Writer's May Habib: "Generative AI challenges the concept of done"

Over the last year or so, as businesses grew weary of magical thinking around generative AI and more interested in practical applications, their attention shifted away from the foundation model builders like OpenAI and Anthropic and toward development tool vendors that can help them build, deploy, and maintain production applications. Writer has raised $326 million on the premise that it can do both.

Writer likes to think of itself as a "full stack" generative AI company, according to May Habib, co-founder and CEO. It has built five different large-language models — with more on the way — and also developed a series of APIs and software tools that work with those models to help companies build apps, which has proven trickier than expected.

"Agentic AI is like the straw that broke the enterprise's back when it comes to generative AI, and people are understanding they can't be doing all of this themselves. They need support," Habib said.

Writer started off as a tool for marketing teams to generate content, but quickly realized that its approach could help companies build applications as well. The company's Palmyra LLMs were designed specifically for enterprise use, and Writer built several security, privacy, and user-interface tools around those LLMs that help business leaders and IT leaders work together on generative AI projects, which Habib said is critical given the hesitancy to put output generated by LLMs in front of customers or partners inside many companies.

"The vast majority of teams that are trying to build generative AI applications are approaching it like they would traditional software development, and we think that's kind of like bringing a knife to a gunfight," she said.

In an interview last week at the HumanX conference in Las Vegas, Habib discussed the company's decision to build its own models, the concept of AGI, and why it has taken so long for companies to get up and running with agentic AI.

This interview has been edited and condensed for clarity.

Why do you build your own models? Obviously that's a complicated, expensive, time consuming process, models are changing rapidly and there's a lot of different companies that have a lot of different approaches.

Let's start with the beginning. The story of Writer is the story of the transformer. We started the company to build models, and with every big step-function change, we've either led it or been like four weeks behind. If DeepSeek proved anything to the world, it's that you don't need billions of dollars to do this, and necessity is the mother of all invention.

We invented using synthetic data. We invented something called the frugal transformer. There's been a ton of just literally architecture after architecture improvement that has allowed us to build state-of-the-art models at a fraction of the cost. Now we've raised $300-plus million, so it's not like, this is no money. But even when you actually look at what [foundation model companies] spent on GPUs, it is not 100x what we have spent, it's 2x to 3x. What they're spending money on is data and data labeling.

So back to, why do we build our own models? Well, they're the only enterprise grade ones. Open source is not really open source, it's open weight. How do we give a Goldman Sachs or a UnitedHealthcare information about our training data if we don't know what's in the training data? We can make data residency decisions. We can make data-retention policy decisions that are enterprise grade in a way that the major LLM labs cannot.

AGI doesn't mean God. It means super, super good.

I think enterprises quickly learned that building, fine tuning, and maintaining LLMs is not their core business. And for five years now, we have kept up and led many of these innovations, with the next one up being self-evolving models, which we believe is the path to super intelligence. These are models that are able to update their training data in real time so they can learn from every wrong interaction, which is going to be very critical in agentic AI.

Do you believe in AGI?

I think it actually doesn't matter. Super intelligence is the term we use to really give shape to this idea that we're going to have, very soon, technology inside of companies that knows more than any one person. But it is an AI system, and it is controllable, and it has to be observable for it to actually work at scale. AGI doesn't mean God. It means super, super good.

With respect to agents and agentic progress in general, what's been holding folks back over the last six months?

Right now, the agents that we're seeing scale fall into really two categories. One is things people couldn't do before and weren't doing before.

When you are creating a process that automates 20,000 different websites for a major hotel chain by connecting to the source of SEO data directly, with no human intervention, and then actually making all those updates right programmatically and only nudging the user when you know one of those updates might break something, that is really, really cool. You just simply would not be doing that with humans. That actually is what makes those categories of use cases much easier. There is no change management, because there was nobody doing the thing. And from a data perspective, you're really able to use systems that are there. You're not inventing what good looks like, right? So there's less of that cold-start data problem.

Then there's a category of use cases that is much harder, and this is in spaces where you already have quite a bit of automation. So, claim adjudication, medical summarization, customer support … where you know five percent or 10% or 15% of the cases require human intervention, and that's taking up 80%-90% of the work.

There aren't these clean handoffs where if you wanted to do the whole thing in code and just hand off an application layer to a business user, that produces a fine application. Generative AI challenges the concept of done.

The reason those are hard is because there are no standard operating procedures that are written down, and a lot of the know-how on how you handle those exceptions sits in people's heads. What makes the second category of use cases tougher is that it's a very manual process pulling out that data, and there's a lot of coordination that needs to happen. Very few enterprises are in the kind of crisis mode required to do something like that very quickly.

I have seen the second category of agentic use cases really, really take off when there has been a crisis of some sort, where a supply chain issue totally broke the current support model, and they needed a load of agents, or they needed 24/7 support for some reason and couldn't manage it with humans.

We heard for years about low code and no code, and that's actually kind of happening right now. Do you see yourself in that category?

We are low code, no code and all code. The reason that all three are important is this is not a traditional software life cycle.

There aren't these clean handoffs where if you wanted to do the whole thing in code and just hand off an application layer to a business user, that produces a fine application. Generative AI challenges the concept of done. You need the ability to continuously be building and monitoring, and you do want to give different users who've got the subject matter expertise to contribute; whether it's a systems integration or a process update, you need to give them tools that fit their level of technicality.