That process someone does by hand every week, and everyone hates.
Sorting emails, moving data from one system to another, generating the same report every Monday, checking documents for the same thing over and over. You don’t need an "automation strategy" — you need to automate that one specific thing, the one you already know how to name because you suffer through it every week.
How I approach it
I start by asking you to show me the process exactly as it’s done today, by hand, on a real case — not described, done in front of me. That’s usually where the part nobody had mentioned shows up (the exception handled "by feel", the step that depends on someone remembering something). I build the automation combining AI where judgment is genuinely needed and simple APIs where it isn’t, and test it against real data before calling it done.
What usually fits
Integrations between the tools you already use, with AI only at the point where judgment is truly needed — classifying, extracting, choosing between options. I don’t replace tools that work; I connect what’s connected by hand today.
The work, in concrete terms
- What we start with
- A complete example of the manual process, its input data, the tools involved and common exceptions. We also need to know who authorises actions and how errors are detected today.
- What it should return
- The specific result previously produced by manual work: transferred data, classified information or a prepared report. Testing should cover the normal path and cases that require intervention.
- What needs to be defined
- Not every step needs AI, and not every decision should run automatically. We agree permissions, validations and human approval points; failures in external tools are also part of the design.
Illustrative example
Every Monday, data from two tools is collected to prepare a report. The automation gathers and structures the information; a person reviews exceptions and approves the result before distribution.