This note is in response to a memo recently sent by a GP to the LPs of the fund he manages. GP argues that whereas SaaS margins had a halo effect until recently, AI is putting strong downward pressure on the price that can be charged due to competition from in house solutions and business competitors. He predicts that businesses that used to be privileged because of the mote of software development costs are now just businesses.
He provides some strategies for leaning into this new reality, such as introducing a professional services or atoms based component to the service: “use software to sell not-software”.
There are two other trends that are emerging that GP doesn’t mention: 1/the toolification of APIs; and, 2/the expertise mote.
The Expertise Mote
We’ve all heard that AIs hallucinate and the their intelligence is jagged. This is a fundamental property of LLMs that cannot be orchestrated away. The implication is that for the foreseeable there needs to be humans in the loop all the way down. GP was able to build Climbing Gym Management software for himself because he had both the technical expertise and the domain expertise as a gym owner and climber.
Whether nor not he is motivated to do this, he is in the position to sell Climbing Gym Management software to other gym owners given his dual expertise. Most existing SaaS solutions are built on this implicit or explicit domain expertise. It’s not that other business owners don’t have one or the other but having both appears to be a kind of mote.
The Toolification of APIs
Legacy SaaS was designed for human users. As those humans become empowered or replaced by agents, protocols such as MCP will become increasingly important to give the jagged intelligences access to deterministic, trusted and secure connections to tools that have access to systems of record and algorithm based resources.
Graffiticode as Smart MCP Server
MCP is an important first step to making SaaS APIs available to AI agents. What it doesn’t do is provide a layer of dedicated AI that is trained to use that tool. The burden in on the general AI (Claude or ChatGPT) to know how to use it with short descriptions and/or skills.
For each tool, Graffiticode provides a formal language that defines its capabilities and an dedicated AI trained to translate human and agent requests into that formal language. These language based MCP tools can be spun up in a matter of days for any SaaS API, complete with dedicated AI built with RAG and other edge training techniques.
The Transition to a AI / Human Hybrid SaaS Future
GP talks about how the viability of SaaS based businesses will include human services. Platforms such as Graffiticode allow the human service component to be built into the services in the formal language design (what is the shape of the solution space?) and training of the dedicated expertise (how do we translate user intent to experts solutions?).
It seems inevitable that every SaaS business that survives the so called SaaSpocalypse will need to serve both man and machine clients. There is clearly a short term advantage to moving fast into this reality with smart tools like those that can be built with Graffiticode.