Analytics playbooks overview
An analytics playbook is a short, AI-generated document of data-backed blocks that is read as a whole. It is the report counterpart to a chat answer.
Background
A normal chat answer to "how is revenue doing?" is a message, perhaps with one table or chart. That fits a single question. It does not fit a request like "review Q3 sales against Q2 by region", which needs several numbers, a breakdown, and a takeaway that read together.
A playbook answers that kind of request in one step. The assistant plans a set of blocks, builds them from your semantic layer, and adds a conclusion drawn from them.
What a playbook contains
A playbook has a title and a list of blocks. Each block is one thing worth showing.
| Block | What it shows |
|---|---|
| KPI | A handful of headline numbers. Each tile shows the current value and, when there is a comparison, its change against it, such as against the previous period. |
| Query | The result of one query, as a grid or as a chart. A reader can switch between the two in the preview panel. |
| Conclusion | An AI analysis written from the other blocks. It is always last and counts as a block. |
Blocks follow reading order: headline numbers first, then detail. A playbook usually has a handful of blocks. It says a few things well instead of covering everything.
A request such as "revenue, and revenue by city" becomes two blocks: a KPI block for the total and a query block for the breakdown, over the same period.
How it works
A playbook is not a separate engine. It is the same AI chat, with one extra capability: building a whole document in one step.
- You turn on Playbook mode and send a request. The mode is permission, not an order. The assistant builds a playbook only if the request clearly asks for an analysis or report.
- The assistant reads your semantic model to learn what you can measure, how it can be broken down, and how entities relate. It asks you only what the model and the conversation cannot answer.
- It plans every block, queries included. Each block is checked as it is added.
- A second pass reviews the plan against everything you asked for. Its objections go back to the planner as advice. After two objections the plan goes ahead.
- Forge shows you the plan in a Review Query card. You confirm it, revise it, or reject it. Guide skips this step.
- The blocks run, and the playbook fills in on screen.
- A conclusion is written from the finished blocks.
Governed data only
Every block is a semantic layer query. A playbook goes through the same permission checks as the rest of the chat. A user limited to the semantic layer only ever sees semantic layer data in a playbook.
Periods are stated in full
A playbook is a document. It can be read weeks after it was made, so a relative period such as "last quarter" is resolved to fixed dates when the plan is made. The block's title and the description behind its info icon state the period it covers, for example "calendar year 2023 (1 Jan to 31 Dec 2023)". The playbook header shows the date and time it ran.
Numbers are not editable
A KPI tile's direction arrow comes from comparing the current and comparison values. Its color combines that direction with the metric's goal, for example "higher is better". A metric with no goal is shown in a neutral color. A reader cannot flip an arrow, because a shared document should not allow a fact to be overridden.
Forge and Guide
Both assistants can build playbooks. They differ in how much of the working they show.
| Forge | Guide | |
|---|---|---|
| Audience | Technical | Non-technical |
| Plan shown before running | Yes, including every query | No, it runs as planned |
| Queries visible in the playbook | Yes | No |
A technical reader can judge a plan by reading its queries, so Forge asks for agreement once before anything runs. A non-technical reader has nothing to check a query against. Guide therefore runs the plan and never shows the query, whether in the chat, in the saved chat, or in a chat shared with someone else.
See Forge and Guide explained for the full comparison.
When a playbook is the wrong tool
- A single figure, table, or chart. The assistant answers with that block directly.
- A vague request. The assistant asks what you want.
- A follow-up to an earlier answer. It is answered in the conversation.
Without Playbook mode, a request for a report or an analysis is answered with single blocks, such as a KPI, a chart, or a table. If the assistant thinks a full document would serve better, it offers to build one instead of building it unasked.
Failure behavior
A run can end in three ways:
- Ready - every block was built.
- Partial - some blocks failed and others answered. The playbook is shown, and each failed block shows an Error status and the reason.
- Error - nothing could be built.
The assistant does not retry a failed playbook on its own. It explains what went wrong and what you could change, such as a narrower question, the blocks you want, or the period.
Related concepts
- AI in Coginiti overview - the full map of AI features
- AI data and privacy - what data the assistant sends, and where, when it runs queries
- Forge and Guide explained - what each assistant can do and how action approvals work
- How to create an analytics playbook - step-by-step instructions for building a playbook in the AI chat
For more information about other exciting Coginiti features visit our documentation