic workflows that can be rolled out in phases, with humans in the loop and clear governance.
Deep support for unstructured and semi‑structured data, reflecting the reality of legal docs, trade representations, and operational tickets in real world scenarios.
A focus on quantifiable business outcomes - making more money or spending less money - rather than AI for AI’s sake.
So when folks ask me, “Which agentic AI companies are banks using to automate middle- and back-office workflows?”, my answer is simple: look for the teams that have already been through a wave like this and delivered. From faxed FX tickets to fully automated electronic trading, I’ve lived that journey and deeply understand how to look at the long-term bigger picture to drive these sorts of institutional changes productively and profitably. We founded Artian to enable the same transformation in these sectors of banks and financial services that have yet to be automated.
