By João Carvalho, CEO and Co-founder, Ubbin Labs
There’s a claim going around that I keep hearing in different forms. It goes something like this: the era of the software platform is ending. Soon we won’t log into tools, click through menus, or maintain dozens of separate systems. We’ll just talk to a large language model, and it will do everything, with all of it integrated behind a single conversation.
I find this idea seductive, and I think it’s about half right. The half that’s right is worth taking seriously. The half that’s wrong is where most of the real engineering lives.
Why the claim has teeth
The honest starting point is that the people making this argument are reacting to something true. Most software was built for a world where humans decide and systems execute. The interface assumes a person sitting there, reading a screen, choosing the next click. APIs were designed for that same world – predictable calls, made by code that a human ultimately set in motion.
Agents break that assumption. When the thing on the other side of the API is reasoning, asking follow-up questions, and deciding what to do next, the old contract starts to creak. We wrote about this on the GFoundry side recently, looking at how AI agents are starting to coordinate across systems through open protocols – the Model Context Protocol being the one most people have now heard of. The short version: the seams between systems are being redesigned for machines that think, not for humans who click.
So if you squint, you can see how someone concludes that the platform – the place you log into, the screens, the buttons – is on its way out.
Where I think it’s wrong
Here’s the part the clean version of the story skips: the work a platform does doesn’t vanish when you put a model in front of it. It moves.
Someone still has to hold the data. Someone has to know who is allowed to see what, and prove it later. Someone has to keep the record straight when three systems disagree about the same employee. Someone has to anonymise personal data before it’s processed, and stand behind that when a regulator asks. A language model, on its own, does none of this. It’s very good at reasoning over information. It is not a system of record, it is not a permission boundary, and it is not an audit trail.
So the question isn’t really “platform or LLM.” The more useful question is: if the human stops being the one clicking through the interface, what does the platform become?
Our current answer is that it stops being the place where people go and starts being the thing agents act through. The screens don’t all disappear – some of them matter more than ever, because a human still needs to supervise, approve, and intervene. But the centre of gravity shifts from the interface to the infrastructure underneath it: the data, the rules, the integrations, the actions that can actually be taken safely.
What we decided to build, and why
I’ll be concrete, because abstractions are cheap. On the GFoundry side we’ve been building toward this with something we’re calling TOAR – Talent Orchestration, Automation and Response. I won’t pitch it here; that’s not what this site is for. But the design decisions are worth explaining, because they’re the decisions any team will face if it believes the “platform becomes infrastructure” thesis.
We made three calls.
First, integrate, don’t replace. The temptation is to declare the old systems dead and rebuild the world. That’s a fantasy when your client runs a Core HR, an HCM and an ERP that took years to settle. So we connect to the sources an organisation already has rather than asking it to throw them away. The data stays where it lives. We orchestrate across it.
Second, open at the seam. We expose the platform’s capabilities to the language models through MCP, rather than betting that one vendor’s model wins. If the interface to intelligence is going to be a protocol, then the protocol is the thing to get right, not the model of the month.
Third, privacy as a precondition, not a feature. Personal data is anonymised before processing, under GDPR. If agents are going to act on people data, that constraint can’t be bolted on at the end. It has to be part of the foundation, or the whole thing is indefensible.
None of these are exotic. They’re the unglamorous decisions that determine whether an “AI-native” idea survives contact with a real client who can’t afford downtime. We’ve been building people systems for about ten years now, with a small team, for organisations that have no patience for our learning curve. That experience makes me sceptical of any story where the hard parts – data, permissions, accountability – quietly evaporate.
What I don’t know
I want to be honest about the edges of this, because I don’t think anyone has them figured out yet.
I don’t know how much of the traditional interface survives. It’s possible that five years from now most routine HR work happens through conversation and the dashboards we build today look quaint. I think the supervision layer – the moments where a human has to decide – stays visual and stays important. But that’s a hypothesis, not a forecast.
What I’m more confident about is the shape of the bet. The platform isn’t dying. It’s changing who it talks to. For a decade it talked to people through screens. Now it’s learning to talk to agents through protocols, while still answering to people for the things that matter. The teams that treat that as an integration problem rather than a replacement story are, I suspect, the ones who’ll still be standing when the noise settles.
We’re building as if that’s true. Ask me again in a couple of years whether we were right!