SFG 31: May Habib on taking on ChatGPT enterprise

‍

In this episode of the Startup Field Guide podcast, Sandhya Hegde chats with the co-founder and CEO of Writer, May Habib. Started in 2020, Writer is an enterprise-grade gen AI platform for larger, security conscious companies. Writer has over 150 customers including teams at UnitedHealthcare, Uber and Accenture. Last valued at over $500M, Writer makes it easy for enterprises to build internal apps that generate a variety of content from sales enablement to blog posts. 

‍

‍Be sure to check out more Startup Field Guide Podcast episodes on Spotify, Apple, and Youtube. Hosted by Unusual Ventures General Partner Sandhya Hegde (former EVP at Amplitude), the SFG podcast uncovers how the top unicorn founders of today really found product-market fit.

‍

If you are interested in learning more about some of the themes and ideas in this episode, please check out the Unusual Ventures Field Guides on building conviction in an early-stage startup, understanding the enterprise buyer, and LLMs in security.

‍

‍

Episode Transcript

Sandhya Hegde

Welcome to the Startup Field Guide, where we learn from successful founders of high growth startups, how their companies truly found product market fit. I'm your host, Sandhya Hegde, and today we will be diving into the story of Writer. Started in 2020, Writer is an enterprise grade Gen AI platform for larger security conscious companies.

‍

Writer has over 150 customers, including teams at companies like United Healthcare, Uber, and Accenture. Last valued at over $500 million, Writer makes it really easy for these enterprises to build internal apps that generate a wide variety of content, from sales enablement to blog posts. Joining us today is May Habib, Writer's powerhouse co-founder and CEO. Welcome to the Field Guide, May.

‍

May Habib 

Thank you so much, Sandhya. I've been such a fan of all of the incredible content you put out for founders, so I am thrilled to be here.

‍

From Qordoba to Writer

‍

Sandhya Hegde

Oh, thank you. You are too kind. So diving in, May, I am sure the origin of Writer all starts with Qordoba, your first startup. Which, I think you started all the way back in 2015, when I first met you. And so I'm really curious, what was the inspiration behind Qordoba at first and how did it evolve into Writer five years later?

‍

May Habib 

Yeah, great question. So for me and Waseem, my co-founder across my two entrepreneurial ventures, we really wanted to figure out a way to accelerate what was happening in localization and translation with Qordoba. And, we used some pretty incredible machine learning during that time to build what was a localization platform, enterprise as well, so a lot of things that we have learned we've been able to carry through to Writer and using encoder decoders in that business is really what led us to Writer.

From a problem statement perspective, we kept seeing customers run into this content quality problem regardless of what team they were on from sales and support to marketing and documentation around what people called like source content, i. e. English language content, and so the idea was, “Hey, could we use these incredible transformers to actually work on source language?”

And it really did mean needing to build a completely different product and a completely different company. The go-to-market needed to be very different. And we started Writer in Q1 2020 to address and go after that opportunity.

‍

Sandhya Hegde

And, at this point, I think OpenAI had just released GPT 3, or maybe it hadn't quite happened yet; it was still a previous version and they had already, famously raised and spent millions of dollars on model training. What was your approach? Like, how did you, how were you contrasting what was happening with people using transformer technology to build startups at the time versus what you and Waseem wanted to do with Writer.

‍

May Habib 

 Yeah, everything on kind of the editing and revision side of our generative technology, we had built in house. When GPT 3 came out, it was just such a step function advance in the quality of long form content generation, which we hadn't gotten to either from a quality perspective, and so there was a short period of time where we integrated that into our product in our stack, knowing what we were going to be trying to do which is build our own with the functionality that we needed to do what we wanted to do. And for us, the GPT-3 didn't make a big difference in our go-to-market or our product strategy because the reception actually was basically the same.

What really made a difference was ChatGPT reducing the risk of exploring and implementing this type of technology, and that really changed the game in enterprise. And luckily for us, by then we had made excellent progress, and on PALMYRA we were able to use all of our own models for serving up all of our generative features and functionality. And from a security perspective, it lined up great with our posture.

‍

Writer's early product vision

‍

‍

Sandhya Hegde

 And, what an amazing 18-month window that was between these two events. I'm curious, if you go back to, the pre ChatGPT era, where your buyer's risk profile is very different, market awareness is very different, how would you articulate Writer's vision then?

What was your product vision, how would you define the customer profile you were going after and what were they looking for that was very different from trillions of parameters, large models…how did you define that early product vision?

‍

May Habib 

 I think we've got this unfair advantage of having had to sell generative AI before there was a word for it. That makes it so that we don't talk product; we talk customer. We don't talk about tokens and parameters; we talk about words and use cases and capabilities of the models.

‍

But, it's a great question. That vision slide that you put, at the beginning or at the end of your seed stage deck? For us, when we were raising, this is Q1 2020, it had that big inflection curve, and it was like AI writing assistant to AI writing. You cut out the assistant.

‍

And we actually put that in our decks to. Now, I definitely remember it was probably the Editor in chief of Mayo Clinic or somebody where it's like, “Kill that slide. They're not gonna want to see that slide.” And so it was, those early months, was a lot of, “What is this? What is it based on? Wait, how did you do that?” And the positioning was AI writing assistant. Now, of course, Grammarly is the comp right? And so even as we were able to do more generative stuff, the go-to-market was just much more natural to go in on content quality. And within technical documentation teams where we could do a lot more than what Grammarly was able to do, just be a lot more prescriptive, a lot more customization. Now, with those folks, those are actually the last folks that want to be creative and generating ideas because they know what they want to say. They just got to say it much more clearly and concisely and have it land, and we were able to do all of those things.

‍

In the content copy and ideation side, Phrasee was in that space, is still in that space, wonderful company. Persado is, I had coffee with Alex this morning, also excellent company and strong product-market fit. And for us, it was just a lot less interesting given where we thought generative was going for us to look backwards in a way, it wasn't bad. This is a big space. We weren't that interested in running A/B variations and like simulating personality and like trying to actually own a conversion number as a product. There's just a lot of complexity and we love being close to the end user. And I think, localization and translation gave us that. As mind numbing as it was sometimes, translators are very specific people, there's something really beautiful to the the hardship of needing to serve very picky end users who are deep into the craft versus admins buying tools for people or someone who is okay with a black box of just, make it do this, and not really hands-on in your interfaces. We stayed in that space.

Now, we were multi-use case and multifunction from the beginning, like functional area. Grammarly was doing this as well. I think they were doing it a lot more in support and because we had just much deeper customization capabilities because we were using generative and not AI 1.0, only NLP techniques and rules based stuff, we were able to address marketing use cases and NLG use cases and support use cases, etc. So the playbook, kinda, 2020, 2021, even when we raised our A was landing marketing and then let the product go across the business. When we did our A, we were probably, like 60, 70 percent of the lands were in marketing but then about half were using us marketing plus one, two, three different areas.

Both those numbers now are much, much higher so most customers use us on three or four different teams. But the series A vision statement was writing is the last unoptimized business process. And, can we approach writing in just a much more methodical, programmatic way?

And again, not having words for generative, right? That's how we explained it is really being able to push a button and have a few parameters and get what you need to do the thing. 

‍

The evolution of the AI buyer

‍

Sandhya Hegde

Like you buyer you care about business process optimization. So how do we help you understand the opportunity or in your language, right? I think, yeah, the challenge of being too early to the spaces, you don't get to kind of align with the tailwind when everyone is okay, this is the thing, I will stop questioning it.

It's just fascinating. So I'm curious. I think a lot has changed about your buyer, right? Maybe actually more has changed about your buyer than your product vision as a company over the last three years. So I'm curious, one thing I think you talked about is this horizontal strategy and the buyer's view of that has probably changed.

I'm also curious how the buyer's view of security has changed or has that, you have always focused on enterprise, what did they ask you about, or were they more educated about, “Oh, is this, are you going to use my data to train your models?” Is that a question you ever got before ChatGPT?

‍

May Habib 

Yeah. 

Sandhya Hegde

I'm curious how the buyer questions have evolved.

‍

May Habib 

It's such a good one, we could probably spend three hours just on this question because I don't know that any other space has had as much like buyer transformation. We were selling to directors of content marketing, directors of product content, directors of content design  — people who really needed the words to be very specific, to do very specific things.

‍

They needed to build, just a much more programmatic, systematic way. Systematic, I think, is the right, systematic, systems of language is the right. Whoever owned the system of language at a company was our buyer. Fast forward 18 months, you've got boardrooms. I've spoken at 25 executive board meetings, right? In the end it's nuts! Just in the last quarter, this is a board level priority and you know the kinds of questions now that we get are folks thinking out loud with us; a COO of a Fortune 100 bank, being like, “Okay, they've got a part of their business where they sell to other banks. Can I take a two week onboarding process and make it a two day onboarding process?”

‍

And we're like we're gonna workshop the shit out of it, but yeah, I think so. And people are thinking so broad and so deep because they’re understanding and the technology, of course, has made huge amounts of progress. Absolutely, the envelope has been pushed by OpenAI and it's great because they forced all of us to catch up. I think the beauty of the 18 months of progress is, we definitely have buyers, especially executives who are like, wow, human beings are no longer a constraint to my growth. They are no longer a constraint to my business model innovation. What can I do?

And, even when what they want to do may not be possible, it's just pulling everybody into this very open mindset. And so many of our customers, people who —  and I would put Accenture, United Healthcare, and Intuit in those categories — the champions that bought Writer in 2020 and rolled it out to three, four teams and the new stuff that we rolled out they leaned in. They are heroes! And they have gotten to be in charge of a lot more stuff and so that is so exciting.

So the buyer now encompasses a much broader set of personalities much higher up in the organization. Our actual buyer, you know the first like few million in sales, much more empowered. And so we've got a much closer relationship with them. It just drives the roadmap and drives everything else that you're able to build.

It's really remarkable. I remember when we took, this is 14 months ago now, the first web pages live, where we talked about PALMYRA and our generative AI functionality, etc. in real depth. And, the team was like, there's no way we're going to put LLM in the subhead, Mae, come on, writing it out.

I'm like, guys seriously, I really think this is going to be fine. And now, we walk in and people want to understand, “That fusion and decoder paper that your team wrote, so instead of RAG or in addition to RAG,” like these are non technical people, it's awesome! And, people are discovering that a proof of concept is possible. Generative AI at scale is hard and expensive.

‍

And we have definitely learned a lot, the platform is the leading generative AI platform, I think. We've packaged things together in ways that make a lot of sense for people who want to do generative AI at scale. But there's no way we could have done that if the buyer had not evolved at the speed that they have.

It'd be 2020 all over again, right? People being like, wait, what is this?

‍

Writer's approach to security and compliance

‍

Sandhya Hegde

“I won't be able to convince my leadership team to lean into this, as opposed to my leadership team is demanding that I deploy this faster.”  Right? Very big difference. So you brought up a really good point which is the huge difference between prototype and at scale and production.

How much does the security and compliance requirements play into that? And at what point did that become the important investment in your roadmap? Was that kind of always a baseline requirement to work in the enterprise? Or was that something that you as a company have evolved your vision around as well?

‍

May Habib 

Yeah, this goes back to us moving from Qordoba to Writer and really starting a brand new company. Like, what we did end of 2019, early '20 as the pillars of our product strategy. The enterprise first, enterprise writing as data security, was a big part of it. We won a lot of RFPs and bakeoffs against Grammarly because we analyzed data transiently and we didn't save down any data that we saw in a browser. We were in there for a millisecond, putting in the edit or the generation and then getting out. And the data was not retained. It was, like in a segregated environment that was all auditable.

‍

All the Chrome extension code was transparent. You literally, if you're an enterprise customer, you could have completely rebuilt. We gave you so much transparency. You could have rebuilt a bunch of the stuff and so we took that into everything that we did. We spent 18 months curating and cleaning the data that went into PALMYRA and it's top of the performance charts, and a big part of that is just the investment that we made.

There isn't proprietary customer data in our foundation model training data. We didn't do secret deals and buy a bunch of stuff. Like it's what everybody else has access to and actually much more refined for copyright and other stuff. So the data awareness, the enterprise approach, all of that has been baked in.

‍

We've been SOC 2 Type 2 since 2020. So we've literally, we're in our third audit cycle, this is a big part of our culture and what we do. In building our own models — we have 17 models, a number of them are on Hugging Face — and in just our approach to how we treat all the content that folks give us for RAG; for building digital assistance; for fine tuning; for index building, all of that is treated with exceptional consideration to privacy. The deployment options that we provide, multi cloud, single tenant and customer managed as well; all of that has been in the DNA from day one.

‍

Writer's approach to addressing hallucinations

‍

Sandhya Hegde

Makes sense. I'm curious about maybe a year ago, once ChatGPT took off the idea of hallucinations and the risk of hallucination, especially when you have a conversational chatbot use case, became front and center for a lot of risk averse buyers. And you're definitely working with kind of knowledge oriented companies, right? Whether it's Accenture or others that really care about accuracy. How did Writer approach the hallucination problem? And, is it a problem for your customers? How do you think about it?

‍

May Habib 

Yeah. Writer is a generative AI platform so we've got the LLM with a knowledge graph set of functionality, right? What other folks call RAG, and it's RAG plus, and AI guardrails together. It's composable, most people are using all three parts to get the generations and the answers at the quality, the scale, the cost, the inference that they need for production grade. Hallucinations, whether the use case is an assistant and a chat interface, or it is an editor input/output style, hallucinations have a lot to do with what we're actually trying to do, what the use case is, like the approach to limit hallucinations. In the kinda knowledge augmented, retrieval augmented generation (RAG) use cases where you're drawing on companies' own data, ontologies become exceptionally important.

‍

And so for us, use case is almost a proxy for what ontology do you call? So it could be this similar set of data that is actually processed and reviewed and analyzed in different ways depending on what the use case is. We've been in pretty exceptional circumstances where folks who've only gone to 50, 60 percent accuracy with a copilot or with Azure AI, et cetera, get to 90 percent plus with us as a result of our kind of multifaceted approach and, at a high level, what I can tell you is we're not trying to boil the ocean.

We narrow in and then apply the techniques that are going to get folks the right answer. We provide the sources so people can trace it back. You've got all of the considerations there, but you gotta be able to do it at scale because there might be some use cases where the timestamp for example, you actually want to privilege, like the last added right?

‍

And there are other use cases where it's actually giving you the oldest one. That's actually the one that's most trustworthy. And so the hallucination control has everything to do with that multi pronged approach use case, right ontology, using RAG very effectively. We actually don't use vector databases for most of these use cases because they're expensive, they're hard to update and just the scaling of high accuracy, we just haven't seen those results relative to our knowledge graph approaches.

‍

In the generation side of the use case, the use cases actually, it's a really similar approach of tracing back the sources, but you're not tracing back the source of an answer. You're actually doing it claim by claim, and that claim could be a fact. It could be a statistic. It could be a quote.

‍

And so there's a mode where we actually give the end user sort of the overlay where they can look at where something came from and the research side of the house is tracing back even down to the weights. There are going to be techniques that we bring to market that allow us to go even deeper than what data did we get from your knowledge graph that has fed these generations or these answers.

So it's multi pronged, the user absolutely has to be in the loop and for kind of high volume use cases in retail, for example, where you're inventing a new workflow and it's at a volume where actually, other than random spot testing and like A/B analysis and just looking at the data, there's no way to tell if there are hallucinations in there, then it's other approaches, right? It's fine tuning. It's turning down the temperature. It's all the other stuff.

‍

Writer's conviction in Enterprise AI

‍

Sandhya Hegde

Makes sense. I'm curious. I know, the last 12 months have been incredible at Writer, so congratulations for that, first of all. But before that, were there moments where you did something that really wasn't working or, it was a great prototype but you were having trouble figuring out the right, the buyer or the retention. I'm curious whether there were moments where you were able to clearly identify a pain point you hadn't thought of and how you responded to that in terms of your product strategy, your company strategy. Any kind of unlocks or surprises along the way that you were able to learn from?

‍

May Habib  

Yeah, definitely. There was definitely a time 14 to 16 months ago for two or three months where we stuck to our guns on enterprise, but definitely privately had some doubts, right? Like tons of people making a shit ton of money in kind of prosumer.

‍

And it makes not a lot of sense but is that something we should be investing in? And, till today, our funnel is 95 percent inbound. People come to us but it's product led sales versus product led growth. People aren't putting in a credit card, not for the majority of our ARR. The self service ARR is a meaningful business on its own, but it's just, it's not what we're focused on doing at all. It wasn't necessarily so obvious every month since the series A, “that's what we should be doing.”

‍

Sandhya Hegde

Yeah, no, I think that's such an important question because you have your conviction around what makes Writer uniquely awesome for enterprise, but then there are all these other companies with millions of people swiping credit cards, checking it out. What was, how did you think about your own decision framework for it?

What was your internal argument for not completely investing in a self-serve prosumer play?

‍

May Habib 

Yeah, I think I'm so biased around how hard it is to do both things at once.

‍

Sandhya Hegde

 Right.

‍

May Habib 

And from my experience, Qordoba to Writer, just really understanding that like to truly succeed you got to burn the fucking boats. And we just weren't gonna burn the enterprise boats. These are our favorite people in the world, like these users and these champions.

And so then, is grass always greener? I think we prayed on, well, I can't speak for Waseem, but I definitely prayed on that, “God, please make sure we're making the right decision here.” Because I do think you gotta commit one or the other, especially at these early stages, 

‍

Where Writer's platform is going next

‍

Sandhya Hegde

Makes sense. Yeah, no, I think that a lot of things that you guys do really well especially on the security compliance side is stuff the prosumer market does not care about and is not going to value you for, will not consider you differentiated for. So yeah, no, I think your conviction has definitely proven to be an asset.

What do you see as the next big horizon for both generative AI as a technology, but more importantly for Writer?

‍

May Habib 

Yeah. I just love, I love what we're doing over the next couple quarters to help bring generative AI to the mainstream. Really, there's going to be a lot of stuff coming out that I'm very excited about. That takes everything that we have learned, we've learned with our customers and makes it much faster, even faster time to value for folks who don't have the same sets of resources or time or engineering capabilities, to do a lot of what the early adopters can do. So I'm really excited about that in the next quarter and two and then, back half of next year, being able to really incorporate what other folks call agents, we call workflows.

But the ability to really take LLMs from helping orchestrate tasks to helping orchestrate work. So it's going to be really fun.

Sandhya Hegde

Yeah, no, I can't wait for it. I think the accessibility problem is like probably the biggest differentiator between enterprise and consumer for me because even as a consumer of these tools myself, I'm happy to do trial and error over a weekend figuring out how to make something cute.

‍

But if I'm at work, I want it to be fast and reliable. I don't want to do any trial and error, right? So that's how I think about accessibility. It's not that like you are catering to an audience that understands it less. It's just that their decision criteria for what software is reliable and easy to use is very different depending on what mode they are working in.

‍

Are you thinking about going multimodal and what does that mean for your customer?

‍

May Habib  

Yeah. Multimodal for us is going to be for our customer first and not prosumer or consumer. So yes, imagery will come but it is going to be in the context of what folks need to know and need to do in the use cases they're using Writer for. So we already parse video, audio, can read charts, read images.

‍

In creating charts, we'll probably do that first. Creating images, it really has to make sense because in an enterprise and in a security environment, high security environment, I don't want to be in the generic blog post banner making business, like you can get that for free anywhere.

‍

It's narrowing in on the use cases that we're going to be able to uniquely do. We work backwards from a world where, just from a product strategy perspective, where excellent generative AI is built into every system of record. We're nowhere close to that, but like we work backwards from that assumption. And consumer grade frontier model access to multimodal functionality is free. And so what do you build in that world? Nobody who is going to build a big business can't be thinking about those two realities.

‍

May Habib's evolution as a CEO

‍

Sandhya Hegde

Maybe that's a good segue into kind of thinking about your journey as a CEO. What everyone who knows you well tells me is you are really exceptional at understanding what is something that you need to learn and who you want to learn it from, which I think is, probably just as important as making the decision, “okay, I want input on this idea or this problem.” I'm curious, how have you thought of yourself as a CEO in terms of what were your strengths when you started? And what were things that you have grown and invested in over all these years? And your advice for other founders who are getting started on the same journey?

‍

May Habib  

I think similar to what we search for in our hires, this kind of sense of curiosity because you're just curious in and of yourself. It has nothing to do with anything extrinsically motivating to most people. You're just curious. I do think so much of where we are today is because, me and Waseem just asked, pretty busy people in the enterprise about their life and their day and their jobs. And really deeply, I think the competitive advantage of our early team, the product team and the early CS team is we know more about content and knowledge workflows across a company than literally any team and I think that's reflected in how we build product. That probably, and it may be a cliche, just not giving up. Having such outsized belief that we will find the answer no matter how long it takes. That's definitely been something that has served us well. 

‍

So it's been the curiosity to work on something interesting and impactful and then scaling that for like real massive at scale kind of contribution to the world. And I think there's part of our culture that really believes in folks with a very steep learning curve, just like keep them on that curve and keep putting new challenges in front of them, and they will find the answers. And at the same time, benefiting from the lived experience.

‍

We are definitely a culture of no prior playbooks. This is a completely, so much of this space is a rip of every space time continuum fabric that has come to define B2B SaaS for the last while and we want people who can think from first principles. 

Sandhya Hegde

Awesome. Thank you so much for joining us today, May. This was amazing, and I can't wait to see where you take Writer. And, so helpful, all the learnings you have shared. Thank you so much for joining us on Field Guide.

‍

May Habib 

Appreciate you. Thank you.

‍