1. List the work that repeats
Choose several tasks people already perform. For each, note what triggers it, which information arrives, the steps taken, the tools involved and the expected result. Ask the person doing the work to explain the exceptions as well as the normal path.
Separate waiting from effort. A task may take several days to finish because it waits for approval, even if the hands-on work takes minutes. Automating the drafting step will not remove the approval delay.
2. Ask whether AI is needed
Stable rules and structured inputs often fit conventional automation. For example, copying an approved form field into a matching record may not need a language model. Reading an unstructured request and proposing a category may benefit from one, provided the result can be checked.
An assistant can help a person find or draft information. A tool-connected agent can also take actions. Each additional action changes the permissions, failure cases and testing the project needs. Start with the least autonomy that can deliver the outcome.
3. Compare the candidates
- Frequency
- How often does the task happen? Infrequent work may not justify the setup and maintenance.
- Input quality
- Are the necessary facts available, current and permitted for use in the selected tools? Poor source information does not become reliable because a model reads it.
- Exceptions
- How many cases require judgement, missing information or a different route? Write down the common ones before designing the workflow.
- Review effort
- How long does it take to check the result? A fast draft is of limited use if verifying it takes longer than doing the task.
- Failure cost
- Can the result be corrected before it affects a customer, record or payment? Drafting for review and acting externally are different risk levels.
- Ownership
- Who will maintain the connections, update instructions and deal with failures? A workflow needs an owner after launch.
Use these criteria to discuss the trade-offs, not to manufacture a precise score from uncertain estimates. The best first project is often the one with clear boundaries and a visible result.
Use the workflow workload calculator to compare current effort with review, exceptions and upkeep. It is a planning estimate to test in a pilot.
4. Define the pilot around one outcome
Consider an incoming-enquiry workflow. A limited pilot could read an approved example set, suggest a service category and prepare a draft summary for a person to review. It would not send replies or update live customer records unless those actions were explicitly added to the scope.
This is an illustrative workflow, not a client result. Its usefulness would need to be measured against the current process, including the effort of checking the category and summary.
- Input: the approved information the workflow receives.
- Output: the exact item it produces, in a defined format.
- Permissions: the tools and records it can access.
- Approval: which decisions remain with a person.
- Exceptions: when it stops and requests help.
- Fallback: how the existing process continues if the system is unavailable.
5. Agree on the evaluation before building
Collect representative cases, including messy inputs and known edge cases. Define what counts as correct and who will judge it. Test examples should cover the actual variety of the work, not only the easiest demonstrations.
Record the current effort first. After the pilot, compare completion quality, review time, exceptions and running costs. A useful result may be better consistency or easier retrieval, not only minutes saved.
Task and responsible person: Trigger and input: Current steps and tools: Frequency and hands-on effort: Common exceptions: Allowed data and permissions: Expected output: Human approval points: Examples used for testing: Success and failure criteria: Software and usage costs: Fallback and maintenance owner:
6. Expand only after the first workflow is understood
A successful demonstration is not the end of the project. Review early use, examine failures and check whether the information and integrations remain current. Add further steps when there is a reason to do so and the owner can support them.
Pax Collective can help with the workflow review, AI assistant or automation implementation. Send a description of your current process and the tools involved; omit sensitive records from the initial enquiry.