The SaaS products winning right now aren't the ones with the best features. They're the ones that can get your data in front of an AI model and do something useful with the result.
For most of its history, SaaS meant putting a decent UI over a database and charging monthly for it. That worked. Then it shifted toward workflow automation — connecting things, reducing manual steps, syncing data across tools. That worked too.
What's happening now is different. The products pulling away are the ones that act as a layer between your data and an AI model. They get context in, get an answer out, and route that answer somewhere your team actually sees it. That's the pattern. It has a boring name: an AI harness.
What a harness actually does
A harness isn't the model. It's the plumbing around the model. It handles getting your data clean and structured, deciding what context is relevant to send, calling the model, and then doing something with the response — sending an alert, updating a record, surfacing an insight.
The model itself is increasingly a commodity. GPT-4, Claude, Gemini — they're all capable enough for most business tasks. The differentiator is everything else: do you have the connectors to pull data from the right places? Is the data fresh? Does the model get enough context to give a useful answer? Does the output go somewhere that changes behavior?
Most companies trying to "add AI" skip straight to the model and wonder why it doesn't work well. It doesn't work because the hard part was never the model.
Where the value accumulates
Connectors are unglamorous but they compound. Building a reliable sync from QuickBooks or Sage Intacct takes time — not because it's technically hard, but because the edge cases never stop. Every customer has slightly different data, slightly different configs, slightly different expectations about what "current" means. Getting that right is a multi-year project, not a weekend one.
Permissions matter too. An AI that can see everything is a liability. The harness has to understand what data each user should be able to query, and enforce that before anything reaches the model.
This is where the moat actually forms. Not in the model choice, not in the chat interface. In the connectors, the data layer, and the logic that decides what gets sent where.
What this looks like at Brooked
Brooked connects to your financial data sources — Snowflake, QuickBooks, Sage Intacct — keeps that data current, and routes it into workflows: alerts when a metric moves, answers when someone asks a question, summaries when a report is due.
That's the harness pattern. The model doesn't know your chart of accounts or your close calendar. Brooked does. That's what makes the AI output useful rather than generic.
What to look for when evaluating tools
If you're buying SaaS that claims AI capabilities, ignore the demo prompts and ask about the data layer. What does it connect to? How often does it sync? What context does it actually send to the model? What happens with the output?
The best tools in this category will have spent more time on those questions than on the chat interface. The ones that haven't will feel impressive for about fifteen minutes before the answers get generic.