Agentic & AI systems

What the assistant was allowed to know

An assistant in the team chat tool, launched against a dozen governed metrics and a knowledge base of what they mean, with every failure fed back as context. The narrow start was the decision. The request queue fell by roughly forty percent. The curation work it created is permanent.

• →Give the agent metrics as tools and context as documents, never raw tables and a prompt
• →Launch narrower than feels reasonable, and let the unanswered questions set the roadmap

When the assistant costs more than the queue

A launched assistant has a sponsor. A retired one has nobody, so retirement happens by neglect or not at all. Four numbers that say whether to widen, narrow or switch it off, the meeting that reads them, and what a good retirement keeps.

When the agent writes the pipeline, review is the platform

For twenty years a data team’s constraint was writing. When a model drafts the pipeline, the model and the query, the constraint moves to review: what a reviewer checks when the author is a model, which standards have to run rather than be remembered, and what the team stops practicing.

The dashboard stopped one step short

A dashboard shows that a number moved and leaves the reason to the reader. Business users now schedule AI briefings that supply the reason and a next step. What that changes for governance, for BI, and for the data team.

An analyst in the chat window

The obvious place for an AI analyst is where the questions already arrive. Putting one there is easy. What decided whether anyone used it a month later was the size of the first step, who was invited to break it, and what it did to the questions.

What breaks when agents meet a real warehouse

The demo works because the schema is clean. Production schemas carry a decade of history, three defensible definitions of revenue, and decisions that need more than a plausible query.