The Broker’s Five-Year Engagement Behind a Trading Platform 

Key takeaways 

• AI has changed how fast trading software gets built. It hasn’t changed how long that software has to keep running, or who stays answerable for it while it does.
• The savings everyone expects from AI are real, but smaller and slower to arrive than the pitch suggests. Most of the cost of a trading platform lives in the years after launch. 
• A convincing demo is the easy part. Software that holds up under live load, survives the market’s edge cases, and stays compliant through years of regulatory change takes experienced people to build and to run. 
• More than half of Devexperts’ development pipeline already runs on AI. The responsibility, though, hasn’t moved an inch. It stays with the engineers and account managers who answer to clients. 
• The events of 2026 have been making the case better than any vendor could. Firms that let build speed outrun human oversight spent the year cleaning up after it, and regulators have started writing the lesson into their rulebooks.

AI is transforming the economics of building software, and trading technology is no exception. Code that once took a team weeks now takes an afternoon, and for anyone pricing out a new trading platform, the arithmetic looks irresistible. 

But there’s a piece the arithmetic conveniently leaves out: time. In the following article, we look at what actually happens after a trading platform goes live, why the engagement between a broker and its technology partner runs for years rather than months, and why AI, for all its speed, hasn’t taken a single name off the list of people responsible for the result. 

Our guide here is Dmitry Zaitsev, Head of Account Management at Devexperts. He has spent close to two decades watching clients arrive, sign, go live, and stay, and he has a clear view of what AI changes in that journey and what it doesn’t touch at all. 

The number everyone arrives with 

Ask Dmitry what’s changed in client conversations since AI became a household word, and he’ll tell you about a number. Clients walk in expecting AI to take 90% of their development budget. 

Can you blame them? The cost of producing a first working version of almost anything has genuinely collapsed. A founder describes a feature, and working code appears within minutes. If that’s the part of the project you can see, ninety percent sounds about right. 

Here’s the catch: building the first version was never where most of the money went. 

Most of the cost of a trading platform lives in the years after launch. It lives in the reliability work that keeps the system up when volumes spike, in the changes that keep it compliant every time a regulator moves the goalposts, in the integrations that have to keep pace as your business grows, and in the unglamorous maintenance that keeps a platform competitive long after the launch announcement has scrolled out of view. AI has barely touched those costs. In careless hands, it has pushed them up, and the tools themselves are getting more expensive by the quarter. 

Dmitry has a vivid way of describing what you get when you push the automation slider all the way and remove the people. Something that runs, yes, but something closer to a film-set prop than a working system: convincing from the right angle, hollow under any real weight. It demos beautifully. Then a live market leans on it. 

Go live once, stay reliable indefinitely 

A brokerage or a prop firm doesn’t buy a trading platform the way it buys office furniture. It takes on a system that will run its business for years, and a technology partner who will run alongside it the whole way. 

“The systems we build normally stay in production for at least three to five years. It’s a long-term operational engagement, and it keeps developing the whole time. This is not a one-and-done project.” — Dmitry Zaitsev, Head of Account Management, Devexperts.

Think about everything that happens to a trading platform across five years. It takes on new asset classes, stretches toward trading hours that creep ever closer to around-the-clock coverage, addresses each regulatory change, and scales to volumes its original design never anticipated. Your business changes, your clients change, the market changes, and the platform has to change with them all. 

None of it can be handed to a model and left alone. 

In the past, this was simply understood. A vendor relationship in trading technology was a tenancy on a very long lease, and everyone priced it that way. Today, the speed of AI-assisted building has created the illusion that the relationship can be as short as the build. The build got shorter. The lease didn’t. 

So where do the humans fit? 

Let’s be clear about one thing: none of this is a case against AI. More than half of Devexperts’ development pipeline already runs on it, and engineers who use it ship faster. AI compresses the work, and that compression is worth real money to clients. 

What AI doesn’t compress is the consequences of getting the work wrong. 

A confident wrong answer about how an order routes, or how margin nets under stress, costs the model nothing. It costs the client money, and it costs the vendor a phone call at an hour nobody enjoys. That’s why a person stays in the loop on anything that carries weight. Someone has to be able to stand behind what the system does, and a model can’t. 

We can’t afford to lose the human touch. The people building these solutions carry a responsibility to the clients who depend on them, and the human still has to be in every interaction with AI, to make sure whatever comes out of it is shippable. — Dmitry Zaitsev, Head of Account Management, Devexperts. 

Regulators, as it happens, have arrived at the same place. FINRA’s 2026 oversight report added a dedicated section on generative AI and reminded firms that the existing rules on supervision and recordkeeping apply to AI-assisted work just as they apply to everything else. In the UK, the FCA has signaled that guidance on audit trails and human oversight is on the way this year. The direction is the same on both sides of the Atlantic: when software touches client money, a human stays answerable for what it does. 

What 2026 has been teaching the industry 

If the case for human oversight ever felt abstract, this year has been busy making it concrete. 

Between December 2025 and March 2026, Amazon worked through at least four severe production incidents after pushing AI-assisted development across its engineering teams. The worst of them, an outage in early March, ran for hours and, by published estimates, cost millions of orders. And the cause wasn’t the model. The cause was a release that went out without the documentation and the sign-off that a human process would have demanded. 

The research backs up what the headlines suggest. A 2026 study presented at the ICSE software engineering conference, drawing on more than 500 practitioner accounts, found that AI-assisted work accumulates technical debt at roughly 3 times the rate of conventional development, with testing being the step most often skipped. Google’s DORA research recorded a measurable fall in delivery stability as AI adoption climbed. And close to half of AI-generated code, by repeated findings, ships with security flaws of the ordinary, catchable kind. 

Underneath all of it sits a quieter problem that deserves more attention than it gets. Building software by prompt is barely eighteen months old as a mainstream practice. That means nobody, anywhere, has yet operated an AI-built system across the five-to-ten-year horizon that a trading platform routinely lives through. The maintainability everyone is counting on has never been tested over the distance. 

Dmitry watched the same year unfold from the vendor’s side, and the lesson he drew is hard to argue with. The firms that got burned weren’t small or careless. Several were among the most trusted names in their markets. They let the speed of the build run ahead of the people meant to vouch for it, and their systems held right up to the moment they were asked to carry the load. 

The relationship outlasts the build 

Speed is worth having, and Devexperts builds faster with AI than without it. Clients feel the benefit in shorter timelines and lower first-build costs, and nobody is handing that back. 

What the speed doesn’t shorten is the commitment. A go-live is the opening of a relationship measured in years, one in which engineers keep the platform sound and account managers keep their word to the client, season after season, release after release. 

For Dmitry, that’s the part AI doesn’t reach. A model can draft the code. It can’t sit across from a client three years in and answer for how the system has held up. People do that, knowing their own name is attached to the result. 

Listen to the full conversation 

Alexander Mamalev, Head of Sales at Devexperts, sat down with Dmitry Zaitsev, Head of Account Management, to discuss how AI changes the way trading technology is built and run, and what it leaves untouched. They get into the misconceptions clients arrive with, and where a human stays non-negotiable across a delivery engagement that runs for years.

Planning a platform that has to last? Talk to us.

And don’t forget to watch the third episode of Devexperts’ podcast on our YouTube channel.