One thing I wouldn't expect a client to tell me is:
“We need an AI agent connected to our CRM with a retrieval layer and an automated escalation workflow.”
If you already know all that, wonderful.
But usually, I'd rather hear:
“Our team spends three hours every morning doing this manually.”
That's a much better place to start.
When I approach an AI consultation, I don't begin by asking what software we should build. I begin by trying to understand how the business actually works.
Show me the process.
Who does what? What happens next? Where does someone copy something into Excel? What requires five WhatsApp messages and a phone call? Where do mistakes happen? What does the owner wish they could see but can't?
And one of my favourite questions:
“What do you do when this doesn't go according to plan?”
Because the happy path is usually easy. The exceptions are where the interesting problems live.
Once we understand the workflow, we can start separating symptoms from the actual problem.
Maybe the company thinks it needs an AI customer-support agent.
But perhaps 70% of the enquiries exist because customers can't find basic information themselves.
Now the problem looks different.
From there, I look at the options.
Could an existing tool solve it? Can we automate part of the current workflow? Does AI genuinely add something? Do we need a small custom application? Or is the process itself the thing that needs fixing?
Sometimes the recommendation might even be:
Don't build anything.
If we do need to build, I prefer starting small.
Build the part that proves the idea. Put it in front of the people who will actually use it. Watch what happens. Find the awkward edge cases. Measure whether it saves time, reduces errors or improves the outcome.
Then decide what deserves to become a bigger system.
AI has made it incredibly easy to build things.
That's precisely why deciding what deserves to be built has become more important.
So you don't need to come to me with an AI strategy, a technical specification or even a solution.
Come with the messy process.
We'll start there.