Agentic analytics

An agent that answers in the channel where the question was asked is a different product from a chatbot on a warehouse. The first is useful when it cites a contract. The second is a faster way to be confidently wrong. Agentic analytics is the first one: a plain-language question in, a sourced answer out, a person still accountable for the interpretation.

What has to be true first

• Governed data underneath, so the source has an owner and a purpose.
• A semantic contract for every metric the agent is allowed to say out loud.
• A refusal list: the questions it must not answer, written from last month’s real meetings.
• A named person who signs the answer. The agent does not own the number.

The precondition is its own essay. The permission list — what the agent may answer, and what it must refuse — is the operational version. Write the refusal list before the pilot. If you cannot fill it, you are not ready to fill the other one.

The failure that gets blamed on adoption

A six-week pilot, a recorded demo, a handful of power users, then a slow return to the dashboard that at least had a person behind it. The post-mortem says people did not change their habits. The actual cause is upstream. The agent was asked to resolve a meaning it was never given, so the room stopped trusting the paragraph.

Skipping to the agent is the current fashion because the demo is always impressive. Layer one and layer two can be a shared document and a named owner. They are slower to film and they are the part that keeps the third layer from becoming another dashboard with a personality.

Where the answer has to show up

Slack, Teams, email, the Monday pack. If people have to go somewhere else to be right, they will keep being wrong in the place they already work. Delivery is part of the definition, not a change-management phase you schedule after the model is chosen.