Editor’s note: This new Relay by Chargebee series features field notes from AI-native SaaS founders as they tinker, build, and monetize their paths towards a new kind of product-market fit playbook. This #2 post captures the story of Sudowrite, in Amit’s words. Dig in, drop a question in the comments, and if you have a suggestion, write in as well.
Here’s an index documenting the top takeaways, to help you find the right sections quickly:
- Model-level blindspots with generic AI tools: Sudowrite’s founding insight
- The modeling-scape enabling Sudowrite’s “non-generic” assistant
- Why human-AI interactions will go beyond chat
- AI SaaS’ non-zero marginal costs and the dance of iterating through interim pricing model(s)
- Designing the highest-value free trial and the non-trivial role that product education will play
- The further-ordained position of pricing in finding PMF
- The AI-led “everyone makes their own software” era might be overstated
- Why the uncertainty of AI feels unlike any other tech shift
Model-level blindspots with generic AI tools: Sudowrite’s founding insight
When GPT 3 came out, it was the first time that a model was capable of doing something useful from our point of view. Those models were really good at generating nonfiction, marketing copy, email copy, that kind of stuff.
We did think about these potential verticals before we started building. There were two problems. One, we just found them uninteresting.
Two, we figured these would be crowded markets because with that type of output, there’s a lot less subjectivity to deal with. If you’re writing website copy, you A/B test it and there’s a good and a bad version, there’s a best and a worst version — with fiction that doesn’t happen.
With fiction, there’s a large gradation of subjective quality. So we felt that there was more interesting work to be done here. You can’t ever get it right. You’re always trying to get it right for the right person at the right moment.
More than that, our audience uses Sudowrite to draft fictional narratives, both inspired by what exists in their lives and what doesn’t. Sometimes these stories involve violence or sex or other topics that model makers consider taboo.
If you are a fiction writer (my co-founder, James, and I, are science fiction writers) using traditional foundational models, this can be a massive challenge, because these models have been trained to censor these topics.
When we started building Sudowrite, we quickly realised that this was a huge problem for writers.
If someone tries to write and the tool says “Sorry, you can’t do that”, it’s incredibly frustrating. Writers are used to writing in Microsoft Word. MS Word would never censor the topics they can write about.
But the generic AI writing tools today do.
Another issue is that a lot of these models, even when they don’t censor what can be written, are built to serve you as helpful assistants — with limited creativity. They’re good at writing memos and emails, they’re good at customer service, but they’re not good at writing dramatic narratives.
So with Sudowrite, we are doing two things:
- building a tool where the writers don’t have to deal with the frustrating limitations on what they can and can’t write about,
- and providing them with a set of tools for writing prose, in a style, tone, and voice that feels like them.
The modeling-scape enabling Sudowrite’s “non-generic” assistant
We are building Sudowrite by innovating on two layers: the prompting layer and the model layer. Today, we use a few dozen models, all of which have been heavily evaluated for their specific edges/advantages.
Some models are better at extracting salient information from a user’s story bible, which is an area of Sudowrite that gathers the core elements of your story and acts as a source of truth as your work develops. Others are better at what we call the tip of the pen, which is actually writing the first draft prose as the final step.
So each step along the pipeline — from a user input to a final version that the user likes — calls on multiple different, specialised models. In addition to these under-the-hood models, we also offer a lot of models on the front end for the user to choose from.
We’ve been surprised that our writers have pretty strong preferences for different models.
People who are just getting into writing with AI can’t be bothered to learn about different models, but those who’ve been working with this stuff for a year or longer often have favorite models.
There’s also a lot of the work we’re doing in training models specific to an author’s voice. We have a beta program where authors can choose to upload work to train a model just for their own use. It’s something they’ve been asking for from literally day one.
Why human-AI interactions will go beyond chat
A weakness of many AI tools is that while the chat and conversational interface is really good for making large changes or creating something whole cloth, it’s not so good at making small edits or manipulations.
That’s something that you can do easily with traditional UIs. You can just select some text and hit bold. But in ChatGPT or Claude, you can’t even bold something today. You’d have to give it a prompt like, “Can you bold the third word in the second sentence?”
And obviously that’s not how anyone would want to do it. Chat feels incredibly refreshing as a UI, and at the same time, like a pit stop on the way to something great.
We’re playing with voice and we’re playing with ways to combine the freedom of chat with the power of direction manipulation interfaces, but it’s tough.
I don’t think anyone has figured it out. People have definitely arrived at better ways of doing things, but I don’t think anyone has figured out what interfaces are going to look like in 5 or 10 years yet.
There’s a calendar and productivity app called Amie that I think has played with stretching the UI in more interesting ways.
I’ve used a few different tools for meeting transcription (I love Granola) and it feels like all of them are adopting AI in a similar fashion: enabling querying of your transcripts through chat.
But Amie’s latest launch is fascinating: they launched a feature that identifies recurring meetings, like a weekly standup, and creates a timeline of those over time to show the evolution of that call over time.
I’m not sure yet if there’s something useful there, but I like that they’re taking a step beyond the obvious and asking interesting questions.
AI SaaS’ non-zero marginal costs and the dance of iterating through interim pricing model(s)
Pricing is definitely more complex for AI products.
We are at our fourth attempt at a pricing model, and it’s still not right.
I feel like I have a bit of an advantage because my last company was in e-commerce, so we had real cost of goods in each transaction. And this is the first time in SaaS that you also have real marginal cost to deliver the final product if you use AI. The era of zero marginal cost SaaS might be over.
Originally our pricing model was flat, pay $5/month (we later charged $10, then $20) and just use it. There was no specific stated limit, but if you went crazy, we would contact you. It soon started showing cracks, though, we had people who were paying $20 a month and costing us $400 a month.
That’s when we decided to switch to AI words. Each plan came with a certain number of output words.
We wanted to find a way to correlate the value users are getting from the product with what they were paying for it. It’s crude because words aren’t really “value”, but it was as close as we could get in a meaningful way.
But as we went to a multi-model world, we had to switch again, from words to credits.
With the level of flexibility that came with offering different language models, and the fact that we let users develop their own custom tools in Sudowrite, using whatever prompts and LLMs they want, it meant that someone could use the most expensive models at each step, inhaling multiple novels as input, and only output a single word, one word that cost us $30 to make.
That was not sustainable. We couldn’t charge by words as an output metric.
Credits was an attempt to more closely align our costs with what we were charging customers. It also allowed us to offer greater flexibility and customisation, as well as lower costs for the vast majority of our users.
Since our costs were better aligned with value, we could offer users a choice of models to use at the tip of the pen as well as allowing them to build custom plugins to do whatever they wanted/needed, specific to their writing workflow.
This usage-based model comes with its own challenge: it’s confusing.
Because we charge different amounts of credits based on the different models you use at different steps along the way, it can be really hard to predict exactly how much pressing a button will cost you.
You can’t always make an accurate guess at the purchase point.
We’ll continue to evolve this. We want to make it easier.
To help with the confusion, we do offer a one-week easy cancel policy because we understand that these pricing models are new for our users.
One useful parallel that I go to is how they do things in the apparel industry. Where, again, choosing the right size (plan) of t-shirt solely from photos is quite difficult.
Ultimately, the solution that worked is that you can buy the t-shirt and you can exchange it if it’s not the right size. I think that’s the kind of solution where we’ve landed too.
You can return or exchange any plan in a week if you buy it and it doesn’t fit.
Designing the highest-value free trial and the non-trivial role that product education will play
All of our plans include free trials. We have had a similar journey of iterations with our free trials as we did with our overarching pricing models.
We started with a flat 2-week trial, and then switched to a usage-based (AI words) trial as we ran into the familiar set of challenges: of finding the balance between value for the customer and cost to the business.
Eventually we decided to switch to a credits-based free trial where users get a limited number of credits to use. People blaze through them at different rates, of course. Some people blow through it in an hour, some people take a few days or weeks.
We want to give people a taste of the capabilities of the product and time to see if it’s right for them, especially because writers’ workflows vary a lot. And it takes time to see if something fits your workflow.
But we’ve had to clamp down on the trial length because a huge percentage of our overall inference costs go to servicing trials. When you’re writing a novel, each inference call might use 10s of thousands or even 100s of thousands of tokens of input for context. These are very real numbers.
Interestingly enough, we have seen that shorter free trials convert better for us. Perhaps because it creates a sense of urgency and our ideal customer, a professional author, is buying our tool to help them with their business and they’re serious about making a buying decision.
Another element in the design of our free trials is education. We do a couple things.
We’ve taught classes for a few years now; we teach at least one class every day and we have different topics throughout the week. It’s a good way for people who are still learning the program to get up to speed without having a long trial potentially.
We’ve also started experimenting with one-on-one onboardings where you can schedule a time with us and actually learn how to use the program with an expert over 20 or 30 minutes and then get more out of the trial.
Creative tools can get complex, and educating people on how to use them and how to get the most out of them will become increasingly important.
The further-ordained position of pricing in finding PMF
Pricing is fundamentally a part of our product. It’s as much about making the product approachable as it is defining who the ideal customer is.
When we started, we thought $5 a month would work for anyone, but we couldn’t build the business we wanted to, with the level of service we wanted to offer, at that price. We worked to find the right price point that works for our ideal customer and the business.
Last year, we did a deep dive to understand who gets the most value from Sudowrite over time. We discovered our core users are professional authors—people who’ve written more than one book and essentially earn their living through writing.
For these writers, our product is an easy purchase. It could probably be three times more expensive and still be valuable because it helps them write much faster and produce better books.
On the flip side, we have aspiring writers and students for whom $10-20 a month is prohibitively expensive.
We’ve consciously decided not to create a lower-cost product. Instead, we’re focusing on being the highest quality solution in our segment, which means we can’t serve everyone.
The AI-led “everyone makes their own software” era might be overstated
As software becomes easier to build with AI, we’ll see more options in every vertical. There will definitely be a lot more software out there.
It’s so much easier to build software today than it was before AI and that’s a relatively recent phenomenon.
I’m hopeful people will make more software for themselves. There is this growing notion of personalized software, bespoke software. I’m a little skeptical that it will be as big as we think.
There was a period where personal publishing was a really big thing and everyone published their own blog. Then it became easier to just post tweets and occasionally post photos on Facebook or whatever.
And because it’s easier, that’s what people did. Most people don’t want to do the work of setting up a blog and trying to get readers even though it’s trivially easy.
I think the same will be true of software for most people—if a solution exists and it’s good enough, they’ll prefer to use that than to create their own software and make the hundreds of decisions that go into making a piece of software.
But I see tremendous potential in allowing people to customize their software. It’s easier for people to see what could be different or better than it is to create something from scratch. It’s easier to edit than it is to write.
Like, “Oh, I hate how my video conferencing software works. I wish it had feature X. I’m not going to build a whole new Google Meet myself just to get X. But if I could prompt it and add that feature easily, I might customize it that way.”
Why the uncertainty of AI feels unlike any other tech shift
We have seen tech advances before, but AI feels very different—it’s incredibly powerful and almost magical.
The pace of change is unprecedented. Every day, there’s a new model with capabilities that seemed impossible the day before. It’s mentally taxing to keep up and imagine what’s possible.
As a founder and a new parent, I’m both excited and uncertain. How do we prepare for a future where AI might transform everything?
It forces us to stay incredibly focused on the present moment and the specific problems we’re solving.
Ultimately, software exists to solve problems. As long as those problems persist and we can build the best solution, I think we’ll be okay.