The model is rarely the hardest part
Putting AI into an established software product is hard for reasons that have little to do with the model. What eight years of production work taught me about Brownfield AI.
When we started Datics AI in 2018, our work was centered on AI and data science. Eight years later, one lesson keeps coming back to me: the model is rarely the hardest part of putting AI into an established software product.
We learned that gradually.
Where the difficulty actually lives
As our work moved deeper into production software, the environment changed. We were no longer working with isolated datasets or clean-slate applications. We were working inside products that already had customers, workflows, permissions, integrations, business logic, data models, and years of engineering decisions behind them.
A pattern kept repeating. The AI itself could work. The prototype could look impressive. Getting it into the actual product was where things became difficult, because the questions changed:
- Can the AI access the right context without exposing the wrong data?
- Can it respect the same permissions and tenant boundaries as the rest of the product?
- Can it take part in the workflow, instead of sitting beside it as another chatbot?
- Can its actions be reviewed, approved, audited, and evaluated?
- Can the team build the next AI capability without starting from scratch again?
That changed the way I think about AI product development. The hard part begins when AI meets the product that already exists.
Brownfield AI
We started thinking less about isolated AI features and more about what it takes to make an established product progressively more intelligent. We now call that focus Brownfield AI:
- Start with the software that already works.
- Preserve what should stay.
- Change what the new capability actually requires.
- Build AI into the workflows, data, permissions, and architecture customers already depend on.
Why this matters for SaaS
I believe a large part of the next phase of SaaS will happen this way. Established products will not suddenly disappear and get rebuilt around AI. They will evolve.
The companies that figure out how to make that evolution repeatable will end up with very different products a few years from now. Not software with AI added to it, but software that has become substantially more capable because of AI.