What data do we need before starting an AI agent project?

The common assumption is that you need a large, clean historical dataset before you can start. For agents, that is usually wrong — you need access and structure far more than you need volume.

What you actually need

Three things. Reachable systems: the data the agent will act on has to be available over an API or a database connection, not locked in a desktop application. Stable identifiers: the agent needs to know that a customer in one system is the same customer in another. Documented process: someone has to be able to describe the task the way they would train a new joiner.

What you do not need on day one

You do not need a data warehouse, a labelled training set, or a fully cleaned CRM. Agents work against your systems live; they are not trained on your history in the way a custom model would be.

The readiness checklist

Pick one process. Confirm every system it touches has an API. Confirm you can create a service account in each. Write the process out as steps, including the exceptions people handle informally. Identify which steps are irreversible. If you can complete that in a week, the project is ready to scope.

The thing that stalls projects

Not data quality — access approval. Getting service accounts provisioned across four systems in a large organisation routinely takes longer than building the agent. Start that request on day one, in parallel with discovery.

Where messy data does bite

Duplicate records. If the same customer exists three times with slightly different names, the agent will act on one of them and appear to have done nothing. Deduplicate the entities the agent will touch, and only those.