Software Project Estimation in the AI Era: From Waterfall to Intelligent Development

Welcome back to another Kmino blog post!

Today, we’re tackling one of the most challenging aspects of software development: project estimation. Whether you’re estimating your first feature or planning a multi-month project, getting estimations right can make or break project success.

As Philip Metzger notes in “Managing a Programming Project”, estimation is fundamentally about understanding both the technical work and the human factors that influence productivity. Combined with today’s AI tools, estimation has become both more powerful and more complex.

From Waterfall to AI: The Evolution of Estimation

Traditional Approaches

For decades, software estimation followed predictable patterns rooted in the waterfall methodology. We’d spend weeks creating detailed Work Breakdown Structures, carefully writing down every task and subtask. We used to rely heavily on metrics like Function Points or Lines of Code, applying mathematical formulas that promised precision but often delivered disappointment.

This approach worked reasonably well for predictable projects with well-understood requirements. However, it suffered from a fundamental flaw: the assumption that software development could be planned like construction projects. Teams would over-specify requirements upfront, spending weeks and even months in analysis paralysis, only to discover that their carefully crafted plans crumbled upon contact with real-world complexity.

The Agile Shift

Agile methodologies revolutionized estimation by embracing uncertainty rather than fighting it. Instead of trying to predict every detail months in advance, teams adopted iterative planning cycles that refined estimates based on actual experience. Story points replaced time-based estimates, acknowledging that relative complexity often mattered more than absolute duration.

As Andy Grove emphasizes in “High Output Management,” the key breakthrough was creating feedback loops to continuously improve predictions based on actual team performance. Teams began tracking their velocity/how much work they could realistically complete in a given timeframe and used this data to make increasingly accurate forecasts.

Enter the AI Revolution

Today, AI tools like Claude and ChatGPT are fundamentally changing the economics of software development. With that in mind, we should be altering the entire “cost structure” of different types of work. A junior developer can now implement complex algorithms they couldn’t have written from scratch, while senior architects can rapidly prototype multiple solutions to compare their metrics.

This transformation creates both unprecedented opportunities and entirely new challenges for estimation. Teams are discovering that their carefully collected historical data from pre-AI development may no longer predict future performance accurately.

AI’s Impact on Modern Estimation

The Acceleration Revolution

AI tools don’t just help write code, they fundamentally alter the economics of software development. Consider a typical API endpoint that once required a developer to manually write the controller logic, input validation, database queries, error handling, and tests. This work might have taken multiple days for an experienced developer, including time for research, implementation, and testing.

With AI assistance, the same developer can describe the endpoint requirements in natural language and receive a complete implementation in minutes. The AI generates not just the core functionality, but comprehensive error handling, input validation, and even unit tests that cover edge cases the developer might have overlooked.

However, this acceleration isn’t uniform across all types of work. While boilerplate code generation has become almost instantaneous, the time required for architectural decisions, stakeholder communication, and complex debugging remains largely unchanged. This creates a new estimation challenge: understanding which tasks benefit from AI acceleration and which remain human dependent.

The productivity boost also varies a lot based on developer experience and the specific AI tools being used. A junior developer might see their productivity increase by 300% on routine tasks, while a senior developer working on complex system integration might see only marginal improvements.

Understanding AI Impact Categories

To estimate effectively in the AI era, teams need to think differently about categorizing work. Traditional estimation often focused on technical complexity or business value. Now, we must also consider how well AI tools can assist with specific types of tasks.

High AI Impact work

Represents tasks where AI tools can reduce development time by 50-80%. These typically involve pattern-based coding where AI can easily recognize common implementations and generate appropriate solutions. CRUD operations exemplify this category perfectly: Creating, Reading, Updating, and Deleting data follows well-established patterns that AI models have seen thousands of times during training.

Unit test creation has become another area where AI provides dramatic acceleration. AI can analyze a function’s implementation and generate comprehensive test cases, including edge cases and error conditions that developers might miss. What once required hours of careful test design can now be accomplished in minutes, though developers still need time to review and refine the generated tests.

Documentation writing, the “biggest pain” of developer productivity, has been transformed by AI assistance. AI can analyze code and generate clear, comprehensive documentation that explains not just what the code does, but why certain design decisions were made. This has turned documentation from a time-consuming chore into a quick review-and-refine process.

In Lunos we were able to test multiple architectures during our smart contract redesign, and that was mostly done through AI-generated unit tests and document writing. We laid down all user flows and Claude suggested multiple approaches for each one of our edge cases. Unit tests that would have taken weeks to write were completed in a couple of days.

Medium AI Impact tasks

See modest improvements, typically reducing development time by 20-40%. Business logic implementation falls into this category because while AI can suggest implementation approaches, developers must carefully review and adapt the suggestions to match specific business requirements.

Third-party service integrations benefit from AI assistance, particularly when working with well documented APIs that AI models have encountered during training. However, these integrations often require understanding nuanced business requirements and handling edge cases that AI might not anticipate, limiting the acceleration potential.

Performance optimization represents an interesting middle ground. AI can suggest optimization strategies and even implement specific improvements, but developers must understand the broader system context to evaluate whether these optimizations are appropriate and beneficial.

Still talking about Lunos, we used Claude to double-check if all of our user flows were correctly implemented. To do that, we asked it to create a document that described all the user flows based on the smart contracts we had. After the document was completed, we were able to verify if any of the business logic was wrongly implemented.

Low AI Impact tasks

Remain largely human dependent, with AI providing minimal time savings. System architecture decisions require deep understanding of business context, technical constraints, and long-term maintenance implications where AI assistance remains limited.

Complex debugging, particularly issues involving system interactions or subtle timing problems, continues to require human expertise. While AI can suggest potential causes and debugging approaches, the detective work of identifying root causes remains primarily developer-driven.

User experience design and stakeholder communication represent areas where human insight, empathy, and context understanding remain irreplaceable, though AI tools can assist with research and documentation tasks.

Practical AI-Era Estimation Techniques

The AI-Aware Three-Point Method

Traditional three-point estimation asked developers to provide optimistic, most likely, and pessimistic scenarios for each task. This approach acknowledged uncertainty while providing a framework for systematic thinking about risk and complexity. In the AI era, we need to evolve this technique to account for the new realities of AI-assisted development.

The updated approach considers three distinct scenarios that reflect different AI engagement levels.

AI-Optimistic scenario

Assumes everything goes perfectly with AI assistance: the tools are responsive, the generated code is immediately usable, and integration proceeds smoothly. This represents the best case outcome when AI tools function at their peak effectiveness.

AI-Realistic scenario

Accounts for the typical AI development experience, including time for prompt refinement, code review and adjustment, and occasional AI limitations. This scenario recognizes that while AI provides significant assistance, developers still need time to integrate, test, and refine AI-generated solutions.

Traditional/Realistic scenario

Serves as a fallback estimate, representing how long the task would take using conventional development approaches. This scenario becomes crucial when AI tools experience outages, prove unsuitable for the specific task, or when the team lacks experience with AI-assisted development for that type of work.

Consider estimating a user authentication feature. In the AI-Optimistic scenario, you might estimate one day, assuming AI generates most of the implementation including security best practices, database integration, and comprehensive tests. The AI-Realistic scenario might extend to 2.5 days, accounting for time to review AI suggestions, customize them for your specific requirements, and handle integration challenges. The Traditional/Realistic scenario might be five days, representing manual implementation from scratch.

Using the formula (AO + 2×AR + TR) / 4, this gives a final estimate of 2.5 days, appropriately weighted toward the AI-assisted reality while maintaining a safety buffer for potential AI limitations.

Developer Experience and AI Adoption

Following Andy Grove’s principle of understanding individual capabilities, modern estimation must account for how different experience levels interact with AI tools. We must think about how effectively developers can leverage AI assistance while maintaining code quality and system coherence.

Junior developers often see the most dramatic productivity gains from AI tools, typically achieving 40-70% time reductions on suitable tasks. AI helps them implement functionality they couldn’t create independently, reducing time spent researching syntax and common patterns. However, junior developers may need additional time for code review and learning, as they’re still developing the judgment to evaluate AI suggestions effectively.

Mid-level developers experience balanced productivity gains, usually seeing 30-50% time reductions on suitable tasks. They’re typically good at recognizing when AI suggestions need modification and can quickly adapt generated code to fit existing architectures. Their experience helps them craft better prompts and identify potential issues before they become problems.

Senior developers often show more modest but strategic gains, typically 20-40% time reduction on suitable tasks. While they could implement most functionality manually, AI allows them to focus their expertise on architectural decisions and complex problem-solving. Senior developers often excel at crafting sophisticated prompts that generate high-quality code, but they also spend more time reviewing AI output for subtle issues and ensuring alignment with long-term maintainability goals.

The Historical Data Challenge

As Philip Metzger emphasized in “Managing a Programming Project”, historical data forms the foundation of accurate estimation. Teams that systematically track actual versus estimated time can calibrate their future predictions with increasing accuracy. However, the AI revolution has potentially invalidated much of this carefully collected data.

A team’s velocity measurements from 2023 may no longer predict their 2025 performance when they’re using advanced AI tools. The tasks that historically took the longest might now be completed quickly with AI assistance, while previously simple tasks might require new skills like prompt engineering and AI code review.

This creates both a challenge and an opportunity. Teams must start collecting new baseline data that reflects AI-assisted development, but they can also use this transition period to develop more sophisticated tracking systems. Modern teams should separate AI-suitable tasks from “AI-resistant” tasks in their velocity tracking, providing more nuanced data for future estimation.

Navigating AI-Era Estimation Challenges

The “AI Will Handle It” Misconception

One of the most common estimation pitfalls in the AI era comes from stakeholder expectations that AI acceleration applies universally across all development work. Business leaders, excited by stories of dramatic productivity improvements, often assume that AI tools will reduce timelines across entire projects proportionally. This misconception can lead to unrealistic deadline pressure and project planning that doesn’t account for the reality of AI assisted development.

The truth is more complex. While AI excels at predictable, pattern-based tasks, it struggles with work that requires deep business context, creative problem-solving, or complex system integration. A project might see 70% time reduction on CRUD operations and documentation, but zero improvement on user experience design and stakeholder alignment sessions. The overall project timeline might improve by 25-30%, not the 70% that initial AI success stories might suggest.

Effective estimation communication requires educating stakeholders about this reality. Rather than promising uniform acceleration, break down estimates by AI impact categories and explain where improvements will and won’t occur. This sets realistic expectations while highlighting the genuine value AI provides.

AI Reliability and Contingency Planning

AI tools, while powerful, introduce new dependencies and potential failure modes that traditional estimation approaches don’t account for. API rate limits can slow development during high usage periods, context window limitations can prevent AI from understanding large codebases, and service outages can force teams back to traditional development methods.

Additionally, AI-generated code quality can be inconsistent. While AI might produce excellent boilerplate code, it occasionally generates solutions that look correct but contain bugs or security vulnerabilities. These issues require human detection and correction, adding time that’s easy to underestimate.

Building contingency planning into the “new AI era” means identifying critical path tasks that depend heavily on AI assistance and ensuring fallback approaches exist. It also means allocating appropriate time for thorough code review of AI-generated solutions, particularly for security-sensitive or business-critical functionality.

The Code Review Evolution

AI-generated code requires fundamentally different review approaches than human-written code. Traditional code review focused on logic errors, style consistency, and maintainability concerns. AI code review must additionally verify that the generated solution actually solves the intended problem, doesn’t introduce unexpected dependencies, and aligns with the broader system architecture.

This expanded review scope typically requires more time than traditional code review, particularly for complex AI-generated solutions. Teams often underestimate this additional review time, leading to estimate overruns when AI assistance is extensive. Experienced teams build this extended review time into their estimates, treating it as a necessary investment in maintaining code quality and system coherence.

Building AI Estimation Maturity

Following Grove’s management philosophy of systematic improvement:

Phase 1: Baseline Establishment (Months 1-3)

• Track AI tool usage alongside time estimates
• Identify team members and tasks with greatest AI acceleration
• Categorize work by AI impact potential

Phase 2: Calibration (Months 3-6)

• Use data to refine AI impact factors
• Develop team-specific AI workflows
• Create estimation templates including AI considerations

Phase 3: Optimization (Months 6+)

• Implement dynamic estimation adjustments
• Train stakeholders on new estimation realities
• Develop predictive models for AI effectiveness

Closing Thoughts

The New Estimation Reality

As Metzger noted decades ago, successful project estimation requires understanding both technical complexity and human factors. AI adds a new dimension: we’re not just estimating development time, but optimizing the balance between human insight and AI acceleration.

Key Principles

Following Grove’s focus on measurable outcomes:

• Treat estimates as predictions with uncertainty, not commitments
• Categorize tasks by AI impact and adjust accordingly
• Build new historical databases that account for AI-assisted development
• Create feedback loops to continuously improve AI impact predictions

The goal isn’t perfect estimates; it’s estimates that improve decision-making while adapting to rapidly evolving AI capabilities.

Conclusion

tl;dr:

• AI is revolutionizing estimation: Traditional historical data needs updating as AI tools change development velocity
• Use AI-aware three-point estimation: Factor in AI Optimistic (AO), AI Realistic (AR), and Traditional/Realistic Scenarios (TR)
• Categorize by AI impact: High impact tasks (CRUD, testing) vs low impact tasks (architecture, debugging) need different approaches
• Account for experience levels: Junior devs see much greater gains than seniors on suitable tasks
• Plan for AI limitations: Tool reliability, context constraints, and additional review time

The future of software estimation lies in understanding both timeless human factors and rapidly evolving AI capabilities.

For more insights on software engineering best practices, make sure to follow us on LinkedIn, and see you at the next one!