Estimating AI projects is notoriously difficult. Traditional software estimation techniques assume that the problem is well-defined and the implementation path is largely known. AI projects have fundamental uncertainties: Will the data be sufficient? Will the model achieve acceptable accuracy? How long will experimentation take? These unknowns make standard estimation approaches unreliable. This guide presents techniques that account for AI-specific uncertainty.

Why AI Estimates Are Usually Wrong

Several structural factors make AI estimation hard:

Research uncertainty. The core modeling work is experimental. You might find a working approach on day one or spend weeks exploring dead ends. This variance is inherent to the discipline, not a planning failure.

Data surprises. Every AI project discovers unexpected data issues. Fields that should be populated are empty. Data that should be consistent has three different formats. Sources that should be accessible require months of procurement. These surprises add weeks to timelines.

Integration complexity. Connecting an AI model to production systems is harder than it looks. Real-time latency requirements, error handling, fallback logic, monitoring, and graceful degradation all add work that is easy to underestimate.

Moving targets. Stakeholders refine requirements after seeing early results. “Actually, the model needs to handle images too” or “We need the response time under 200ms” can invalidate weeks of work.

Estimation Techniques

Three-Point Estimation

For each major work item, estimate three scenarios:

  • Optimistic (O): Everything goes well. Data is clean, first approach works, integration is straightforward.
  • Most likely (M): Normal challenges. Some data cleanup needed, second or third approach works, moderate integration effort.
  • Pessimistic (P): Significant obstacles. Major data issues, extensive experimentation needed, complex integration.

Expected duration = (O + 4M + P) / 6

This formula, from PERT estimation, weights the most likely scenario while acknowledging the tails. For AI projects, the pessimistic scenario should be genuinely pessimistic - not worst-case but “things go badly.”

Example for a document classification model:

  • Optimistic: 6 weeks (clean data, proven approach)
  • Most likely: 10 weeks (moderate data work, some experimentation)
  • Pessimistic: 18 weeks (major data issues, multiple approach pivots)
  • Expected: (6 + 40 + 18) / 6 = 10.7 weeks

Phase-Based Estimation

Break the project into standard phases and estimate each:

Discovery and feasibility (2-4 weeks). Data assessment, technical spike, architecture design. This phase is relatively predictable.

Data preparation (3-8 weeks). Highly variable. Depends on data quality, volume, and the number of sources. Always longer than expected.

Model development (4-12 weeks). The most uncertain phase. Bound it with time caps: “We will invest up to 8 weeks in model development. At week 4, we assess progress and decide whether to continue or simplify.”

Production engineering (4-8 weeks). Deploying the model, building APIs, implementing monitoring, setting up CI/CD. More predictable if the team has done it before.

Validation and launch (2-4 weeks). User acceptance testing, stakeholder review, gradual rollout. Often compressed under schedule pressure, which is a mistake.

Total range for a typical enterprise AI project: 15-36 weeks. This range is wide because uncertainty is genuinely high.

Reference Class Forecasting

Look at how long similar AI projects actually took, not how long they were planned to take. If your organization or industry peers have completed similar projects, their actual timelines are the best predictor of yours.

Common reference classes:

  • Document processing and extraction: 3-6 months
  • Recommendation systems: 4-8 months
  • Customer churn prediction: 2-4 months
  • Conversational AI (chatbot): 4-9 months
  • Computer vision (custom model): 4-8 months

These ranges assume a competent team with reasonable data availability. Adjust upward for first-time teams or significant data challenges.

Budget Estimation

AI project costs fall into several categories:

Personnel. The largest cost. An AI team of 3-5 people (ML engineer, data engineer, software engineer, part-time PM and domain expert) costs $100K-$200K per month fully loaded in most markets.

Compute. Training costs vary enormously. Fine-tuning a small model might cost $100; training a large custom model can cost $10K-$100K+. Production inference costs depend on volume and model size. Budget $2K-$20K/month for compute during development, and model production costs based on expected query volume.

Data. Data acquisition, labeling, and storage. Human labeling costs $0.05-$5.00 per item depending on complexity. A 10,000-item labeled dataset might cost $500-$50,000.

Tools and services. Vector databases, monitoring tools, MLOps platforms, API services. Budget $1K-$5K/month for tooling.

Contingency. Add 25-40% contingency for AI projects. This is not padding - it is realistic accounting for the uncertainty inherent in the work.

Presenting Estimates to Stakeholders

Never give a single number. Always present a range. “We estimate 12-18 weeks” communicates uncertainty honestly.

Explain what drives the range. “The low end assumes our data is clean enough to use directly. The high end accounts for significant data preparation work. We will know which scenario we are in after the 2-week discovery phase.”

Identify decision points. “After the feasibility spike in week 3, we will have a much tighter estimate. I recommend a go/no-go decision at that point.”

Separate confidence levels. “We are 90% confident we can deliver the MVP in 16 weeks. We are 50% confident we can hit the stretch accuracy target in that timeframe.”

Estimation Anti-Patterns

Anchoring on the first number. Once someone says “8 weeks,” all subsequent estimates orbit that number. Start estimation by listing unknowns, not by guessing durations.

Estimating without a spike. Providing timeline estimates before doing any technical investigation. Always invest a few days in a feasibility assessment before committing to a project timeline.

Forgetting the non-model work. Model development is typically 30-40% of total project effort. Data preparation, production engineering, testing, and deployment are the majority. Estimates that focus only on model work will be dramatically low.

Assuming linear scaling. “The pilot took 4 weeks for one use case, so five use cases will take 20 weeks.” In practice, additional use cases share infrastructure but each has unique data challenges. Estimate each independently.

The most important estimation principle for AI projects: be honest about what you do not know, build in decision points where you learn more, and update estimates as uncertainty reduces.