1 week-1 idea: introducing Boss.ai!

This article follows Should we all learn to write better? and Use GenAI to improve your writing.

Most of you can work remotely, at least partially. Your co-workers or managers might live in a different country. Your agendas are filled with Zoom calls or Google Meets. You company is probably trying hard to reduce the number of Zoom meetings. Like Shopify, now showing meetings’ costs to discourage excessive scheduling:

More remote work and fewer meetings require greater use of asynchronous communications. Which means that writing documents is an increasingly significant part of our jobs. And writing becomes a more and more crucial skill to master for all of us.

Unfortunately, from my experience as Ankorstore's VP of Product, the majority of employees struggle to write effectively, with:

• Typos, grammar errors, etc. (ie. No proofreading)
• Unclear & lengthy wording
• Poor storytelling

As I mentioned in my previous article this is not Ankorstore's isolated problem. These issues were famously encountered by Amazon, within political administrations like those of Churchill and Obama, across the finance sector, and in pretty much every company of every industry and with everybody.

It is estimated that enhancing employees' writing skills could bring trillions of euros to the global economy. GenAI, with its capability to generate flawless text, appears to be a breakthrough technology to help solve this issue.

The boss.ai Product

I already showed how to use the chatGPT interface to improve our writing. However, this method is time-consuming as it requires extensive copy-pasting.. So, here is an idea of a way simpler Product:

I believe that using a Chrome extension to access a GenAI algorithm within our existing tools (e.g. Google Docs, Notion, etc.) would provide the smoothest user experience. From there, the AI could write comments, similar to how a manager would suggest improvements:

These comments can be about typos, writing style, but also to suggest shorter paragraphs, or to improve the content (ex. “specify why clients find the feature painful”).

For the MVP, our AI can be prompt engineering on top of chatGPT API:

With just one click, anyone could have their documents reviewed by an AI and improve their writing before sharing with managers or stakeholders. By helping employees improve, managers will spend less time deciphering documents, and ultimately companies will make better decisions!

How Boss.ai can be better than existing Bn$ alternatives

The SaaS market for 'AI text generation' is crowded. Dozens of startups are operating in various fields such as email assistants, internal document search engines, programming, and marketing copy generation - the more crowded. The segment we target - writing assistance - is actually the less crowded.

In all these verticals we can find companies with already >$100m of funding. Most notably:

• Jasper.ai (copy generation): 131m$ raised, 1.5Bn$ valuation, >80m$ of revenues
• Grammarly (writing assistant): 400m$ raised, 13Bn$ valuation

I tried Jasper.ai and it saved me so much time that I’m paying 39$/month already, just for this:

• Jasper’s Chrome extension, so I can use it from my Google doc
• When I select a paragraph, I can have Jasper review my wording

Being able to use GenAI directly from my Google documents saves me a massive amount of time. Instead of copy-pasting content in chatGPT, figuring out the right prompt, and copy-pasting back into my document, I can improve my Google Docs’ wording in seconds.

Interestingly, Notion also released a similar feature recently:

I also tried Grammarly, which is great to help you fix basic typos & grammar errors, even if my chatGPT prompts already found more grammar errors than Grammarly...

Yet, Jasper, Grammarly, or Notion are far from perfect solutions for our use case:

Proofreading:

• Jasper doesn’t have a proofreading option.
• Notion does, but you can’t control what it rewords, nor how.

Concise wording:

• Grammarly doesn’t?reword entire paragraphs
• Jasper’s or Notion’s do. But the (re)wording doesn’t “look like my writing”. Causing consistency issues across the document.
• They don’t tell me why they want to correct me. I cannot learn.
• Since the outputs from GenAI/chatGPT aren't perfect, I always need to compare my original version with Jasper's, to decide how to best improve it. It’s time-consuming, and a painful UX.

More importantly:

• None of these tools help improve my content clarity (eg. ??specify how many clients encounter the problem??)
• It doesn’t help my storytelling (ie. they don't assist in better structuring my thoughts)

We looked at this chart already, mapping the different written communication problems. Here is an illustration of where current solutions help, and fail:

Existing tools only scratch the surface of what could be done to improve employees’ communication skills and companies’ decision-making. Which is expected! These softwares focus on different problems to solve, and different audiences.

There is an opportunity to build a tailored product and a much better solution for our Tn$-niche!

Let’s deep-dive how to build such superior product, with

1. A better user experience
2. A better algorithm

1. Building a better user experience with AI-generated comments

Jasper’s target audience is content marketing teams (social posts, blog posts, etc.)

Marketing content writing is a very different process than writing strategic memos or roadmaps. They have different pain points, and require different solutions.

Marketing’s Jobs-to-be-done is to keep a community (ie. potential clients) engaged (ie. remembering your brand and product). To engage a community, Marketing needs to share new attractive content, frequently. Generating fresh content ideas daily can be challenging and time-consuming, making it difficult to maintain a consistent pace.

Jasper gives you always new ideas and writes them for you, saving Marketing teams hours every week. Just look at reviews:

Jasper's Product Market Fit is about generating content.

Google is also testing a new AI feature in Google Docs. Here again, to generate text from short prompts:

All email & copy generation Products (like Jasper, or Google) are doing the same: generating text from prompts.

However, let’s look at how we write Product briefs or Strategic memos to drive internal decisions:

• The writer already has insights (data, user feedbacks, features ideas, etc.). His goal is to convince a thoughtful and senior audience (managers, stakeholders, etc.).
• These unique insights require specific company knowledge, and cannot (yet) be invented by an AI…
• … these unique insights should represent 80% of the end-document’s content
• The writer’s goal is to properly explain the insights, and illustrate how they, when combined, suggest a course of action

An AI generating paragraphs out of short prompts cannot be the solution: in a strategic memo context, the difficulty is less about "generating text", and more about "articulating complex insights" together. If the writer doesn’t have great insights, an AI-generated memo will just be a long (but well-written!) bulllshit...

This is why none of the existing AI text generation solutions will work for our use case. Text generation from short prompts cannot be the solution.

On the other hand, I believe integrating AI-generated comments presents a more natural user experience:

• The writer drafts his thinking
• The AI suggests how to make them clearer through comments
• After a couple of iterations with the AI, the writer will land on clear, concise and well-articulated content

AI-generated comments would only replicate how a mentor works with a mentee in real-life: the mentor gives feedbacks, but wouldn't write everything for the mentee.

2. Building better algorithms

In addition to UX, the quality of algorithms is crucial. I experimented with chatGPT and achieved good results by using customized prompts: I gave examples of good comments to teach chatGPT what to do. Then, I selected a few actual documents written by my PMs to test the effectiveness of my prompts.

My proof of concept (PoC) significantly reduced the size of the texts:

My prompts also gave very good feedbacks to make the content clearer: I would have made most of the comments myself as VP Product:

For my team, incorporating such feedback before I even read would drastically improve the quality of their first draft.

For me as a manager, a (way) better document would save me hours of reviewing and allow me to focus on the content.

This is worth a lot! I would have been ok to pay for such service, perhaps somewhere in the 10€/month/PM range.

And this is only with an MVP-grade Product.

Beyond the MVP: custom-built models

First, AI-generated comments could potentially outperform my simplistic use of 'ChatGPT prompts'. Relying on ChatGPT as the foundational AI model may quickly reveal limitations. Paul Graham intuitively suggests:

ChatGPT responses are generic.

To make it very simple: the model learns from all sorts of documents, almost regardless of their quality. But the model output is only as good as the quality of the data it learns from! Since ChatGPT learns from a broad spectrum of data, its responses cannot excel in niche use cases like ours.

Even if you prompt ChatGPT with specific role-based instructions, such as 'you are the best CPO in the world, please comment on this document', the results, while interesting, will not be outstanding.

How can an AI effectively coach if it's merely average at providing feedback? It must be exceptional.

To solve similar problems, several companies invest hundreds of millions into their custom-built GenAI models.

Bloomberg is an interesting example of such trend. They are training their own LLM, BloomberGPT, using a mix of public documents (±50%) and their own financial documents (±50%). The goal is to outperform generic models like ChatGPT on Finance-specific tasks.

What if we tailored a model, by learning specifically on strategic documents or product briefs? Wouldn’t it give great feedback to teams all around the world, and help companies make better decisions?

Beyond improving the ??AI-comments algorithm??, boss.ai could also cover more use cases:

• give feedback about the data: the AI could comment graphs, how relevant are the insights, challenge data analysis…)
• improve the story telling: is there a clear summary? Are arguments in the right order? Are they redundant? …
• Managers could give boss.ai specific instruction to review in their teams’ documents (eg. ??always make sure it respects the company template??, ??I want documents to be no longer than 4 pages??, …)
• Etc.

This would require the development of custom AI models.

But it doesn't mean investing Billions into a completely new LLM! I believe this is achievable with a couple of engineers fine-tuning existing open-source pre-trained models.

Conclusion

We have a multi-trillion-dollar opportunity in front of us, by building the first employees’ AI-assistant to write better strategic memos, roadmaps, product briefs, or any internal document.?

With a simple MVP, I assume we can already test the Product Market Fit:

• a Chrome extension
• Leaving comments on Google docs, or Notion documents
• With an algorithm that proofreads and suggests ways to be more concise and clarify the content
• Using chatGPT API, and custom prompt engineering

It would take just a couple of days to build.

We can focus this MVP on a sub-segment, to limit the complexity:

• Product Managers are good candidates, as they write such memos every week
• Engineers are interesting as well: it’s a 5-10x larger segment than PMs, write a substantial amount of documentation, and their natural writing skills are low (hey, they are engineers!)

Distribute it for free to a small group of 10 to 20 early adopters and observe:

• Are they using it on a daily basis? If not, why?
• Are they sharing the tool with other colleagues?

Within about a month, we should know if it has the potential to become a product worth $10 to $100 million. But based on my research, and existing benchmarks, many being Bn$ business, I bet it will.

Who wants to try?