The 2-Minute Script That Spawned a Week of Work
Mike Waszazak sat down to do something simple: migrate years of scattered notes into Obsidian. He opened a Claude Code session in VS Code, described what he wanted, and the AI wrote a migration script in roughly two minutes.
Simple job done. Time to move on.
But before he could close the terminal, Claude said something unexpected:
“You have 116 projects in these exports. I’m seeing clusters — Product-Alert has 11 projects, Product-Classroom has 12, Leadership has 8. Want me to analyze the themes?”
That single suggestion triggered a chain reaction. What followed was a daily notes template, automated task filtering, a tagging system, visual styling for note types. One week later, Mike had a completely different setup than anything he’d planned.
He called this experience “scope creep as discovery” — and it flips everything we know about software estimation on its head.
The Great Inversion
Traditional project management treats scope creep as Public Enemy Number One. We write detailed specifications, freeze requirements, and charge extra for every change. The core assumption is: the client knows what they want, and our job is to deliver it efficiently.
But AI agents reveal a different truth: most of the time, nobody knows what they want until they see what’s possible.
Mike’s experience maps to three conditions that turn scope creep from a liability into an asset:
1. Cheap iteration
Before AI: Trying a new feature meant weeks of development. Every wrong turn cost money. Scope management was about protecting the budget.
With AI: Trying, seeing if it sticks, and adjusting costs minutes instead of sprints. The first task system Mike built was over-engineered. He scrapped it within two days. No one cared. The folder structure took three iterations. Some features were built and never used. None of this mattered, because the cost of exploration was trivial.
2. Persistent context
Unlike human developers who require handovers, the AI’s context accumulated across sessions. It remembered the folder structure, the naming conventions, the design decisions. When Mike explained something twice, it was encoded into a custom skill — a permanent capability that never needed re-explanation.
This compounds over time. Each piece of work connects to the last. Capabilities build on capabilities.
3. Emergent opportunities
The most striking example of emergent value was completely unplanned. Mike asked Claude to build a skill for capturing daily accomplishments. He mentioned wanting “coaching-style feedback” — expecting something generic.
Claude responded with this appended to his daily note:
“Strategic Alignment: Strong day. You mentioned the leadership operating model deck in passing, but don’t undervalue it — the shift from consensus-based decision-making to directed ownership during transformation is high-leverage work.”
Then it added:
“Content Opportunity: The leadership model shift could be a LinkedIn post. Angle: ‘Consensus is great for steady-state. It’s terrible for transformation.’”
The output exceeded the input. That’s the difference between a feature and a collaborator.
Why Your Estimation Models Are Breaking
This isn’t just an interesting workflow story. It’s a fundamental challenge to how software agencies price and scope projects.
The fat-tail problem
With traditional development, the work distribution follows a rough curve: 80% of features take 80% of the time. The tail is manageable.
With AI agents, the distribution inverts. The 80% that’s fast becomes very fast — a two-minute script replaces a day of coding. But the last 20% — the emergent features, the discovery-driven tangents, the things nobody knew they wanted — becomes completely unpredictable. You can’t estimate what you didn’t know existed when you wrote the proposal.
The hybrid estimation paradox
Mike’s experience suggests a three-phase model:
Phase 1 — AI Discovery Sprint (Time & Materials): Let the agent explore the problem space. Surface friction points, identify patterns, discover the actual scope. Don’t try to estimate this in advance — you can’t.
Phase 2 — Human Estimation (Fixed, informed): Once the AI has surfaced the real requirements, a human can estimate the next phase with reasonable accuracy. The scope is no longer theoretical — it’s been discovered through making.
Phase 3 — AI-Assisted Build (Hybrid): Build with guardrails. Use AI for speed, humans for judgment. Track estimate vs. delivery to catch deviations early.
This is fundamentally different from the traditional “estimate everything upfront, then build” model. Agencies that try to force-fit the old model onto AI-assisted work will either lose money on fixed-price projects or scare clients away with inflated buffers.
What This Means for Agency Pricing
If you’re a software agency using AI agents in your delivery, you need to answer a different question than before. Not “how do we limit scope?” but “what’s worth exploring next?”
Here’s a practical framework for pricing in the agent era:
-
Discovery is a billable phase, not a freebie. The AI will surface things the client didn’t know they needed. That’s value, not scope creep. Price it as Phase 1.
-
Build in an exploration buffer. If you’re estimating a fixed-price project, add 20-30% for emergent discovery. You’ll need it.
-
Track estimate vs. delivery obsessively. When the AI surfaces something unexpected, you need to know immediately whether you’re inside or outside the agreed scope. Tools like Apropo’s estimate-vs-delivery tracking turn subjective “scope creep” arguments into data-driven conversations.
-
Educate your clients upfront. Tell them: “We use AI agents in our process. This means we’ll discover things during the project that neither of us knew about at the start. That’s a feature, not a bug — but it means we need to agree on how we handle those discoveries before they happen.”
The Bottom Line
Scope creep wasn’t always the enemy. It became the enemy because iteration was expensive. AI agents have collapsed that cost.
Agencies that embrace “scope creep as discovery” — and build their pricing models to accommodate it — will deliver more value than those that cling to fixed-scope illusions.
The ones that thrive won’t be the agencies using the most AI. They’ll be the ones that know exactly when to let the agent explore and when to put boundaries in place. And they’ll have the data to prove their estimates were right — even when the scope changed.
This article draws on the experience of Mike Waszazak, documented in his January 2026 piece “Scope Creep as Discovery” (Bootcamp/Medium), and applies his insights to the specific challenges of agency estimation and pricing.