As of 13 August 2026, AI can only partly automate your business workflows.

What to actually do

1. Hand it to a person

   The route this page recommends

   A person who owns the outcome does this end to end, worth it when the failure is dear.

2. Use a tool built for this

   Second choice

3. Do it yourself

   The distant third

   A chat interface, power-user skill, and roughly 1 hour until you can act on the result.

   How to actually do it

   1. Open a document and write the current workflow from trigger to final outcome, naming the responsible role, the systems used and the exceptions that people handle manually.
   2. Gather three successful examples and examples of failed, duplicated or incomplete cases, removing unnecessary personal data before pasting them into the chatbot with the prompt.
   3. Open the documentation or administration pages for each application involved and record the available triggers, actions, fields, permissions and test-environment options.
   4. Paste the workflow, anonymised examples and confirmed application details into the chatbot, then ask it to produce the workflow map, configuration and test plan requested in the prompt.
   5. Create the automation in a test environment or with copied records, using restricted permissions and approval steps for external messages, payments, deletions and other irreversible actions.
   6. Run every normal and failure case from the test plan, compare each resulting record and message with the expected result, and record any mismatch in the workflow document.
   7. Ask a colleague who understands the process to check the rules, permissions, exception handling and rollback plan, then correct the configuration and repeat failed tests.
   8. Enable the workflow for a small, observable batch, inspect its logs and outputs, and only then move it to routine use with an owner and a defined manual fallback.

   Prompt

   Act as a business-process analyst and automation engineer. Help me design a safe, maintainable automation for this workflow:
   
   Business goal: [describe the outcome]
   Current process: [list every step in order, including who performs it]
   Systems involved: [name each application and what it stores or does]
   Trigger: [what starts the workflow]
   Inputs: [fields, files, messages or events received]
   Rules: [conditions, approvals, exclusions and deadlines]
   Output actions: [what must be created, updated, sent or logged]
   People responsible: [roles, not personal data]
   Known examples: [include at least three successful examples and any failure examples]
   Constraints: [permissions, privacy requirements, retention rules, volume and budget]
   
   Produce:
   1. A plain-English workflow map with the trigger, each action, each decision and the owner.
   2. A table of required fields, their source, their destination and whether each is mandatory.
   3. The simplest implementation plan using existing application features where possible.
   4. Any code, formulas or configuration needed, clearly labelled and separated from explanation.
   5. A list of assumptions and unanswered questions. Do not invent application features, credentials, business rules or customer information. Mark anything that needs confirmation as NEEDS CONFIRMATION.
   6. A test plan covering normal cases, missing data, duplicates, permissions failures, timeouts and incorrect inputs.
   7. A rollback plan and an audit-log specification.
   8. A staged launch plan that starts with a copy or test environment and does not send messages, alter records or delete data until a human has approved the test results.
   
   Keep actions idempotent where possible, minimise personal data, request approval before irreversible or external actions, and explain exactly what I must configure manually. Do not claim that the workflow is production-ready until every NEEDS CONFIRMATION item and test has been resolved.

   Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.