Make things people want (2004-2024)
“Make things people want” has been the startup world’s credo for almost 20 years. For good reasons: despite the formidable progress in simplicity generated by cloud hosting (AWS was created in 2004) and modern languages (Ruby on Rails in 2004), which made it possible for YC (2005)’s kids to fight against the old giants, software remained long and expensive to build. The efforts towards its creation still had to be mutualized among its users.
Those were the SaaS golden years: Write the code once, and rent it to the people for whom it works despite the fact it wasn’t tailor-made for them.
Making custom software perfectly fitted to each user’s need was just not an option, so people with niche needs were condemned to good enough solutions. Often modulo higher marketing, sales, implementation and customer success costs.
The quest for custom software
Creating software easily by just asking a robot, thus unlocking custom software, has been a dream all over this period, for us included. It was the promise of many “no-code” platforms but none of them really managed to get steam, and the various “AI on code” attempts failed (cf. our investment in source{d} in 2016). The technology just wasn’t there.
It all changed in 2024: GPT, Claude, Poolside and the others are so good that they already make software engineers 30%+ more efficient in many companies. Amazon said they already reclaimed the equivalent of 4500 human years of developer time.
Everyone in the industry is convinced the last human-written line of code will happen in a matter of years; in the future, you will “just” have to ask AI to build the software you want.
What the hell is an Agent
AI-native softwares will quickly get 10x better than SaaS, for 2 reasons: First, the high level of customization they allow make them better for normal people, and much better for people with niche needs. Second they natively deal with AI (call OpenAI APIs, basically), which gives them an immense advantage.
One could think that traditional SaaS also have access to the APIs of OpenAI and its peers, and it’s true. But it would be ignoring the high and growing complexity involved in calling the right model in the right way, to maximize the smartness of the answer while limiting the cost of the query.
When accessing GPT o1 through the API instead of ChatGPT, you can decide how many dollars you want to give the system to produce your answer (it was already possible to structure chains of thoughts by yourself, basically asking "try harder" to any LLM and paying for it, but GPT o1 makes it native).
As a result, some people in high-stakes industries like pharma happily pay $1000+ per question (btw if you are working on a ChatGPT for rich people please ping me!)
The job of formulating the request, sizing the optimal spend, then helping the model to make the most of the dollars it can use is a real one. One which traditional SaaS have no clue about, and will need to learn while rebuilding their whole products from scratch.
It’s essential that those modern pieces of software mix traditional (even if AI-generated) code to AI API calls, and don’t just ask everything to GPT & co. In part because the real world is complicated and many things are still better managed through well-thought interfaces and databases, but fundamentally because powering AI is expensive. There simply isn’t enough electricity and GPU power on earth to handle every single request with GPT, so it makes sense to save it for what it does really well (languages generation and interpretation, mainly).
Those modern pieces of AI-native custom software are commonly called Agents. In this taxonomy, complex systems imbricating many of them towards a goal are called Agentic Applications.
It’s not that simple
It would be easy to deduct that the software industry is dying. That everyone will churn from the Salesforce and Workday of the world to craft the Agents of their dreams on top of a Snowflake-like database, as Klarna famously announced it was doing.
The software industry as we know it is probably dying, but anyone who has ever been in a relationship knows that “What do you want?” is not a trivial question. It depends on your technical constraints, social context, and knowledge.
As Elon said: “One of the really tough things is figuring out what questions to ask. Once you figure out the question, then the answer is relatively easy.”
To design Agents, you need Architects
Figuring out which precise Agents the user wants to ask AI to bring to existence and how to make them play together to form Agentic Applications is not an easy job. We’re currently seeing a whole new software category gaining steam by doing precisely that.
It doesn’t have a name yet, so I propose Architects.
Architects’ fundamental value proposition to customers is to do these 3 things:
- Collect data in an AI-native way, dramatically increasing its quality and quantity
- Make sense of this data to take stellar decisions
- Design and track how these data and decisions are propagated in your whole information system and whole organization
Under the hood, Architects do this helping their customers to design Agents (custom mix of traditional code and AI API calls) and have AI code them. Designing the right Agents couldn’t be done well without 1) the AI-native data collection, which provides plenty of information on who the customer is and 2) some sort of focus on a topic (“vertical”), which allows the Architect’s own Agents to learn from all of its customers and get better over time.
We are lucky enough to have invested in several Architects recently, among them Attention and Maki:
Attention helps sales teams to automate work out of their client interactions (calls and emails, human or robot-led) by extracting insights and suggesting the best workflows to push it to the right person, at the right place, and at the right time.
Example: During a discovery call, if the prospect says a specific integration is important to him, add it to both Linear and Salesforce and create an alert that will pop in Slack when the feature is ready.
Maki helps HR teams to design AI-native candidate experiences (can involve talking to a robot at your own schedule, passing custom-made tests to detect specific abilities…), fine-tune its algo trained on millions of candidate answers with his own internal data to take the best hiring decisions, then design and track the candidate journey through the client information system and whole organization.
Example: After a robot-led screening video interview which gave the candidate a taste of the company culture, decide if the candidate moves to next step and update Workday.
Architects are silent SaaS-killers. And winner takes all
From a business standpoints Architects have 2 very exciting traits:
Data Network Effects: The more customers they have and data they see (some of it would not even exist without them), the more their own algorithms are powerful and able to recommend the right decisions and Agents to build.
Incremental Expansion: They silently grow their perimeters until the day the old System of Records (Salesforce, Workday…) are obsolete and don’t make sense anymore vs a simple database. Giants everyone thought were impossible to displace will go through an inexorable chinese torture and end up obliterated.
It’s anybody’s guess to know how all this will impact the tech industry size, but it will probably end up an order of magnitude bigger, since more people will be able to benefit from it more intensely. Relatively speaking, the layer of the cake captured by Architects will likely be slimmer than legacy SaaS, but it will be part of a much, much bigger cake.
The rest of the cake will largely be enjoyed by large AI model makers such as Poolside or OpenAI. And humanity as a whole.