What are AI Playbooks?
A Fusion AI Playbook is a written description of one job, such as checking this morning's forward curves, onboarding a new dataset or building the month-end market summary, that Fusion AI can carry out for you step by step, the same way every time.
You write the playbook once, in markdown. It says:
- What the person running it fills in: the curves, the dataset, the month, the people to email
- The steps, in order: what each step must do, in plain language, and the tools it may use
- The checks each step's result must pass before the run moves on
- Where a person approves before anything on the platform changes
- What the run hands back: a report, a script, a summary, a count
Fusion turns the playbook into a simple form. Anyone on your team fills it in and presses Run, and Fusion works through the steps on your platform data, exactly as written, keeping a full record of every step, check and approval.
Why use a playbook rather than a chat?​
A chat with Fusion AI is ideal for questions and one-off work. Some jobs, though, are repeated again and again and have a right way to be done. Asking for them in a chat each time has drawbacks:
| In a chat | In a playbook |
|---|---|
| Ask twice and you may get two different approaches | The playbook sets the order of the steps and the tools each step may use |
| The quality depends on how well the question is asked | The instructions are written once by the person who knows the job best |
| Nothing checks the answer unless you do | Each step is checked against rules you write, and retried if it falls short |
| Changes happen when you confirm them in the conversation | The run stops at an approval gate, and named approvers are emailed |
| The record is the conversation | Every run records each step's result, every check, every approval and the tokens used |
| You need to be there | Runs can carry on in the background or run on a schedule |
How a playbook runs​
When you run a playbook, Fusion follows it one step at a time:
- Focused steps: Fusion works only on the current step, with only that step's tools plus a few read-only ones. A step that reads curves cannot save a script.
- Structured results: each step records the values it was asked to produce, such as a script, a table or a count. Later steps and checks use these values.
- Checks: rules such as "every curve was found" or "the script validates" are verified after the step. If a check fails, Fusion tries again and is told why, jumps to another step, or the run stops.
- Transitions: rules such as "if nothing is wrong, end here" choose the next step.
- Gates: the run pauses for a person to approve or reject, typically before something is saved, created, run or sent.
- Budgets: each run has a token cap, and runs respect your tenant's Fusion AI budget and limit.
Where you find playbooks​
Playbooks live in the Fusion AI extension in the portal:
| Where | What you do there |
|---|---|
| Playbooks tab, Library | Browse the OpenDataDSL library of ready-made playbooks. Run one as it is, or copy it to make it your own. |
| Playbooks tab, My playbooks | Write and edit your tenant's own playbooks, including any added by an extension. Draft a playbook from a description with Fusion. |
| Tasks tab | Fill in a task from a playbook and run it, in the page or in the background, or give it a schedule. Approve or reject runs waiting at a gate. |
| Studio (the chat) | Ask Fusion to run a playbook. It fills in a task from the conversation and shows it as a card you can run. |
| Fusion AI Playbook Runs insight | See runs by playbook and user, success rates, failures and their reasons, and tokens per run. |
Key terms​
| Term | Meaning |
|---|---|
| Playbook | The markdown document describing the job. Every save creates a new version. |
| Task | One use of a playbook: the values filled in on the form, tied to the playbook version it was created from. |
| Run | The execution of a task: the state of each step, its results, checks, approvals and tokens. One task has one run; clone a task to run it again. |
| Step | One unit of work in the playbook, carried out by a Fusion assistant with a set of tools. |
| Check | A rule verified after a step completes. |
| Gate | An approval point after a step. |
| Transition | A rule that chooses which step comes next. |
| Output | A value the run hands back when it finishes. |
Next steps​
- Creating playbooks: the ways to write a playbook and the editor
- Playbook syntax: the full reference for the playbook format
- Running playbook tasks: forms, background runs, schedules and approvals