Start with repeated work that uses sources you can provide, produces a result someone can check, and keeps mistakes easy to contain. Compare candidates across your teams before choosing which tasks to automate with AI first.
I would use the worksheet below to make that choice. It is a proposed checklist; its usefulness and any savings need to be tested on your work. It starts with the problem, the people involved and the goal, then asks whether automation is worthwhile.
Compare the work your teams do
Ask each team for repeated work that takes time away from customers or delays a decision. Describe the input, the finished result and the person responsible for it. "Sort incoming requests by department" is specific enough to examine. "Improve operations" needs a clearer description.
For each candidate, fill in these five rows using examples from the people doing the work.
| Question | Evidence to bring | What makes it ready to test |
|---|---|---|
| How much work is there? | The actual count over a stated period and the time spent completing it | Enough repeated effort to justify investigating a change |
| How consistent are the steps? | Written steps and examples of exceptions | A clear path for routine work and a person to handle exceptions |
| Are the inputs ready? | Current documents, records and permission to use them | Agreed sources that the test can use |
| How will someone check it? | A definition of an acceptable result and the check itself | A named reviewer who can detect an incorrect or incomplete result |
| What happens after a mistake? | Who could be affected, how it is caught and how it is corrected | Review before consequential actions and a practical correction path |
Even a high-volume task needs a result that someone can check. If a wrong result could harm someone or commit the business, settle the approval and correction process before testing that action.
Examples to discuss with each team
These are proposed workflows. Your test would establish whether they help.
For sales enquiries, prepare a reply from current product information. The salesperson checks the requested details, prices and commitments before sending it. A missing answer goes back to the salesperson.
For invoice entry, extract details into a draft entry. Compare them with the invoice and the relevant order before a person approves the entry. Keep payment approval as a separate decision.
For operations reports, assemble a draft from agreed records. Check that the figures reconcile and that the report covers the required period. First resolve which file is authoritative when records disagree.
For incoming requests, suggest a category and destination. Let the reviewer correct the routing, and give urgent or unclear requests an explicit path to a person.
These examples fit the kind of business process automation I build. The right starting point depends on your inputs and checks, rather than the department name.
Use a measured result to choose what comes next
Before a build, write down the current completion time and the quality standard. Choose examples that include routine cases, exceptions and incomplete inputs. Keep the same standard when you compare the AI-assisted work.
Count preparation, checking, corrections and failed attempts. Compare the full cost per accepted result, as well as the time taken to finish the work. Include the human effort around the generated result in that comparison.
Agree the next decision in advance: expand work that meets the standard at a worthwhile full cost, revise a specific weakness, or keep the existing process. Record what happened so the next team has evidence to work from.
Frequently asked questions
Do we need AI for every repeated task?
Start with the simplest way to do the job. A fixed calculation or a rule that copies a field may suit ordinary software. Consider AI where the work involves interpreting varied text or preparing a draft, with a check on the result.
Who should choose the first workflows?
Bring together the people doing the work, the person responsible for its outcome and whoever will check the results. They can compare the benefit across teams and agree who owns each decision.
Can we start before our documents are in order?
Start with the documents needed for the workflows you want to test. Have the people responsible agree which versions are current. Use the test results to decide what needs organising next.