AI Agents for Project Discovery: A Practical Guide

Most AI implementations fail because they try to boil the ocean. Teams attempt to build general-purpose AI assistants that can do everything, and end up with systems that do nothing well.

We took a different approach: build AI agents for specific, high-value workflows. Our first target? Project discovery and intake.

The Problem

Business Systems teams spend 30-40% of their time on intake:

• Clarifying requirements
• Understanding scope
• Identifying impacted systems
• Estimating effort

Most of this is repetitive pattern-matching that an AI can handle.

Our Solution: The Discovery Agent

We built an AI agent that conducts structured discovery conversations. Here’s how it works:

1. Conversation Flow

The agent follows a decision tree, but with natural language:

START → Understand business outcome
      → Identify impacted processes
      → Map to systems
      → Capture success metrics
      → Estimate complexity
      → Generate Problem Brief

2. Prompt Engineering Patterns

We use several key patterns:

Chain of Thought: Force the model to reason step-by-step:

Before suggesting a solution, I need to understand:
1. What business outcome are you trying to achieve?
2. What process does this affect?
3. Which systems are involved?

Few-Shot Examples: Include examples of good Problem Briefs:

Here's an example of a well-structured Problem Brief:
[EXAMPLE]
Now, based on our conversation, here's your draft...

Guardrails: Prevent hallucination with explicit constraints:

Only suggest systems that exist in our technology catalog:
- Salesforce
- HubSpot
- Snowflake
- ...

3. Human-in-the-Loop

The agent generates a draft Problem Brief, but a human always reviews before it’s finalized. This catches edge cases and builds trust.

Results

After 3 months:

• Intake time reduced 60% — from 2 hours to 45 minutes average
• Higher quality briefs — more complete, consistent format
• Better stakeholder experience — 24/7 availability, immediate response

Lessons Learned

1. Start narrow — One workflow, done well, beats ten done poorly
2. Guardrails matter — Constrain the model to your domain
3. Keep humans in the loop — AI assists, humans decide

---

Want to explore AI agents for your workflows? Let’s talk.