Automate the work that repeats every week.

The most frequent process with a written-down rule, not the one that annoys people most. Return is frequency multiplied by time saved, so a short task repeated daily almost always beats a long one done monthly, and the daily task is usually the one nobody complains about. Count both before choosing.

Often more so, because a small team has no slack when one person is away and no operations function to absorb the work. The case here is usually continuity and error rate rather than headcount saved, which is a different argument from the one made to a large company. Below a few hours a week of a genuinely repetitive task, the honest answer is usually no.

A single well-defined workflow is typically 2 to 4 weeks including the measurement before and the parallel run after. What extends it is access rather than engineering: a system with no interface, a portal that has to be operated by hand, or a document set nobody has collected examples of. We ask for those in the first week.

It should stop and tell someone, which is a design decision rather than a given. A workflow that fails silently is worse than no workflow, because the output is trusted and wrong, so failure paths and alerting are built before the successful path is signed off. Anything below a confidence threshold goes to a person instead of proceeding.

Somebody on your side, named before the build starts, and that constraint is what keeps the design simple enough to be maintainable. Logic stays readable and credentials sit in your own accounts, so a form change is an edit rather than a support request. Where nobody can own it, we would rather scope something smaller than leave a dependency behind.

Wherever the workflow is hosted, which is why it is an architecture decision rather than a policy one. Under the UAE PDPL, processing personal data through a hosted service makes that provider a processor you are accountable for, so where documents carry personal or commercially sensitive content we self-host the workflow engine and keep the data inside your environment. Entities registered in DIFC or ADGM run their own regimes, which we scope to specifically.

When the process is wrong, when it is still changing monthly, or when nobody will own the result. Automating a broken process delivers the wrong outcome faster and more consistently, which is worse than the manual version because it is trusted. In those cases the useful work is redesigning the process, and we would rather say so than sell the build.