The problem with traditional coding is that everything must be rigidly defined. If you want to make a report that displays various types of data, you have to identify what those data points should be, where they originate, and how to display them.
AI flips this paradigm completely by letting you loosely identify a problem and then try on solutions. Simply state in plain English what kind of report you're looking for. If your harness can access the data needed for this, it will. It can then guesstimate which data is relevant, and, regardless of how it was formatted originally, convert it into data points and throw together sample visualizations. You then provide English feedback if needed and iterate until you have the report you want.
When you see something like this happen for the first time it literally feels like magic if you've ever spent considerable time coding pre-AI. Perhaps we should break it down a bit further:
- Can access the data: this translates to data that resides in a programmatically accessible system: CRMs, ECMs, project tracking, database, your file system, share drive, email over MCP. If you can write code to access the system, first teach your harness how to do this and record that in memory or via a skill. Without this step your plan fails before it gets started.
- Data format: I say this is largely irrelevant because it is. Pre-AI it was hugely important how you manage data. Does it live in siloed one-off documents without metadata or a well-structured DB with rules, triggers, and maintenance? The former was nigh impossible to derive useful reports from until AI. A file is a package of data. AI knows the package structure and can open it to extract relevant data. You still need to be able to identify roughly where your data lives (see 1), but with enough tokens you could ask your harness to build that robust reporting database from your loosely unstructured files. Iterating at this step is the most expensive depending on volume and ability to filter/sort/search your data. Recommend sampling various categories of data until you find what you need. Give an example to lead it in the right direction. It's not literally magic, but if you give a rough sketch, it can fill in a surprising amount.
- Visualizations: this part is nearly free. You have your data from 2 and now AI can take guesses at how you want to see it to be useful.
Some people may shy away from 1. If I give it access to my data, bad things can happen. This is true in theory but unlikely in practice, and you can make it impossible if you prefer:
- Only grant read access: most systems support some form of permissions whether it be an actual user you act as or a scoped API token.
- Development environment: prove out your implementation somewhere expendable where a mistake doesn't impact real data before asking to replicate where the real data lives.
- The auto-classifier: many harnesses have a sub-agent that reviews what they are doing, automatically denying any risky behavior. You can configure this behavior to help make it more accurate. Once you have completed 1, ask your harness to help you set up auto-approval for your read-only system access and denies for writes. In my experience this works remarkably well.
Any combination of the above may be useful depending on your case.