Automation works best when the process it serves is understood. If a team spends hours moving information between systems, the first question is not which tool to buy. It is why the handoff exists, which rules govern it and what happens when the information is incomplete.
Start with one outcome
Choose a result that a business owner can recognise: orders entered without re-keying, invoices routed to the correct approver or support requests assigned with the necessary context. Avoid an objective such as “use AI across the business”; it cannot tell you whether a particular workflow improved.
Record a baseline using representative work. Measure elapsed time, manual handling, exceptions and rework separately. Waiting for approval is a different problem from typing data into another system, even when both contribute to the same delay.
Map what people really do
Follow a recent item from its arrival to its final record. Ask the people doing the work to show the screens, spreadsheets, messages and checks they use. Capture unofficial steps as well as the documented process; they may be compensating for an important missing rule.
- What event starts the workflow?
- Which system holds the authoritative record?
- What must be checked before the next action?
- Who decides when a case does not match the normal route?
- What proves that the workflow has finished correctly?
Separate rules from judgement
Stable rules can often be implemented directly: matching an order identifier, validating a required field or routing a request by an agreed category. Ambiguous interpretation may suit an AI-assisted step, provided the output can be checked. Important approval decisions should remain with an accountable person unless a separately reviewed policy permits otherwise.
A practical design often combines deterministic validation, a model for a bounded interpretation task and a review queue for exceptions. This is more useful than asking a model to control the entire workflow.
An audit worksheet
| Area | Record | Why it matters |
|---|---|---|
| Inputs | Fields, formats, permissions and missing data | Defines what the automation can rely on. |
| Rules | Checks, routing and approval thresholds | Makes behaviour testable. |
| Exceptions | Types, frequency and owner | Creates a workable human fallback. |
| Integration | APIs, identifiers and rate limits | Reveals technical constraints. |
| Outcome | Baseline, success measure and review date | Shows whether the investment helped. |
Worked example: incoming purchase requests
Imagine requests arriving by email, being copied to a spreadsheet and then entered into an ERP. A useful first project could capture the request, validate the cost centre and required fields, create a draft ERP record and send incomplete cases to the requester. The budget owner still approves the purchase.
This boundary removes duplicate entry without confusing data capture with spending authority. The pilot should include duplicate emails, unknown suppliers, missing attachments and API failures, not only the clean examples used in a demonstration.
This is an illustrative scenario, not a client case study.
Design for retries and failures
Use stable identifiers so that retrying a request does not create duplicate records. Track the state of each item, make unsuccessful actions visible and provide a safe way to resume after a dependency fails. An unattended automation that silently loses items creates more work than it removes.
Limit access to the systems and fields required for the task. Define what is logged and how long it is retained; operational visibility does not require keeping every sensitive document indefinitely.
Decide what to automate next
Review the pilot against the baseline, including the effort spent correcting its output. Expand only after the normal path and exception path both work. Sometimes the audit shows that simplifying an approval or improving a form is the better first change.
ANAWAZ’s automation and systems integration service helps turn these findings into a scoped implementation. For interpretation or document tasks, explore our AI engineering service. Describe the workflow you want to improve.