Scoping
Most AI agent projects don't fail during build. They fail during scoping — or more precisely, they fail because they skip scoping entirely and treat "we should use AI agents" as a sufficient technical spec.
Six months later, the team has a sophisticated demo that can't handle the real input distribution, doesn't integrate with the production data sources, and has no evaluation framework to tell them whether it's actually working. The project gets declared a pilot, moved to quarterly review, and quietly dies.
The fix is almost always a proper scoping phase. This article describes what that looks like in practice.
The most common scoping mistake is starting with the technology instead of the workflow. "We want to build an AI agent for customer support" is not a scope. It's a technology preference attached to a department name.
A real scope starts from the workflow: what are humans doing right now, step by step, and what specifically do we want the agent to do instead? The answer has to be precise enough that you could write a test case for it.
"We want an agent to handle customer support" is not a scope. "We want an agent to classify inbound support emails by issue type, retrieve the relevant account history, draft a response for tier-1 issues, and route tier-2 issues to a human queue with the context pre-filled" is a scope.
The difference is not semantic. The second version tells you what APIs you need, what data the agent requires, what success looks like, and where the human handoff point is.
We run every Discovery engagement against these questions. If you can't answer them clearly, the project is not ready to build.
What event starts the workflow? An inbound email, a new record in a database, a scheduled time, a user action? The trigger defines your system's entry point and dictates the integration architecture.
What does "done" look like? A structured record written to a database, a message sent, a file created, a human notification delivered? Vague outputs produce vague systems.
What information does the agent need to complete the task, and where does it come from? Internal databases, third-party APIs, user input, documents? Each data source is an integration that needs to be built and tested.
What should happen when the agent can't complete the task? Escalate to a human, retry with different context, log and skip, alert an operator? This is the question most teams skip, and it's where most production systems break down.
How will you know the agent is working? This is not a qualitative question. You need a measurable metric: accuracy rate, processing time, escalation rate, cost per completion. If you can't name a metric, you can't evaluate the system.
Does any output require human review before acting? Are there regulatory constraints on automated decision-making in this workflow? Does legal need to approve the implementation? These questions don't kill projects — discovering the answers mid-build does.
What's the current state? How long does the workflow take manually? How many FTEs does it consume? How often does it fail or require rework? The baseline is how you calculate ROI and how you set a meaningful performance bar for the agent.
A structured two-week scoping phase — what we call Discovery — typically costs $4,500 and produces a technical spec and a fixed-price build proposal. That sounds like overhead. In practice, it's the cheapest risk mitigation available on an agent project.
The alternative is starting a build without the answers to these questions. The median outcome is a project that runs for three to four months, produces a system that doesn't handle the real input distribution, and requires substantial rework before it can go to production. The total cost is usually 3–5x what a well-scoped build would have cost.
We've turned down build engagements where Discovery revealed that the workflow wasn't actually a good fit for agents. That's not a failure — that's the Discovery working correctly.
Ready to scope a project? Our Discovery engagement is $4,500 flat and gives you a technical spec plus a fixed-price build proposal. Book a scoping call to start.