Artificial intelligence project management vs. traditional software

Imagine this scenario: your company decides to implement an Artificial Intelligence (AI) system to optimize a critical process. The project manager, accustomed to large-scale implementations of traditional software, opens the schedule, defines a closed scope, ties the deliverables together in a linear Kanban board, and promises maximum accuracy in record time.

IA June 2, 2026 6 min read By: Skyone

If you're managing your data and artificial intelligence projects the same way you manage your ERP or legacy software projects, I have a reality check: your schedule is lying to you.

In the world of traditional technology, success is usually binary: the code works or it doesn't, the screen goes up or it doesn't. But when we cross the border into the world of AI, the logic changes drastically. We leave the realm of determinism and enter the field of probability, where one plus one doesn't always equal two.

To understand why so many AI projects fail before they even leave the drawing board, and how technology leaders can overcome this challenge, we've compiled the key practical lessons discussed on the Builders Podcast by experts who are experiencing this transition firsthand.

The reality check: why do traditional methods fail in AI?

Market data is unforgiving. Research from consulting firms like McKinsey indicates that around 70% to 85% of AI projects fail or don't make it past the first testing phase (the famous Proof of Concept or PoC).

The reason for this alarming rate is not a lack of cutting-edge technology, but rather a profound misalignment of expectations and methodologies. AI development more closely resembles a scientific and experimental process than traditional software engineering.

The change in the "definition of done"

In traditional software projects or ERP modules, the definition of "done" is clear: the workflow moves from development to validation and culminates in the deployment of structured code.

In Artificial Intelligence, the definition of "ready" is the validation of a hypothesis. The process is inherently cyclical and iterative: implementation, testing, recall analysis ( a metric that assesses how much the model is making mistakes or needs further testing), and then returning to the beginning. Often, at the start of the journey, both the developer and the client simply don't know what the final result will be.

The 3 customer profiles (and the risks to the project)

Dealing with AI requires surgically managing human expectations. In the business ecosystem, three typical stakeholder that project leaders need to learn how to manage:

The miracle believer: This is the enthusiastic profile of someone who has heard about the potential of technology in the market and believes that implementing AI will be quick, easy, and magical. They have no idea of the amount of work and internal structuring required on their own end for the model to yield results.
The skeptic of immediate ROI: believes in the potential, but demands an extremely fast Return on Investment (ROI), without understanding the evolutionary timeline of AI. Machine learning models require a mature journey: you gain 30% efficiency first, then 60%, until you reach maturity and excellence.
The political surfer: wants to implement the technology at any cost just to sell internally or to the market that the company "has AI." This is considered the most dangerous profile, as they don't know exactly what they want the solution for, ignoring technical limitations and generating very high levels of future frustration.

Promising 100% accuracy in an AI project on "Day 1" isn't a sales strategy; it's asking for a job. Since we're dealing with probabilistic models, a realistic approach from the very beginning is vital.
Expert's note

The hidden raw material: is your data ready?

The major dividing line between the success and failure of an AI agent or model is the quality of the data. The technology will read and learn strictly based on the information it is given.

The most common scenario in companies is that the client requests a complex solution and, when questioned about the database, discovers that it doesn't exist, is incomplete, or completely unstructured. For example: if you create a model designed to address Human Resources (HR) pain points and feed it with an inconsistent database, or try to extract financial answers from it for which it wasn't trained, the delivery will fail.

Before designing robots or complex workflows, put the ball on the ground and audit your raw material: business data.

How to adapt management: CAPEX, OPEX and Dual-Track Agile

To prevent time and resource management from being stifled by the traditional deterministic model, large companies utilize real-time adaptation frameworks:

Accounting separation (CAPEX vs. OPEX): More mature markets make a clear distinction. The data structuring, research, experimentation, and initial model training phase is treated as CAPEX (investment). From the moment the model is established, monitored, and generating scalable gains, the project enters the OPEX (operating cost) phase.
Dual-track framework: using the agile model divided simultaneously into two tracks: Discovery (experimentation) and Delivery (value delivery focused on software).
Value stream mapping: focusing on continuous improvement and constant process refinement, rather than viewing errors as definitive failures. In the age of AI, learning from anomalous model behavior is part of the evolution cycle.

Conclusion: AI doesn't replace people, it replaces activities

Successful Artificial Intelligence projects are not created to replace the human factor, but rather to eliminate operational bottlenecks and 100% manual or bureaucratic activities. By freeing professionals from repetitive tasks, such as spending hours manually formatting reports or presentations, leaders gain valuable time to focus on strategy, interpersonal relationships, and new business opportunities.

Managing AI requires letting go of static scope and embracing probabilistic cadence. Your timelines will only stop lying when they reflect hypotheses, structured data, and real-world experimentation.

Do you want to thoroughly master management methodologies for the new technological age?

Listen to the full debate and check out all the hacks, HR/Finance database examples, and practical lessons shared by Bruno Marcos (Data Engineering Coordinator at Skyone) and Sidney Rocha (Services Director at Skyone) in this unmissable episode.

🎧 Click here to listen to the full episode of the Builders Podcast on Spotify!