⚠️ This is an engineering write-up of an internal proof of concept, not financial advice and not a product announcement. Trading CFDs is high-risk and most retail accounts lose money. Read the disclaimer before drawing conclusions about your own setup.
For most of the internet era, customers have accessed digital products through a website (or mobile-friendly website) and, more recently, a mobile app.
A third has now emerged: AI agent.
Instead of manually opening an app, finding the right page, and working through a series of screens, we tell an agent what we want to achieve. The agent then does those steps with the service on our behalf.
The 45-second version of the agent journey: one prompt, from sign-up through to a closed trade.
One common example is how this changes the way people search for information.
And I believe that in the last 3 months it has grown even more significantly.
Traditional search still handles some activity, but the behaviour is now different. Instead of typing a few keywords in a search engine and opening ten tabs, people simply give the AI assistant the context, ask follow-up questions and have it compare the sources for them, and the AI will do the searching for them, for example, when booking flights or planning holiday trips.
Bill Gates made a broader version of this argument in 2023 when he wrote that “agents will be the next platform”. His view was that people would eventually stop choosing a separate app for every task and tell their device what they wanted in everyday language. I do not entirely agree with his view, but I think it will always stay half and half since humans are always stubborn to change; there are still lots of people who prefer using pen and paper to fill out a form instead of doing it digitally.
Regardless, I wanted to see what that might look like with a product that has more complexity and consequences than researching a holiday or booking restaurant reservations. So, for the PoC, what’s the best way to test it out? Our own product, of course.
Full disclosure: I work at Fintrix Markets, which gave me a convenient test subject. I wanted to see whether an AI agent could complete the full trading journey just like a human would:
I have recorded the whole thing, and it’s safe to say that the agent made it to the end, but it needed far more direction than I expected. That roughness taught me more than a polished demo would have, but the main conclusion is that we’re building our product for the purpose of human interaction and not for autonomous agents.
The demo was janky
The agent could complete registration, but it needed plenty of hand-holding on what to do next.
It completed the journey, but using it felt like supervising a junior employee on their first day. I had to point it in the right direction more often than I wanted.
Some of this came from the agent itself. Today’s agents still make mistakes, lose context and need instructions that seem obvious to a human. We even provided a detailed onboarding process in the llms.txt on our website for AI agents to consume, but the agent still could not understand the difference between calling an API and filling in a form, or why it needed to confirm customer details again.
Human checkpoints need to feel deliberate
KYC is another good example that an agent still cannot do on its own.
An agent can follow the onboarding steps, but it still cannot become me. The regulated entity still needs to verify the real customer. I still need to provide identification, take a photo and complete a liveness check. There are also declarations and agreements where an agent again prompted me to confirm multiple times. Our vendor Sumsub has a plan to implement KYA (Know Your Agent) that will bind our KYC to our agent, so we just need to do this verification once (or periodically when our details change) and the agent can use it for different purposes.
A required human checkpoint should feel like part of the journey, not like the agent has become stuck.
But currently it’s limited because the infrastructure supporting agents is still limited.
Giving the agent a wallet
I also wanted the agent to fund the account. For the PoC, I experimented with x402 and USDC. I created a crypto wallet the agent could use through Coinbase’s developer tooling via CDP. The wallet worked, but that immediately raised questions about authority.
Who owns the wallet? How much money can the agent move? Can it send funds anywhere or only to approved businesses? Which payments need my approval? What happens when the agent misunderstands an instruction?
Moving USDC from one address to another is relatively straightforward. Giving software useful financial authority without giving it unlimited authority is much harder. For my test, I simply gave it a temporary wallet with a very small amount of money in it, similar to a debit or prepaid card.
Trading was the straightforward part
Once the account was available and the agent had the right access, placing a trade was fairly simple. Our trading platform vendor, MT5, now supports MCP, so we simply provided instructions about how to set up the MCP in the terminal. But then again, this is another handover to a human, albeit a very simple one; still another annoying step, but you only need to do it once.
The agent could open a position, check it and close it. Trading already has many machine-friendly interactions. The surrounding customer journey created more friction. Registration, identity, agreements, funding, authentication and approvals all carry assumptions about how a customer interacts with a business.
Designing for people and their agents
Digital products may increasingly serve the customer through two different interfaces. Humans will need websites and mobile apps. They need visibility, control and a place to complete actions that require personal judgement or consent. However, agents need access to the product’s capabilities, and this can be done via API, MCP, etc.
An agent does not care where the deposit button sits or whether the funding journey uses three screens instead of four (I remember spending hours with stakeholders and our UX team to get this right). It needs to know what actions are available, what information they require, which permissions apply, when the customer must step in, what happened after an action and what it can do next.
That is a different kind of customer experience (or should I say agent experience, but wait, should we now treat the agent as a customer?). For the agent, the best interface is simply no screen at all.
Website. Mobile. Agent
Same product, same capabilities, three ways in. Only two of them need a screen.
Websites and mobile apps will remain important for now, but personal AI agents could become another major way customers access the same businesses. We can already see this with the number of agencies and services out there that offer AEO on top of traditional SEO (although fundamentally both are the same; you do one and you get the other).
Today, I might open an app and book a car to the airport. In the future, I may tell my agent to arrange the trip and make sure I arrive by 7am; the agent then checks the traffic, notifies my self-driving car to be ready at 5:55am and pings me at 6am to get into the car and be driven to the airport.
Currently, I might open a banking app to pay a bill, but in the future, I can ask my agent to handle all my bills, tell me which subscriptions I can cancel because I never use them, alert me to suspicious transactions, track how my budget is going day by day, etc.
Rather than visiting a brokerage website, completing forms, depositing money and manually placing a trade, I might say:
“Compare the top five forex brokers in my region with a minimum leverage of 500:1. Pick the best one according to my trading style, open an account, deposit $10,000 and trade using the usual strategy that we’ve been using. Give me a daily report every morning at 8am. Escalate any issues to me.”
My Fintrix Markets experiment taught me a lot of lessons. At its current stage it still needed too much prompting and too many handovers that felt accidental. Some human involvement is required and should remain. The limitation comes from the infrastructure around agents (KYC, wallets, permissions, etc.). But I am sure that as other supporting technologies mature (which, at this pace, won’t take long), the friction will be removed.
So I am proud to share that Fintrix Markets is almost agent ready. We have already got a few things on our backlog that will improve the agent’s journey further.
I’m sharing the rough demo because it shows where the experience breaks today. It gives a concrete list of what an agent-ready product still needs to solve. And maybe others can help point me in the right direction.
Disclaimer
This post describes an internal proof of concept run by the Fintrix Markets tech team. It is engineering content, not financial advice, and nothing here is a recommendation to trade any instrument or to use any particular tool. The agent-driven journey shown in the video is experimental and is not a supported way to open or operate an account. Trading CFDs is high-risk and most retail accounts lose money.