From our SaaS audits
What we find inside SaaS companies
Patterns from the technical due diligences we run for investors and boards.
AI
Three Codexes walk into a codebase: The fellowship of the god class
Handing an AI your engineering conventions cleans up the code without fixing the architecture. A 5,000-line god class shrank to 500 lines across tidy domain files, and an independent reviewing agent still returned FAIL: authority never left the central object.
How to become an AI-native SaaS company in the next 3 years
Becoming AI-native in three years does not need a three-year technology roadmap. It needs a usage policy, a map of your data, one bounded experiment, and both the enthusiasts and the sceptics in the room. Build the ability to change, not the overhaul.
AI chat is a brain in a jar
Hitting a wall with AI usually means hitting the wall of the chat box, not the wall of AI itself. Give a model memory, tool access and a browser, and the interfaces we built for humans become the integration. The bottleneck stops being the model and starts being your imagination.
Your second brain knows more than you do
An AI-assisted note system captures faster than you can understand, and the gap compounds into internalisation debt: a dense graph of links attached to a thin mental model. The test of a knowledge base is not how much it holds, but how much you could still explain with the file closed.
Curl says it's fast. Your users disagree.
Synthetic checks like curl and Lighthouse measure a request your users never make. A HAR file captures the real logged-in flow, and handed to Claude alongside the codebase it turns a wall of JSON into a ranked list of fixes that get implemented in the same conversation.
A back-end engineer can do what now?
A back-end engineer set out to redesign a React app with an AI agent and came out actually knowing React. The unlock wasn't the prompting, it was understanding the code well enough to steer it. As the cost of entry collapses, knowing the technology matters more, not less.
Ask Claude what you're not tracking
Analytics only shows the events you remembered to add, never the ones you forgot. Let Claude instrument features as it builds them, then run a skill that audits your tracking against your own docs and live data, flagging what's missing, noisy, or quietly broken.
The integration loop
Integrating a tool with Claude rarely needs an MCP server. A boring script wrapped as a named, version-controlled skill does the job, and the real value shows up when two skills combine: one tool tells you what happened, another tells you why.
The strangest hobby at the table
Writing code by hand used to be the job. Soon it will be a hobby, like woodworking or growing your own veg. What a company pays for now is judgment: knowing what to build, catching the answer that looks right and isn't, deciding when it's good enough to ship.
Your company needs more than an LLM subscription
Handing everyone an LLM subscription gives them a model. A shared, self-hosted agent gives them a configured environment with company context, connected tools, and clear boundaries. Why a company should own that layer rather than rent it, and how to start small.
Everyone has a plan until they get punched in the mouth
Coding agents made implementation cheap, but validation never got cheaper. That shift moves the bottleneck from writing code to deciding what is worth building, and the response is more autonomy and faster iteration, not more detailed upfront plans.
The executable code style guide
A written style guide is a museum piece nobody enforces. An LLM changes that: it reads rules in plain prose and applies them on every pull request, reaching judgment calls no linter could express. The style guide stops being a document and becomes a runtime.
My email agent invented a prompt injection, then fell for it
An autonomous email agent hit a missing script, spiralled through 25 pointless shell calls, then fabricated email content including a prompt injection, and acted on it. The fix is not more warnings. It is structural validation before the model ever sees the data.
This tool is useless
Hand two developers the same tool and you get opposite verdicts. 'Useless' is rarely a fact about the thing. It's a fact about the person holding it. The rare skill was never operating the tool. It's imagining the door a tool opens before the use is obvious.
Documentation has a new reader and why it should belong inside the codebase
AI is now a reader of your documentation. When context lives outside the repository, gaps appear and assumptions creep in. Moving documentation closer to the code makes the system easier to reason about, for people and for machines.