Comparing the Use of Agile in Traditional IT and AI Projects - MICHAŁ OPALSKI / AI-AGILE.ORG

Agile has been a cornerstone of modern project management in the IT industry for over two decades. Since the Agile Manifesto was introduced in 2001, it has revolutionized how teams work by promoting iteration, close collaboration with clients, and adaptability to change. Today, as artificial intelligence (AI) emerges as one of the most influential fields in technology, a pressing question arises: does Agile work as effectively in AI projects as it does in traditional IT initiatives?

This article provides a detailed comparison between traditional IT projects and AI projects from an Agile perspective, exploring their similarities, differences, challenges, and real-world examples.

Agile in Traditional IT

In classical IT projects—such as developing web applications, ERP systems, or mobile solutions—Agile has become the de facto standard. Development teams frequently rely on frameworks such as Scrum, Kanban, or SAFe to accelerate delivery, mitigate risk, and ensure continuous value creation.

Key characteristics of Agile in traditional IT include:

• Clear business requirements – Product backlogs are usually based on relatively stable user expectations.

• Short iteration cycles – Each sprint produces a working increment of software.

• Straightforward testing – Unit tests and automated QA processes provide quick verification of correctness.

• Incremental value delivery – Features can be introduced in small, functional segments that benefit the end user immediately.

Agile thrives in traditional IT because software development processes are generally predictable and well supported by established practices and tools.

Example:

A retail company commissioning a new e-commerce platform can plan features such as shopping carts, payment integration, and search functionality in a series of sprints. Each sprint delivers a tangible piece of functionality that stakeholders can review, test, and refine. This predictability and steady progress align perfectly with Agile’s principles.

Agile in AI Projects

AI projects differ fundamentally from traditional IT projects. Instead of rule-based programming and predictable business logic, teams work with data, statistical models, and learning algorithms. Applying Agile to AI introduces unique challenges that traditional IT teams seldom encounter.

Key characteristics of Agile in AI include:

• Experimental nature – Product backlogs often contain hypotheses rather than defined requirements. For example: “Test whether Model X improves classification accuracy by 10%.”

• Unpredictable outcomes – A sprint may end not with a deployable model but with the conclusion that the dataset is insufficient or that the algorithm performs poorly.

• Strong dependency on data – In traditional IT, code is the fuel. In AI, data is the fuel. The quality, completeness, and accessibility of data determine project success.

• Testing complexities – Verification relies on performance metrics (accuracy, precision, recall, F1-score) rather than binary pass/fail results.

• Interdisciplinary teams – AI projects often include data engineers, machine learning specialists, analysts, and domain experts, requiring additional coordination and shared understanding.

Example:

Consider a healthcare startup building a diagnostic AI to detect early signs of lung cancer from CT scans. Unlike a web app, progress is not measured by building “login functionality” or “search.” Instead, a sprint might be dedicated to preprocessing thousands of images, experimenting with convolutional neural networks, or improving recall for rare cases. The outcome may be inconclusive, requiring multiple cycles before any usable functionality emerges.

Similarities and Differences

Aspect	Traditional IT	AI Projects
Requirements	Relatively stable	Often experimental, framed as hypotheses
Iterations	Each sprint delivers working software	Iterations may end with inconclusive results
Testing	Automatable, rule-based	Metric-driven, probabilistic outcomes
Dependencies	Mainly code and technologies	Data quality and availability are critical
Teams	Developers, QA, business analysts	Interdisciplinary: ML, data, domain experts

How Agile Needs to Adapt for AI

To make Agile effective in AI contexts, certain adaptations are necessary:

• Hypothesis-driven backlogs – Backlogs should contain testable assumptions rather than rigid user stories.

• Research-oriented iterations – Teams must accept that not all sprints will produce deployable code but may instead yield valuable insights.

• Redefined “Definition of Done” – In AI, “done” might mean “tested three model architectures and documented results” rather than “feature is live in production.”

• Stronger focus on MLOps – Continuous integration and delivery must be extended with monitoring of data pipelines and deployed models.

• Transparent communication with stakeholders – Business partners must understand that progress in AI is less predictable than in traditional IT.

Example:

In a financial services company developing a fraud detection AI, the backlog might include user stories such as “Investigate whether adding transaction geolocation data improves detection accuracy.” The sprint goal could be purely exploratory. Success might mean learning that the improvement is negligible—valuable knowledge that prevents wasted investment in unnecessary data acquisition.

Case Study 1: Traditional IT – Banking App

A major bank adopts Scrum to deliver a new mobile banking app. Requirements such as “allow users to transfer money” or “add fingerprint login” are well understood. Each two-week sprint produces fully functional features that are tested and integrated into the live app. By release, the product meets user expectations and regulatory compliance requirements, and Agile enables fast adaptation to minor changes like UI adjustments or security updates.

Case Study 2: AI – Predictive Maintenance in Manufacturing

A manufacturing company wants to use AI to predict when machines will fail. The project backlog includes items such as “clean historical sensor data,” “test random forest model,” and “evaluate LSTM neural networks.” After several sprints, the team learns that many sensors produce inconsistent data, delaying progress. The Agile process helps the team adapt by prioritizing data engineering tasks before model development. Instead of shipping a finished feature each sprint, the value is in building cumulative knowledge that eventually results in a predictive system saving the company millions in downtime costs.

The Human Factor in Agile AI

One of the most underestimated differences between traditional IT and AI projects lies in human collaboration. Agile thrives on cross-functional teams, but in AI projects, diversity of expertise is even more pronounced. Data scientists often think in terms of experiments and models, while business stakeholders expect concrete outputs.

Bridging this gap requires:

• Regular knowledge-sharing sessions.

• Visualization of metrics in ways business users can understand.

• Education of stakeholders on the probabilistic nature of AI results.

Example:

In a customer support chatbot project, the data science team may report an “F1-score of 0.72.” Business stakeholders, however, may not understand what this means in practical terms. Translating it into “the bot correctly answers about 7 out of 10 queries, but struggles with rare cases” aligns expectations and maintains trust.

Agile Beyond Development: Continuous Learning in AI

In traditional IT, once software is deployed, Agile ensures ongoing improvements. In AI, deployment is just the beginning. Models degrade over time due to data drift (changes in input data patterns) or concept drift (shifts in relationships between inputs and outputs).

Agile for AI therefore extends beyond development:

• Monitoring models in production is as important as initial training.

• Retraining cycles must be part of the backlog.

• Feedback loops from end users and systems provide continuous learning opportunities.

Example:

An e-commerce platform deploying a recommendation engine must monitor whether suggested products remain relevant as new inventory is added and customer behavior evolves. Agile ceremonies such as sprint reviews and retrospectives can incorporate model performance discussions to ensure continuous alignment with business goals.

Conclusion

Agile remains a highly valuable methodology, but applying it to AI projects requires greater flexibility and an appreciation of AI’s experimental nature. While traditional IT leverages Agile for predictable, incremental delivery, AI projects shift the focus toward experimentation, data quality, and iterative learning.

In essence, Agile in AI is a research-oriented adaptation of Agile—less about delivering features in every sprint and more about building knowledge that leads to long-term breakthroughs. Organizations that embrace this shift and tailor Agile practices to the realities of AI stand to gain a competitive advantage in a rapidly evolving technological landscape.