Everyone Can Call the Same API

Model releases reshuffle the leaderboards every few months. The products built on top of them mostly don't change.

That gap is something we think about constantly. Finiti exists to put filings in front of regulators, which means the questions that keep us up aren't about which model we're calling. They're about where AI belongs in the workflow and where it doesn't, how a statement traces back to its source, and whether a lawyer will actually trust what comes back.

Roshan Nanu leads AI at Finiti. He studied neuroscience before moving into AI and has built AI platforms in healthcare and industrial settings, so he has watched this problem play out in two fields that are at least as unforgiving as ours. He also makes the calls on where AI sits in our product.

We asked him three questions.

As AI models become more capable and more widely available, what will actually set AI-native companies apart?

Every few months a new model release reshuffles the leaderboards, and every few months the products built on top of those models look roughly the same as they did before. That's a good sign the model isn't where the difference lives anymore. Everyone can call the same API.

I think the companies that pull ahead will be the ones that treat the model as a component and compete on three things around it:

Specialization. General-purpose tools are good at a little of everything. The products that win in a given field will know that field deeply: its documents, its rules, its edge cases, and what "good" looks like to the people doing the work. That knowledge takes years to build and doesn't come with a model upgrade.

Trust. This has two halves. The first is evaluation: if you can measure how your system performs on real work, your product stays stable as the underlying models change, and you can adopt a better model with confidence instead of hoping nothing broke. The second is a product that earns trust from its users, by behaving consistently and showing its work, so they're comfortable relying on it.

UI/UX. The best AI in the world doesn't matter if people have to leave their tools to use it. Products need to meet users where they already work, make reviewing the output quick and natural, and feel good to spend a day in. I've built AI platforms in healthcare and industrial settings, and in both the hardest part was never the model. It was getting the output in front of a clinician or an engineer in a form they'd actually trust and use. ChatGPT is the clearest example of this. It didn't go viral because it was a leap in technology; GPT-3 had been available and in use for more than two years before it. It took off because it was finally the right interface for that technology. The job for every AI-native company is finding the user experience that's unique, simple, and right for its specific domain and workflow.

None of these come in the box. A competitor can call the same API tomorrow; they can't call up years of domain knowledge, a tested evaluation practice, and a product their users have already built habits around.

For companies in regulated industries like finance, healthcare or legal, the bar is much higher. What does it take to build AI systems they can actually trust, and how should the industry think about accuracy and provenance?

Most of what a disclosure team does in a given week doesn't need a language model. It needs the right numbers pulled from the right places, checked against last quarter, and put in the right format. That's automation, not intelligence, and it's a useful place to start, because there are really two separate problems here: building an AI system that's trustworthy, and building one that a user can actually trust.

Building a trustworthy system

Remove AI wherever you can. Much of a regulated workflow is better served by deterministic automation than by a model. The more you constrain where AI acts, the easier it is to evaluate what it does, and the better the outputs get.

Give AI the right context and tools where it does act. When a model has exactly the information and tools a task requires, it's far less likely to go wrong. Most failures I've seen come from bad or missing context, not from the model itself.

Building a system users can trust

Make provenance come from the system, not the model. Provenance that relies on a language model citing its own sources correctly is weak and prone to failure. If a system drafts a section of a filing, the system itself should track every piece of context that went into it, so every statement can be traced back to the source or precedent that informed it, whether or not the model remembered to say so.

Make verification easy. This is the most important part, and it's a UX problem. If checking an AI's output takes even half the time it would have taken to research and write it from scratch, the tool isn't saving anyone much. Finding the source behind a statement should feel natural, not like an audit.

Thinking about accuracy in terms of risk

The industry should think about accuracy the way it already thinks about human error: as risk to be managed. There are two ways to earn trust. You can reduce risk by only using AI where the stakes are lower, or you can increase accuracy by adding checks and redundancies. Until accuracy is proven to be very high, there should always be a human in the loop, and the system should make their review easy.

That gives two bars to clear. The first is whether the system is more accurate than a person doing the same task. The second, and harder, is whether a reviewer becomes comfortable enough to start auto-approving certain tasks or spending less time on review. That second bar is where AI actually changes how the work gets done.

Looking at the AI boom over the last few years, what lasting impacts do you think will permanently change how people behave and interact with the world around them?

I studied neuroscience before working in AI, and one finding from memory research has stuck with me: people tend to stop remembering things they know they can look up. We saw it with search engines; we got better at remembering where to find a fact than the fact itself. AI is pushing that further, from offloading what we remember to offloading what we do.

The biggest shift I see is from doing to reviewing. More and more knowledge work starts with a draft someone else produced, and the professional's job becomes judging it: is this right, is anything missing, would I put my name on it? That's already true for code, and it's coming quickly for legal, financial, and compliance work.

That shift changes what makes software sticky. The applications people stay with will be the ones that do two things well:

Make review fast and natural. When the job is reviewing, the review experience is the product. Tools that make it easy to see what changed, where it came from, and whether it's right will earn daily habits. Tools that make review feel like a second job won't.

Arrive with domain knowledge already built in. Nobody wants to teach a tool their profession before it becomes useful. The products that already understand the field on day one, its language, its standards, what a good draft looks like, will be the ones reviewers trust first and keep using.

People aren't going to stop checking the work. They're going to spend more of their time doing it, and they'll gravitate toward the tools that respect that time.