Start with repetition, not novelty
List activities people repeat every day or week: assigning enquiries, categorizing forms, reconciling records or summarizing long requests. Note how frequently the task occurs, how often it is corrected and who owns the final decision. A workflow that happens twice a year may not justify integration work; one repeated dozens of times a day may.
Separate a simple rule from an AI problem
If the system only needs to check whether a field is blank and route a record, ordinary workflow rules are often better. Consider AI where the system must interpret messy text, classify variable documents or find answers in approved knowledge. Even then, keep deterministic checks around the decision.
Score the candidate workflows
Compare potential projects against five criteria: frequency, time spent, cost of mistakes, data readiness and implementation risk. A good pilot has enough volume to measure an improvement, but low enough risk that mistakes are recoverable. Do not automate sensitive approvals first unless robust controls are already in place.
Specify what happens when the system is unsure
Define missing-data behavior, confidence thresholds, human review and exception alerts before building. For an invoice workflow, a missing total should create a review task rather than silently fill in a plausible number. For a support workflow, an unsupported answer should move to a person.
Measure a real outcome
Choose a baseline you can check: handling time, rework rate, ticket reassignment, backlog size or the portion of documents that need review. Run a contained pilot, compare the baseline and investigate exceptions. A new automated step is only an improvement if it reduces effort without introducing unacceptable errors.
Build the second project from what you learned
Once the team trusts the first workflow, use its integration patterns, monitoring and governance as a foundation. That is more dependable than automating ten disconnected processes at once.
A simple scoring worksheet
Create a short list of three workflows and rate each one from 1 to 5 for repeat frequency, process stability, quality of source data, ease of reviewing mistakes, and value if solved. Then separately flag risks: regulated data, payment instructions, irreversible writes and downstream customer impact. A high-value workflow with unacceptable risk is not a good first experiment. Choose the candidate that teaches your team something useful without exposing an essential service to uncontrolled decisions.
- Write the triggering event and the accountable process owner.
- Record the current time per task and the correction rate.
- Identify what data the automation may read, retain and change.
- Define one safe human override and one clear stopping condition.
Pilot design: where rules end and AI begins
For example, a service enquiry workflow can use deterministic rules to validate an email address, enforce consent choices and assign a destination. AI can help summarize an unstructured message or propose a category, but the system should record its evidence and refrain from inventing missing facts. Limit write permissions until the team has reviewed enough examples. A reviewer should be able to reverse a mistaken classification without reverse-engineering the entire workflow.
The handoff is part of the product
Specify what an employee sees when the model cannot answer, times out, or produces a low-confidence result. Include a link to the original message, the suggested classification, the reason for escalation and the next responsible person. Track both speed and quality: a faster process that silently misroutes priority enquiries is not an improvement. Document the fallback so operations can continue while the automation is unavailable.