The retrofit problem
Teams asking to add AI to an existing product usually discover the blocker is not model capability but evidence: the system knows its current state and nothing about how it got there. You cannot automate a judgment you have never recorded a human making.
The fix is cheap during construction and expensive afterwards, which makes it worth doing before there is any AI on the roadmap at all.
Log decisions, not just state
Most systems store the result: the ticket is closed, the order is approved, the lead is qualified. Store the decision as an event instead — who or what decided, when, on what inputs, and what the alternatives were. That event stream is the training and evaluation data for every future AI feature in the product.
Capture the inputs the human saw
A decision without its context is unusable for automation. If an agent approved a refund, the record needs the account history and the message they were reading, not just the outcome. Reconstructing that later from timestamps across four tables is the work nobody budgets for.
You can’t automate a judgment you never recorded anyone making.
Make overrides first-class
When a person changes what the system suggested, treat that as a meaningful, structured event rather than an ordinary update. Overrides are the highest-signal data any product generates, and they are the mechanism by which an AI feature improves after launch.
Keep an addressable text layer
The unstructured material — notes, messages, documents, descriptions — needs stable identifiers and clean retrieval, not just storage in a blob column. Retrieval quality later is mostly determined by whether this was treated as data or as exhaust.
Version everything that shapes behaviour
Policies, prompts, thresholds, and rules should be versioned with an effective date, so that a decision made in March can be explained by what the rules were in March. Without it, every audit and every regression investigation becomes archaeology.
None of this is AI work. It is ordinary good instrumentation that happens to be the exact substrate AI features need, which is why the systems easiest to add AI to are usually just the well-built ones.