AI project discovery8 min read · 01 / 20

Why the best AI projects start with a messy conversation.

I have left discovery calls feeling excited by the opportunity and unsettled by the lack of a clear answer. That discomfort is often where the real AI project begins.

Written by Mitesh DabhiAI Adoption Specialist
Hand-drawn diagram showing a messy client conversation becoming a structured brief and focused AI project through signals of repetition, delay and dependency.
What feels like a messy conversation can become a focused AI project if I resist the urge to jump to the tool.

I know the feeling: the call ends, the notes are full, and the answer is still nowhere in sight.

I have experienced versions of this moment many times. One person describes a repetitive task. Another mentions a spreadsheet that only one employee truly understands. A manager talks about delays. A team member explains how they copy the same information between systems every day.

Everyone is describing a real problem, but the stories do not yet connect. I feel energized because I can sense the opportunity. At the same time, I feel uncomfortable because there is no single requirement, no obvious product and no neat answer to take away.

I used to believe a strong discovery call should end with a crisp feature list. Now I know that pressure can make us choose a tool too early. The unresolved moment after the conversation is not the end of discovery. It is where the real discovery starts.

01 / After the call

The conversation felt messy because the workflow was messy.

After the call, I stopped trying to summarize what everyone had said. Instead, I started tracing how the work actually moved from one person to another. The official process sounded clean: receive a request, review it, prepare a response, approve the work and deliver the result.

  • Searching through previous emails
  • Copying data from multiple platforms
  • Asking a senior employee for missing context
  • Reformatting information manually
  • Waiting for approval
  • Correcting errors caused by incomplete inputs

That was the first turning point for me. The confusion was not a sign of poor discovery. It was evidence that the real workflow lived between systems, handoffs and people’s memories. The conversation had exposed what the process document could not.

02 / What I listened for

I stopped hunting for an AI idea and listened for three signals.

Once I let go of needing an immediate solution, three patterns became easier to hear: work that repeated, work that waited and work that depended on one person’s knowledge. Those signals did not tell me to automate everything. They showed me where to look more carefully.

The patterns that brought clarity

The conversation started making sense when I focused on the friction instead of the technology.

01

Repetition

The same task keeps returning, using similar information and producing a familiar output.

Reports · summaries · reviews · classification
02

Delay

Work regularly waits for information, formatting, review or manual coordination between people.

Handoffs · approvals · missing inputs · rework
03

Dependency

One experienced person holds the context, knows where everything lives and keeps the process moving.

Tacit knowledge · exceptions · decision history
03 / Finding structure

A blank page turned the conversation into something we could build.

I did not begin with an architecture diagram. I opened a blank page and organized what I had heard into a simple brief. I wanted to understand the work before deciding how AI should enter it.

01Business problem02Current workflow03People and systems involved04Inputs and outputs05Time currently spent06Failure points07Human judgement required08Automation opportunities09Risks and open questions

As those boxes filled, I felt the shift from uncertainty to clarity. The requirement had been there all along, scattered across the conversation. Structuring it made the opportunity visible and gave the team something concrete to question, correct and improve.

The question I asked myself“What had the business already told me that I had not structured yet?”
04 / Choosing restraint

The most useful answer was smaller than the flashy one.

Once the brief became clear, it was tempting to imagine a large AI agent that handled the entire process. But the team did not need a magical system. They needed one painful part of the work to become calmer, faster and more reliable.

Faster turnaroundFewer manual handoffsBetter consistencyImproved visibilitySearchable knowledgeStructured approvals

That change in perspective mattered to me. I moved from trying to prove what AI could do to helping the team decide what should improve first. The project became smaller, more practical and easier to trust. It finally felt like something people could use, not just something impressive we could demonstrate.

What I carry forward

I no longer judge a discovery call by how quickly it produces an answer.

I judge it by whether it reveals how work really happens: where people wait, what they repeat, whose knowledge holds everything together and which decision is harder than it should be.

The uncertainty I feel after a messy conversation no longer worries me in the same way. It reminds me to slow down, listen again and structure what I heard before reaching for a model or platform. That is usually where the useful idea is hiding.

The strongest AI projects do not begin when I find the right model. They begin when I understand the real human problem.