Describe the current process
Who receives the data, copies it, checks the result and hands it over? Collect real examples and typical errors. Before choosing a tool, understand whether the problem is collecting information, processing it or passing it to the next person.
Separate rules from judgment
Moving fields between tools often does not require AI. Understanding free text may need a model and review. We define what the system prepares, what a person checks and which actions need additional approval.
Try a limited pilot
Choose a small set of examples, including incomplete and unusual inputs. Measure time including review and corrections, not just the first response. Agree what happens when the process fails and who takes over.
Example pilot: preparing incoming enquiries
Imagine a team reads messages each morning, identifies the requested service and copies details into a spreadsheet. A pilot could prepare a short summary, suggest a category and flag missing information. A person then checks the suggestion before contacting the customer. The first test does not need automatic replies. This illustrative workflow lets you evaluate preparation quality on a task the team already understands.
Measure the whole task
First record how long the usual manual process takes. Then include input preparation, waiting, review and correction in the new workflow. A quickly generated answer has not demonstrated useful savings if checking it takes longer than the original task. Record the types of errors too. Some are easy to correct; others call for a different process or a more reliable source.
Agree what happens when the workflow stops
Define where failed processing appears, who owns the follow-up and how work continues manually. Check duplicates, incomplete documents and temporarily unavailable services. Record which data the system actually needs and who can access it. A useful pilot ends with instructions the team can follow and a decision about further development. A list of AI tools used is not a deliverable by itself.
What we checked on our own website
We used Midaz's enquiry intake as a specific process: the form, the saved original message, an assigned task and the visitor acknowledgement. The check aimed to establish that the responsible person could see what was submitted and continue the conversation. A button-click count was not a substitute for that outcome.
We checked the Bosnian and English forms, task content and incoming email and WhatsApp replies. In controlled repetitions, processing the same event did not create an additional task, while a separate message received its own task. Repeating one event and receiving a new request from the same contact need different handling.
The workflow went live on 2 October 2026. These checks establish the path from message to task; they do not establish a percentage time saving or sales growth. That would require comparing actual handling time and outcomes over a relevant period.
For your first pilot, make a similar list: a complete request, missing information, repeated processing, another message from the same contact and a temporarily unavailable service. Write the expected outcome and the person who handles the exception for each case. Start with rules for the data handoff, then add AI only where interpretation is needed.
Frequently asked questions
Which process is a good first candidate for AI automation?
A recurring task with known inputs, a clear output and someone able to assess quality. An example is a team that reads messages every morning, identifies the requested service and copies the details into a spreadsheet.
Does every process need AI?
No. Moving fields between tools often does not require AI. A model can be useful when free text has to be understood, but even then the result should be checked. Before choosing a tool, find out where the problem arises: in collecting information, processing it or passing it to the next person.
What is a pilot?
A limited first test of one process on a small number of examples, including incomplete and unusual inputs. It is agreed in advance what the system prepares, what a person checks and what happens when the process fails. It ends with instructions the team can follow and a decision on what is worth building further.
How do I know whether a pilot helps?
Record how long the usual manual process takes, then include input preparation, waiting, review and corrections in the new workflow. A quick answer has not shown a benefit if checking it takes longer than the original task. Record the types of errors too: some are easy to correct, others call for a different process or a more reliable source.
What happens when the automation stops or gets something wrong?
That should be agreed before you start: where a failed run appears, who receives the task and how work continues manually. Check duplicates, incomplete documents and temporarily unavailable services. Also record which data the system actually needs and who can access it.
Bring one process and a few examples. The first goal is to show that the solution helps with that specific job.