⋅ X min read
I spent the years before I started Ben thinking about physics, not benefits. One thing that stays with you from that world is a particular kind of scepticism about the word "transformation."
That's not a belief physicists like me hold, it's how nature works. Most change isn't gradual. A system sits in the same state for a long time, then crosses a critical point and tips into something structurally different.
Water at zero degrees isn't slowly becoming ice, it's water, then at the critical point, it's ice. Anyone who's watched a puddle freeze over in winter has seen it happen in real time.
I think the benefits industry is sitting right on top of one of those thresholds now, and almost nobody within it seems to have noticed.
"AI-native" has stopped meaning anything
To be clear, here's my definition. AI-native means the system doesn't just follow rules you wrote in advance, it reasons over what's actually happening and decides what to do next. Old software is deterministic: it does exactly what it was told, every time. AI-native software notices when something's changed and adjusts, without someone having to go back and rewrite the rules.
But nobody's using it that way. Everyone in this category, to some degree, says they're AI-native. And why wouldn't they? A decent strapline costs about 50 cents in tokens to write. A working prototype can be built for under $50. Of course everybody says it, the barrier to claiming it has basically disappeared. It's become the kind of phrase that's stopped meaning anything, the way "cloud-based" did a decade ago, the way "online" did before that.
And most of what companies mean when they say it is small: a chatbot sitting on top of the same platform that existed before, answering questions about a policy document a bit faster than a search bar used to. That's definitely not nothing. But it's also not the threshold I'm talking about.
A chatbot can be added to almost any platform, however old, however patched together, however many acquisitions and legacy modules it's built from.
That's exactly why it's the first thing everyone shipped. It's cheap to bolt on and easy to demo. But it doesn't touch the actual problem, which is that most benefits platforms got stitched together, module by module, over a decade or more, so a policy document, the eligibility rule it drives, the pricing logic and the payroll instruction all sit in different systems that were never designed to talk to each other.
Foundation models are table stakes now. So is feeding it a policy PDF. Any of us could do that ourselves in Claude or ChatGPT this afternoon. It'll read the document just fine. But it doesn't have a way of seeing how that document connects to everything depending on it, because nobody ever captured and structured the rest of the business in a way AI could actually read.
So the model does the only thing it can: it answers questions.
The threshold that actually matters
The actual threshold, the one that changes what's structurally possible, isn't about the model at all. It's about whether the data underneath was ever built to be something AI could act on, not just read. That's an architecture decision, and it's one that has to be made from the beginning.
Here's what it looks like once it's actually built. A reconciliation error gets caught before it reaches a payslip, not discovered afterwards in a month-end scramble. An employee's homepage updates the moment their circumstances change, because the platform already knew, nobody had to go and tell it.
There's a name for that distinction, and I think it's the one that matters most right now. AI that just chats back is one thing, that's most of what's out there today. AI that acts on your behalf, works through next steps, and comes back to you only when it needs a decision, is another thing entirely.
That's an Agent. And I think 2026 is when that shift stops being a demo and starts being how the work actually gets done, and now context, not the model, is what's holding most back.
You cannot retrofit it in eighteen months because someone asked for an AI roadmap. I'd go even further: most of the platforms currently telling reward leaders they're "AI-native" made the opposite decision, when a bolted-on chatbot looked like the fastest way to say yes to the same question.
It's worth answering the question you’re probably asking in your head right now too: large companies have all the data, shouldn't that mean they win this race?
Not so fast. Having the data is one thing. Turning it into something structured enough for AI to act on, and changing how thousands of people work day to day, is another. It's a consequence of when these companies were built and what they got right for that era. But it means something uncomfortable is coming for this industry.
In a few years, "AI-native" is going to stop being a marketing claim and start being a question with a real, checkable answer: can this platform's AI act on a genuinely connected model of how the business's benefits work, or can it only describe fragments of it back to you. A lot of companies currently answering "yes" to that question are going to find themselves, very publicly, unable to answer it a second time.
What to actually ask your vendor
For anyone reading this who runs a benefits programme, here's the practical version of all this.
The question worth asking any vendor, including us, isn't whether they have AI. Everyone will say yes, and for the next year or two, most demos will look roughly the same. The question that actually separates the two kinds of company is whether their AI is reasoning over one connected picture of how your benefits work, or reading pieces of it and hoping the gaps don't matter.
Ask them to show you, live, what happens when a policy changes. Watch whether the system rebuilds the logic and shows its working, or whether someone goes and does it by hand afterwards while the chatbot keeps answering questions about the old version.
Here's an easier version anyone can run in five minutes.
Put the same plain-English question to your platform's employee-facing search: how do I claim on my dental cover? A connected system answers with that employee's actual cover. A bolted-on one recites the policy document back at you.
I built Ben on the bet that the bolted-on kind of software was always going to lose, eventually, even if it took the market a few years to start asking the right question. I think we're close to that inflection point now. And I’d rather say that plainly, in public, before it's obvious to everyone, than wait and let the market discover it on its own.
Water doesn't announce it's about to become ice. Neither will this.
