Business automation transfers a defined part of a workflow to software: receiving an event, checking information, applying a rule, moving a record, or asking someone to approve the next step. A dependable automation also knows what to do when the expected path fails.
Begin with work that repeats often, has identifiable inputs and outputs, and creates measurable friction. If every case requires a different judgment, automate the surrounding collection and routing first. Keeping a person at a decision point can be the right design.
Match the method to the work
On a small screen, scroll the table sideways to read all columns.
| Method | Useful for | Boundary to understand |
|---|---|---|
| Rules and scheduled jobs | Routing complete forms, reminders, and scheduled reconciliation. | Rules need explicit conditions and a response when data is missing. |
| APIs and connectors | Moving structured data between supported services. | Permissions, rate limits, versions, and connector licensing still matter. |
| Robotic process automation | Interacting with an application through its user interface. | Screen layouts, sessions, dialogs, and software updates can interrupt the flow. |
| AI-assisted steps | Extracting proposed fields from varied documents or drafting a summary. | Model output needs validation; plausible text is not proof that a field is correct. |
An event-driven cloud workflow typically starts with a trigger and then runs actions. Microsoft’s cloud-flow introduction demonstrates this structure. Use a direct supported interface where it fits; a user-interface robot remains an option when the application does not expose the required operation.
Worked example: route an invoice without paying it automatically
The following design is fictional. It prepares an invoice for review and accounting entry; payment authorization remains a separate controlled process.
- Receive and identify. Preserve the source document and assign an internal intake ID. Record when and through which approved channel it arrived.
- Extract proposed fields. Read supplier ID, invoice number, date, currency, totals, and purchase-order reference. Preserve the source alongside extracted values.
- Validate. Require essential fields and check arithmetic. Compare the supplier and purchase order with approved records. A new bank account in the invoice must not silently replace verified payment details.
- Route. Ask the appropriate budget owner to review the document and any discrepancies. Missing information goes to a named exception queue.
- Record the decision. Store who approved or rejected, when, and which document version they reviewed. Changed amounts require a fresh decision.
- Create and reconcile. Submit the approved record to the accounting system, keep its returned ID, and confirm that intake and accounting records agree.
For an invented invoice with a $200 subtotal and a $12 tax line, the stated total should be $212. This checks addition only: it does not establish that the tax is legally correct or that the goods were received. A document that totals correctly can still be a duplicate or fraudulent request.
Define a business duplicate check, such as supplier identity plus invoice number, and decide how to handle corrections and credit notes. That check is different from identifying repeated delivery of the same technical request.
Design the failure paths before the happy path goes live
A timeout can mean the destination accepted a request but the reply was lost. Repeating a create operation blindly can produce two records. Where supported, use an idempotency key that stays the same for retries of the same operation, and confirm the destination’s documented behavior. Stripe’s idempotency documentation is a concrete example; its rules and key lifetime are specific to Stripe.
On a small screen, scroll the table sideways to read all columns.
| Situation | Appropriate response | Evidence to retain |
|---|---|---|
| Temporary service failure | Retry within a bounded policy, with increasing delay where appropriate. | Attempt count, response, and next retry time. |
| Missing or invalid input | Stop that item and request correction. | The field and reason that failed validation. |
| Unknown result after a timeout | Query or reconcile before repeating an operation with side effects. | Stable request key and any destination record ID. |
| Approval overdue | Escalate to a designated backup reviewer. | Current owner, age, and escalation history. |
| Partial completion | Resume from a verified checkpoint or perform a documented compensating action. | Which steps actually completed, not just the overall run status. |
Microsoft’s Power Automate error-handling guidance covers scoped error handling, retry policies, termination, and notifications. Keep alerts actionable: identify the item, failure, owner, and recovery route. Repeated noise without ownership teaches people to ignore the queue.
There is no universal undo across multiple services. Deleting a newly created record may not reverse an email already sent or an order already dispatched. Limit credentials to necessary actions, and define a stop switch that pauses new intake while preserving in-progress records for reconciliation.
Separate released time from cash savings
These are teaching assumptions, not market prices or guaranteed savings. Salaried payroll may stay unchanged; the benefit can be more time for other work. Cash payback requires actual cash reductions or additional contribution, not merely multiplying minutes by a wage. Include testing, training, supervision, usage charges, and ongoing changes. If exception rates rise, recalculate before scaling.
Run a pilot you can stop and explain
Test representative cases in a sandbox: a normal document, missing currency, an arithmetic error, a duplicate delivery, a service timeout, and an approver who is unavailable. For AI extraction, compare each field with a human-checked reference and include unfamiliar layouts. An average success rate can hide a critical error in bank details or totals.
Then shadow a small real workflow without letting the automation commit consequential changes. Compare its proposed actions with the existing process. Authorize a limited live phase only after ownership, access, recovery, and reconciliation are clear.
Download the workflow pilot checklist (plain text) to record the process owner, test cases, stop procedure, and acceptance criteria. Revisit it whenever a connected system, business rule, or AI component changes.