The difference between an AI project that delivers and one that stalls usually isn't the technology. It's the scoping. Too narrow and you build something nobody uses. Too broad and you never finish. Here's how we scope AI projects that actually get to production.
What You Need to Know
- AI project scoping is fundamentally different from traditional software scoping. Uncertainty is higher, the solution space is wider, and you can't fully specify the output before you start.
- The biggest scoping mistake is skipping discovery. Spending 2-4 weeks on structured discovery saves months of building the wrong thing.
- Budget for iteration, not just delivery. AI projects need tuning, testing, and refinement that traditional software projects don't. Build 30-40% contingency into timelines.
- Define success criteria before you start, not after. And make them business outcomes, not technical metrics. "Reduce claims processing time by 40%" not "achieve 95% accuracy on test data."
65%
of enterprise AI projects exceed their initial timeline by more than 50%
Source: McKinsey, The State of AI in 2024
The Scoping Framework
We scope every AI engagement around five dimensions. Miss any one of them and you're building risk into the project from day one.
1. Problem Definition
This sounds obvious. It isn't. "We want to use AI for customer service" is not a problem definition. "Our customer service team spends 60% of their time answering the same 20 questions, and response times average 4 hours" is a problem definition.
Good problem definitions are specific, measurable, and tied to a business outcome. They also identify who experiences the problem and how often.
The question to ask: "If we solve this perfectly, what specific metric changes and by how much?"
2. Data Assessment
Every AI project depends on data. Before committing budget, you need to know:
- What data exists? Documents, databases, APIs, spreadsheets, emails, chat logs.
- What state is it in? Clean and structured, or messy and inconsistent? (Usually the latter.)
- Can we access it? Technical access, legal access, and practical access are three different things.
- Is there enough? For RAG-based systems, you need a knowledge base. For ML models, you need training data. The quantity and quality requirements are different for each approach.
We've walked away from projects at this stage. Not because the problem wasn't real, but because the data foundation wasn't there and the investment to build it exceeded the project budget.
3. Solution Architecture
This is where you decide the "how," and it's where AI scoping diverges most from traditional software. The solution space for an AI problem is typically wider than for a traditional software problem.
A customer service AI could be: a RAG-based chatbot, an agent-assisted tool that drafts responses for humans to review, an automated triage system that routes queries, or a knowledge base that helps agents find answers faster. Each has different complexity, different accuracy profiles, and different adoption implications.
The question to ask: "What's the simplest approach that would deliver 80% of the value?" Start there. Complexity is always available later.
4. Team and Timeline
AI projects need a different team composition than traditional software:
- AI/ML engineering for model integration, prompt engineering, and pipeline development
- Data engineering for data pipelines, quality, and integration
- Domain expertise for validation, testing, and defining "good enough"
- Product/UX for the interface layer, because the best AI with a bad interface doesn't get used
- Change management for adoption, which is the part that gets cut first and matters most
Timeline-wise, we structure AI projects in phases:
| Phase | Duration | Output |
|---|---|---|
| Discovery | 2-4 weeks | Problem validation, data assessment, solution recommendation |
| Foundation | 4-8 weeks | Core infrastructure, data pipelines, initial model integration |
| Capability | 4-8 weeks | Working AI capability with user interface |
| Refinement | 2-4 weeks | Tuning, testing, edge case handling |
| Deployment | 2-4 weeks | Production deployment, monitoring, handover |
Total: 14-28 weeks for a single AI capability. Significantly longer than most stakeholders expect.
5. Success Criteria
Define these before you write a line of code. They should be:
Business outcomes, not technical metrics. "Reduce average claims processing time from 45 minutes to 15 minutes" is a success criterion. "Achieve 95% accuracy on the test set" is an intermediate technical metric that may or may not correlate with business value.
Measurable with existing instrumentation. If you can't measure the current state, you can't measure improvement. Sometimes the first investment is in measurement capability.
Time-bound. "Deliver value within 90 days of deployment." Not "eventually deliver value."
The 90-Day Rule
If an AI project can't demonstrate measurable business value within 90 days of deployment, something is wrong with the scoping, the implementation, or the problem selection. This doesn't mean full ROI in 90 days. It means visible, measurable progress toward the target outcome.
Common Scoping Mistakes
Scoping for the demo, not production. A demo that works on 10 curated examples is not evidence that the system will work on 10,000 real-world examples. Scope for production data quality and volume.
Underestimating data work. In our experience, data preparation and pipeline development consume 40-60% of total project effort. If your scope allocates 10% to data, it's wrong.
Ignoring adoption. The best AI system with 10% adoption delivers less value than a mediocre system with 80% adoption. Budget for change management, training, and iterative UX improvement.
Scope creep via "while we're at it." AI projects attract feature requests like magnets. Every stakeholder sees a potential use case. A clear, documented scope with an explicit "not in scope" section is essential.