Choosing the approach
When is an AI assistant enough, and when do we need automation?
An AI assistant can be enough when a person brings the context, directs the work and reviews the result. Consider automation when a recurring process has a clear trigger, defined steps and someone responsible for exceptions. Start with what your existing tools can do.
Separate help with a task from running a process
Someone comparing documents for an occasional meeting can work directly with an assistant. They choose the files, ask follow-up questions and decide what to use. The conversation lets them adjust the request as they learn more.
A report that must arrive every Monday raises additional questions. What starts the work? Which files are current? Where does the result go? What happens when an input is missing? Automation needs answers to these questions even when AI performs one of its steps.
Choose the simplest approach that meets the need
| Situation | Approach to consider | Reason |
|---|---|---|
| An occasional task with changing questions | A person working with an AI assistant | The person can supply context and adjust the work as it develops. |
| A repeatable calculation or fixed transformation | A formula or feature in existing software | The required rule can be stated directly. |
| Repeated movement of information under clear rules | An integration or automation | The trigger, destination and expected result can be defined. |
| A recurring process involving variable text | A tested AI step within a controlled workflow | Interpretation may help, while checks and exception handling remain explicit. |
| An unresolved decision with competing priorities | Human discussion and judgement | Someone must decide what the work should achieve. |
Use repetition as evidence, not a reason on its own
A repeated task may still be easy enough to do manually. Compare the effort you expect to save with the work of setting up, checking and maintaining an automation. Include the people who will deal with failures. Frequency matters alongside complexity, consequence and how often the process changes.
Try the method with a person first if you are still discovering the steps. Keep a note of the decisions they make. Those decisions show which rules can be specified and where a person still needs to take over.
The same PDF task can need different arrangements
For an occasional comparison, a colleague could supply approved PDFs to an assistant, request a table with page references and check the result. Saving the instructions and one checked example may be sufficient for the next person.
If that comparison becomes a regular shared service, the surrounding work needs attention. Agree how files arrive, how the correct versions are selected and where checked results are stored. An automated draft might wait for review. Missing files or unreadable pages should produce a visible exception rather than a complete-looking report.
Test the arrangement with duplicate files, changed layouts and interrupted runs. Decide whether a failed step can safely repeat. The person receiving the result should be able to tell whether the process finished and which material it used.
Name the owner before adding a build
A custom build may be justified when existing tools leave a clear gap, such as a required connection, access boundary or way of working. Write that gap down before choosing how to build it. Include the ongoing support, expected running costs and what the team will do if the system is unavailable.
Assign someone to review failures and changes to the inputs, tools or process. Make the first decision revisitable: what evidence would justify more automation, and what would make you simplify it? The appropriate level of automation can change as the team learns.
Related reading
Start with the problem.
Bring a decision, a stuck piece of work or something that may need building. We can work out whether CURN is the right fit.
Book a call