Start with a task your team can show you from beginning to end. A recent enquiry, an approval that stalled or a weekly spreadsheet can reveal more than a broad request to “bring AI into the business.”
The first project should have a defined responsibility, usable inputs and someone who can decide whether the result is right. It also needs a result worth measuring. That combination is more useful than choosing whichever demonstration looks most impressive.
Describe the work before choosing the tool
Write down what starts the process, where information arrives, the decisions people make and where the result goes. Include the handoffs and the exceptions that usually require a call or an email.
For example, “prepare approved timesheets for payroll review” gives a project a boundary. “Automate payroll” could include approvals, calculations, deductions, payments and disputes. The second phrase hides several different responsibilities in one brief.
Compare a small set of candidates
Choose three recurring tasks and examine them using the same questions:
| Question | What a useful starting point looks like |
|---|---|
| Does the work recur? | Enough examples to understand the normal cases and exceptions |
| Is the effort visible? | A team can show time spent, delays or rework |
| Are the rules explainable? | People agree what a correct result looks like |
| Are the inputs usable? | Records have an owner and can be accessed appropriately |
| Can exceptions be handled? | Unusual cases reach a named person with useful context |
| Can the scope stay narrow? | One output or responsibility can be tested independently |
These are discussion questions, not an automatic ranking formula. A frequent task may still be a poor first project if every case requires a different decision or the information cannot be trusted.
Check whether ordinary automation is enough
Moving a record, checking a required field or sending an overdue reminder may need clear rules rather than a language model. AI may be useful where inputs vary, such as classifying an enquiry or extracting information from documents. Its output still needs checks appropriate to the consequences of a mistake.
An illustrative pilot might classify incoming enquiries and prepare a queue for staff. Starting by letting it independently make commitments to customers would introduce different risks and review requirements. A smaller responsibility can be tested against labelled examples before the scope expands.
Resolve ownership and difficult cases
Name the process owner, the person who maintains the information and the team that receives exceptions. Discuss missing records, contradictory inputs, system outages and repeat submissions before building the happy path.
Where an action needs approval, make the approval an explicit workflow step. Tools can support this: n8n documents human review of selected AI tool calls, for example. That capability still needs the correct approver, permissions and operating arrangements; it does not replace the organisation’s judgement (n8n, accessed 26 October 2026).
Measure the same work before and after
Collect a baseline for the chosen task. Useful measures might include processing time, waiting time, rework, unresolved exceptions and staff effort. Separate time spent actively working from time a record spends waiting for a decision.
During the pilot, count the work still required from people, including checking answers and resolving errors. Avoid treating every automated step as a realised saving. Decide in advance what evidence would justify expanding, revising or stopping the project.
Agree a first-project brief
Summarise the trigger, inputs, decisions, permitted actions, exception route, output, owner and success measures. Keep a few ordinary examples and a few difficult ones beside the brief. They will help the team check whether the proposed solution matches the work.
If the organisation’s information, skills or controls need preparation, our free organisational AI readiness assessment can help identify those foundations. It does not determine whether a particular workflow offers sufficient value or is technically feasible.
Explore workflow automation, find a practical SME starting point or discuss one candidate process.
Source
n8n (accessed 26 October 2026), Tools Agent: human review. Supports the specific product capability described. The candidate-selection method and pilot advice are Spiritus’ editorial recommendations; no validated benchmark or guaranteed result is asserted.