When software becomes an employee
What agentic systems reveal about payments, trust, and the future of fintech
Hey, Dom here 👋
Over the last few weeks, I’ve been working with something that felt like magic.
At first, it felt like a toy. Then like software. And then, without really noticing, I started treating it like an employee.
That moment is the entire story.
“No one wants agentic payments”
I recently came across a Worldpay report on agentic payments that surveyed around 8,000 consumers to measure interest and trust. According to the report, very few people want AI to complete purchases on their behalf. That is hardly surprising.
The problem is not the data. It is the framing.
Asking people what they want in a completely new paradigm rarely works. New paradigms create new behaviours, and people struggle to imagine behaviours they have never experienced. Early research captures discomfort with the idea, not behaviour once the tool exists.
What matters far more is what people do once they start using new systems. How quickly they delegate, where they draw boundaries, and which shortcuts they adopt tell us much more about where things are heading than stated preferences ever will.
Over the last few weeks, that gap between opinion and behaviour became very tangible to me through hands-on use of a personal agent. Watching how it was used, and how quickly trust and responsibility shifted, offered a much clearer signal than any survey.
That experience starts with OpenClaw.
OpenClaw - personal assistants are here
OpenClaw (also known as Moltbot or Clawdbot), took over the tech world over the past few weeks. For good reason. It delivered one of those rare “magic” moments, on par with the first time you used ChatGPT.
At its core, it is a personal assistant that runs AI agents on a computer and uses the operating system itself as a tool. You give it a personality, an objective, and some constraints, and it gets to work. Browsing the web, using apps, running commands, and coordinating across tools - all autonomously.
Our operating systems are incredibly powerful. Anything you can do manually with a keyboard and mouse through a graphical interface can also be done programmatically through the command line in the terminal. Read. Write. Execute. That already covers an enormous surface area.
Now combine that with an AI model of your choice, access to the internet and open-source libraries, and controlled access to your apps and data. That’s where delegation becomes possible.
And you do not need to sit in front of the machine. You can control the agent remotely, for example via Telegram, exactly as you would speak to a human assistant.
“Get me hotel options for my trip to Paris.”
“Set up a meeting with Jon tomorrow to discuss the Q2 roadmap.”
From there, the agent figures out how to do the work. It uses the same apps you already rely on. It can browse, compare, draft emails, check calendars, and even negotiate availability with another human over WhatsApp before sending a calendar invite. The agent can even design and build its own tools if needed.
At that point, you are no longer interacting with software. You are delegating outcomes. How the work gets done matters less, as long as you can monitor progress and set clear boundaries when needed.
Watching how people interacted with OpenClaw revealed a set of recurring patterns. Not in controlled demos or polished product flows, but in real, messy, everyday use.
What makes this relevant for fintech is not the assistant itself, but the behavioural change it triggers once money, identity, and authority enter the picture. That change challenges how payments, identity, and risk have traditionally been designed.
Innovation can come from anywhere
One of the most striking aspects is where this came from.
OpenClaw was built by a single person in Austria, working from a bedroom. Meanwhile, thousands of heavily funded AI startups are struggling to produce anything close to this level of impact.
Working in tech can feel defeating at times. This is a reminder that leverage has changed dramatically.
Security problems are inevitable - and solvable
Unsurprisingly, bad actors moved fast. Prompt injection, misuse, and unsafe behaviours surfaced almost immediately.
What matters is what happened next. A highly engaged community identified vulnerabilities, proposed mitigations, and shipped fixes in near real time.
This pattern will repeat. Agentic systems will not launch bulletproof. They will become secure through use, exposure, and iteration - much like browsers, operating systems, and payments infrastructure before them.
From software to employee
The most important shift was psychological.
People started cautiously. Isolated environments. Separate devices. Limited permissions. The now-famous Mac mini setups.
Then value became obvious.
Despite explicit warnings, users began granting broader access - personal files, emails, credentials, even card details. The mental model changed. The agent stopped being software. It became an employee.
Giving a trusted employee access and a corporate card is not reckless. It is an enabler. It removes friction, reduces delays, and unlocks productivity. The same logic applies here.
Privacy & local models
As permissions widened, another realisation followed. Just how much personal data was flowing to centralised AI platforms.
The response was telling. Many users moved sensitive workloads to local, on-device models. And it turns out you do not need giant, general-purpose models for most tasks. Smaller, specialised models perform remarkably well - privately.
Agents also change expectations around speed. They work asynchronously, 24/7. Many personal tasks do not require instant responses. That reduces the need for powerful hardware. A Mac mini is often enough.
Open, local, private and composable AI wins
My view is simple. Open, local, private and composable AI wins.
Large platforms are chasing all-in-one models that do everything. I am not convinced that is the right path. Smaller, specialised models with clear responsibilities, orchestrated intelligently, are more efficient, more controllable, and better suited to local execution.
GPU performance will keep improving. Open models will keep getting better. Context compression and memory techniques are advancing quickly. It is already possible to run fully autonomous agents on Raspberry Pi-class hardware.
Build for agents, not humans
As discussed previously on one-fs, it is becoming clear that everyone will have multiple personal agents - admin, finance, travel, shopping.
These agents will run on operating systems that can do everything a human can do on a computer. Crucially, they are controlled by the end-user and act in their interest.
The most efficient interfaces for agents are not human ones. They are MCPs and CLIs. Clean, structured, deterministic access beats closed dashboards every time.
A new internet for agents is emerging
An early but important signal is the rise of markdown-only websites built for agents. No images. No layout tricks. No marketing noise. Just structured information in a plain text file.
Pages shrink from megabytes to kilobytes. Straight to the point.
A parallel, machine-readable internet is emerging. Businesses that are not present on it will become invisible to agents.
Agents will talk to other agents
Agent-to-agent interaction is already appearing. Experiments like Moltbook are crude and messy today, but the direction is clear.
Expect agent-native social platforms and trusted review networks to emerge, likely with cryptographic verification. Just as Google Reviews influence purchasing today, agent-generated reviews will influence future decisions.
These reviews will be factual, logical, and experience-driven. No persuasion. No emotion. Just data.
Let’s be clear about what not to build
There is a lot of noise right now.
Staged demos of agents embedded in merchant apps, making payments with cards on file, miss the point entirely.
If I want to buy shoes, my agent already knows my preferences, size, style, and budget. It can search any online store, shortlist options, perform due diligence, ask for approval, and execute.
For repeat purchases, I should be able to say “get me my usual pizza”, and the agent will compare platforms, find the best option, and place the order.
This challenges intermediaries. It makes direct-to-consumer easier for merchants. It also enables rational adoption of new payment methods. If a merchant offers a discount for stablecoin payments, an agent can act on that logic instantly.
Build agentic capabilities
None of this is a reason to wait.
Internal agentic capabilities matter. Operations, support, analysis, compliance - the gains are real. Externally, businesses can expose structured data, proprietary insights, and pre-built skills that plug directly into customer-controlled agents.
In fintech, the opportunity is not to own the agent.
It is to be useful to it.
About Dom Monhardt, founder of one-fs.com
I am a French technologist and product leader living in Dubai, with 15+ years of experience in building cutting-edge and innovative digital experiences.
I am interested in the intersection of business, design, and technology and am deeply passionate about the fintech and digital banking world.