What Does Rigorous AI Agent Discovery Look Like Before Any Build Starts?

Quick answer

Rigorous AI agent discovery covers four areas before any vendor selection or build begins: mapping the actual conversation volume and intent distribution from historical data, auditing the data and systems the agent would need access to, defining measurable success criteria the business agrees on in advance, and identifying the organizational owner accountable for the agent post-launch. Skipping discovery is the single most common reason AI agent projects overrun budget or underperform after launch.

The projects that overrun budget or underperform after launch almost always share one root cause: the business jumped from "we want an AI agent" straight to vendor selection or build, without a discovery phase that actually defines what the agent needs to do and against what data.

Map real conversation volume and intent distribution

Before writing a single requirement, pull six to twelve months of historical conversation logs (support tickets, chat transcripts, call summaries) and categorize them by intent and volume. This produces the actual shape of the problem, which is almost always different from what stakeholders assume from memory. It also directly informs the team sizing and budget conversation covered in /blog/team-sizing-budget-ai-agent-engagement, since intent complexity drives effort far more than raw volume does.

Audit the underlying data and systems

An agent is only as good as the data and tools it can reach. Discovery should include a hard look at knowledge base freshness, CRM data quality, and which systems have (or lack) an API the agent can call. This is the same data readiness question covered in depth in /blog/data-preparation-for-ai-agent-onboarding, and it needs to happen before scoping, not after the build has already assumed clean data exists.

Define measurable success criteria the business agrees on upfront

Discovery should produce a written definition of success: a target containment rate, an acceptable escalation rate, a latency ceiling, a CSAT floor. Defining these after the agent is built invites goalpost-moving and makes the ROI calculation covered in /blog/calculate-ai-agent-roi-before-greenlighting impossible to run honestly, since there is nothing concrete to measure against.

Identify the accountable owner before build starts

Every AI agent needs a named internal owner accountable for its quality after launch, not just during the build. Discovery is the right time to identify who that is and confirm they have the authority and bandwidth for the role, rather than discovering post-launch that no one is actually watching the agent's performance.

How discovery output should feed the next steps

The output of discovery, intent map, data audit, success criteria, and owner, is exactly the input a vendor RFP process needs to produce comparable bids, covered in /blog/ai-agent-vendor-rfp-bakeoff-checklist. Vendors asked to bid without this groundwork end up guessing at scope, which produces estimates that are not actually comparable to each other and that frequently need renegotiation once real requirements surface mid-build.

The cost of skipping it

Teams that skip discovery to save two or three weeks routinely lose far more than that in rework once the build phase surfaces requirements discovery would have caught, from missing API access to an intent category nobody had sized. Discovery is cheap relative to the cost of re-scoping a project that is already midway through build.

FAQ

Q: How long should a discovery phase take?

For a mid-sized deployment, two to four weeks is typical. Shorter than that usually means the intent analysis and data audit are being skipped rather than compressed, which shows up later as scope surprises during build.

Q: Who should run discovery, the vendor or the client?

Ideally a joint effort. The client has the historical conversation data and system access; the vendor or agency has the pattern-matching experience to know what data will actually matter once the agent is live. Discovery run entirely by one side tends to miss the other's blind spots.

Q: What's the single most commonly skipped discovery step?

Auditing whether the underlying data (knowledge base articles, CRM records, product catalogs) is actually clean and current enough for the agent to use reliably. Teams often assume the data is fine because humans have been working with it for years, but human agents silently compensate for gaps that an AI agent will surface immediately.