Workflow guide
How to build your first business AI workflow
Choose a recurring task with a known input and a result someone can check. In Sontairo, define the trigger, data, AI step, and action separately, resolve setup issues, then review the first runs before adding more responsibility.
Sontairo · Updated
1. Choose a task with a clear finish
A good first workflow has an input you already trust and a result with an obvious owner. For example, a weekly operations briefing can use a defined set of records and produce a summary for one team. “Run the business” leaves too many decisions unspecified to verify a result.
Write down what starts the work, which information it needs, what a successful output contains, and who handles an exception. If you cannot describe those four things, narrow the task before building it.
2. Choose an agent, automation, or Concierge
| Your task | Start with | Why |
|---|---|---|
| Work you direct in a conversation | An AI agent | Instructions and connected tools for interactive work. |
| A repeatable process with a trigger | An automation | Configured inputs, logic, actions, and run history. |
| Questions from website visitors | Concierge | Business content, a visitor conversation, and a next step. |
Use BizOps when the process needs structured customer, job, or pipeline records. Agree on field meanings before asking AI to interpret them.
3. Separate information, reasoning, and actions
For a weekly briefing, start with a schedule, add configured components that gather the required information, then pass those results to an AI component with a fixed output format. Use an explicit notification component to deliver the result.
Example workflow
Schedule → Gather defined inputs → AI summary → Team notification
Example output: completed work, blocked items, missing data, and links to the input records. This is a suggested workflow design, not a preconfigured integration or a measured customer result.
Unattended AI components cannot ask follow-up questions or discover and call tools. Supply the necessary information up front. Connector, webhook, and notification components carry out actions, making the intended behavior visible in the workflow.
4. Check the setup before activating
Connect success and error paths, complete the required fields, and select Check setup. Resolve missing permissions, connectors, and schedule details. Check the previewed run times in the selected timezone.
Use a low-impact first output, such as an internal briefing, while you evaluate the process. Compare its statements with the actual inputs and inspect how it handles an empty dataset or missing field. Decide what should happen on failure rather than assuming that every component succeeds.
5. Review runs and publish deliberate changes
Activation publishes an immutable workflow version. Subsequent edits remain a draft until you publish changes. This lets you refine instructions or components without silently changing the version that is already running.
Run history records component status, duration, inputs, and outputs. Inspect where a run failed, whether the summary used the expected data, and whether its notification reached the intended destination. Repeated failures can pause a workflow after its configured retry threshold.
Decide whether the workflow is worth expanding
Track completed runs, exceptions requiring a person, corrections to the output, and time spent reviewing it. Compare those results with the same task done manually. Expand only when the output is useful and the review effort makes sense for your team.
A workflow that produces a plausible summary of incomplete inputs still needs fixing. Improve the input and failure handling before adding more agents or more actions.