Work with what you have
Use the inboxes, documents, databases, CRMs, ERP systems and internal tools already carrying the work.
Start with the job, not the AI.
I define the process, permissions and approval points before expanding what an agent can do.
What a useful agent can do
An agent becomes useful when it has a clear job, controlled access to tools and defined limits.
Handle repeatable steps
Read, check, prepare, update and pass information between systems without pretending every decision is routine.
Stop for decisions
Keep payments, access changes, exceptions and anything risky with a person.
Leave a trail
Record actions, sources and failures so the work can be inspected.
Small first. Useful first.
I begin with one recurring job that is currently handled by hand.
Show me the job
Walk me through what starts it, where the information comes from and where it needs to end up.
Find the repeatable part
Separate routine handling from decisions that still need experience or judgment.
Build it narrowly
Make one process reliable before expanding its tools or responsibilities.
Keep it understandable
Provide clear controls, visible hand-offs and enough documentation for people to understand how it works.
Where an agent may not fit
Some processes are better handled by people or conventional software.
The process changes every time
If nobody can explain the job twice the same way, there is nothing stable to build around yet.
A mistake cannot be recovered
Irreversible actions need tight permissions, human approval or no agent at all.
Normal software already solves it
If conventional software already solves the problem well, adding an agent would create unnecessary cost and complexity.