Automation
Follow defined rules and steps.
Use for repeatable checks, updates and routing.
ExampleFlag a cost change and route a review.
Respond to changing costs, investigate exceptions and route actions for approval, with workflows and AI agents operating within your business rules.
Follow a cost update through checks, investigation and a proposed response, with approval at the required checkpoint.
Submitted is not approved, published, or recovered margin. Illustrative figures.
Follow defined rules and steps.
Use for repeatable checks, updates and routing.
ExampleFlag a cost change and route a review.
Investigate a goal using permitted tools.
Use when the work requires examining information and deciding what to inspect next.
ExampleExplain a margin decline with supporting evidence.
Add investigation inside a defined process.
Use when routine steps need an agent-assisted assessment before the next action.
ExampleAssess affected accounts before preparing a pricing response.
| Property | Automation | Agent | Combined workflow |
|---|---|---|---|
| Who chooses the path | You do, on the canvas, before it runs | The agent does, at runtime, a step at a time | You draw the path and hand one step over |
| Repeatability | Follows configured rules and branches | Reasons toward the goal, so the route can differ | Defined either side of the reasoning step |
| Starts on | A manual run, a system event, or a schedule | Somebody asking, or a step inside a workflow | Whatever starts the workflow |
| Runs on | Price lists, rule sets, datasets, views and the newsfeed | ||
| Governance | The same for all three: the creator’s permissions, your lifecycle policy, one run record per execution | ||
A condition identifies a subset of records rather than testing them one at a time, and a supported action then operates on that subset, so a flag across many items is one write rather than one per item. Two limits are worth knowing. A step that works on each member separately, such as moving a family of records through a lifecycle, takes on at most 500 in a run, because each member is its own round of reads and writes inside a single worker. A step that does one piece of work over the whole set is not capped that way. Record counts still affect how long the underlying queries take.
An agent investigation ends in something you can read and check, not a wall of tool calls. The calls are still there when you want them.
Margin fell 2.2 points. Landed cost accounts for nearly all of it, concentrated in three suppliers, and eleven customers carry most of the exposure.
| Cost | −2.4 | Landed cost, freight surcharge |
|---|---|---|
| Mix | +0.3 | Shift toward higher-margin lines |
| Price | −0.1 | Realized price broadly held |
| Net | −2.2 | Points of margin, Q1 to Q2 |
Saved as a note, citing the records above. A note is a draft artifact: nothing was published and no price moved.
Illustrative execution using supported tools. Every tool named below exists in the product; the run itself is a composition for this page, not a recording.
search_pricing_modelsreadNortheast price book, Q1 and Q2engine_query_dataread1,842 itemsdesign_new_analysisreadCost −2.4 pts, mix +0.3, price −0.1check_data_provenancereadLanded cost, freight surcharge, 3 suppliersengine_query_datareadTop 11 carry 68% of the exposurecreate_notedraftNote saved, nothing publishedCalculations are performed by the query and analysis tools, not by the language model. The model decides which tool to call and writes up what came back.
A workflow acts under its own name, with its creator’s authority. Roles are read fresh from the creator every run, so a permission you remove applies to the next run, and a workflow can never reach further than the person who built it. If that person loses their role, it refuses to run and records why.
Your lifecycle policy decides, not the workflow. A step that is assigned to named people cannot be performed by a workflow at all: it is refused, and the work waits for them. Steps nobody is assigned to run on the creator’s authority like any other action.
A run is a job, and cancelling the job stops it before it takes its next step. Work already completed stands, because stopping is not undoing. A run that dies with its worker is closed out as failed rather than left open.
Every run keeps its trigger, each effect with the step and record it touched, start and finish, and on a failure the step that raised and its error. Notes record why a run did less than it looks like it should have, which is how a run that reached nothing is told apart from one with nothing to do.
Four shapes that recur. Each is something you would build, not a template waiting in the product.
Built on a dataset or view that carries those dates. Automations run on price lists, rule sets, datasets, views and the newsfeed; there is no agreement or rebate entity they attach to directly.
Workflows and agents use Revomo’s shared records, supported actions and governance controls.
Walk through how Revomo can combine defined steps, agent-assisted investigation and approval controls around your commercial work.