Productivité & Carrière

Get Started in AI Project Management: Your First Concrete Step Today

AI Project Management: The Essentials in One Article — Real Code, Diagrams, and Concrete Steps, Excerpts from a 35-Lesson Course.

REHOUMA Haythem

• 13 min read
Get Started in AI Project Management: Your First Concrete Step Today

The best way to learn AI Project Management is by doing. This article gives you a leg up with practical excerpts from a 35-lesson course — enough to get your first result today.

tl;dr
  • Introduction to AI PM
  • AI Project Lifecycle
  • Scoping and Business Case
  • Building an AI Team
  • Agile Methods for AI
~$ cat ./parcours.md # AI Project Management — 10 chapters
01
Introduction to AI PM
→ Course presentation→ Who does what? Overview of AI roles+ 1 more lessons
02
Lifecycle of an AI project
→ From idea to production→ MLOps vs classic DevOps+ 1 more lessons
03
Framing and business case
→ Finding the right problem→ ROI and value measurement+ 1 more lessons
04
Building an AI team
→ Data Scientist, ML Engineer, PM roles→ Organizing a multidisciplinary team+ 1 more lessons
05
Agile methods for AI
→ Scrum applied to AI→ Research sprints vs delivery sprints+ 1 more lessons
06
Data management
→ Data governance and GDPR→ Data quality and versioning+ 1 more lessons
07
MLOps from a PM perspective
→ The ML pipeline seen by a PM→ Monitoring models in production+ 1 more lessons
08
KPIs and value measurement
→ Technical vs business KPIs→ A/B testing for AI models+ 1 more lessons
🏁
Final project (+ 2 chapters along the way)
→ You leave with a concrete, demonstrable project

Data Catalog and Controlled Access

NOTEObjective — Understand what a Data Catalog is, why it saves AI projects in large enterprises, and how to manage data access securely without becoming a bottleneck.

The problem: where is the data?

In a 500+ employee company, you can easily count 50 to 500 critical tables scattered across: central data warehouse, S3 data lakes, business relational databases, Salesforce, Excel files on SharePoint, manual SI exports… A new Data Scientist spends on average 2 to 4 weeks finding and obtaining access to the right data.

WARNINGHidden cost — A team of 4 DS at €75k/year losing 3 weeks finding data equals ~€17k per project. Across 10 projects, the Data Catalog pays for itself.

What is a Data Catalog

A Data Catalog is a centralized, searchable directory of the company’s data. It answers the questions: where is this data, who owns it, is it fresh, is it high-quality, who has access, how do I request it?

What a catalog contains

Market tools

Typical dataset record in the catalog

NOTEDataset: fct_transactions_clients
Description: All customer transactions since 2018, aggregated daily.
Owner: Data Finance Team (data-finance@compagnie.com)
Business Steward: Marie Dupont (CRM Department)
Sensitivity: Sensitive (personal data)
Freshness: Daily update at 2am
Quality (Sept 2025): 99.2% completeness, 0.3% duplicates
Access: On request via ServiceNow form, DPO + owner validation within 5 days
Upstream lineage: SAP, Salesforce
Downstream lineage: fraud_v3, churn_v2 models, finance dashboard

Access management: 4 sensitivity levels

Level Example Access Delay
Public Product catalog, published figures Everyone Immediate
Internal Non-sensitive internal metrics Any employee on request < 24h
Sensitive Personal customer data Manager + DPO 3-5 days
Secret Strategy, M&A, health data CODIR + retrospective audit Case by case
TIPPM Tip — From the scoping document onward, identify the sensitivity of each targeted dataset and anticipate delays. If you need a “secret” dataset and discover it only at sprint 5, you lose 2 months.

Anonymization and pseudonymization

Pseudonymization

Replace identifiers with pseudonyms. Reversible if the mapping table is kept.

Ex: customer_id 12345 → HASH_a3f8b2c

Anonymization

Make re-identification impossible. Irreversible.

Ex: aggregation by age decile + rounding amounts to €100

Outside GDPR if done properly (k-anonymity, etc).

WARNINGPitfall — Many teams think they anonymize by hashing an email. That is only pseudonymization. True anonymization is harder (k-anonymity, differential privacy).

The data sandbox trick

Instead of granting access to production data, many companies create a sandbox:

TIPResult — DS prototype in the sandbox, then request prod access only for the final phase. Net acceleration: from 4 weeks down to 4 days to start a POC.

Real case: Carrefour and its Unity Catalog

CRISP-DM, TDSP Frameworks and Modern Alternatives

NOTEObjective — Discover the 3 most widely used frameworks for structuring an AI project, understand their strengths and weaknesses, and know which one to adopt depending on your context (large enterprise, startup, agency).

Why a methodological framework

A framework is not a fixed Gantt chart. It is a shared lens for all stakeholders: you, the data team, the sponsor, the auditor. Without a framework, everyone calls “phase 2” what the neighbor calls “phase 4”. With a framework, everyone speaks the same language.

TIPTip — Choosing a framework and sticking to it, even if imperfect, is a thousand times better than improvising. Discipline in the path matters more than the exact choice.

Framework 1 — CRISP-DM (the classic)

CRISP-DM (Cross Industry Standard Process for Data Mining), published in 1999 by a European consortium (IBM, Daimler, NCR, OHRA), remains the most used framework worldwide in 2026 (KDnuggets survey).

The 6 phases

NOTEStrength of CRISP-DM — Very explicit about the feedback loop: if evaluation fails, return to Business Understanding. It is intellectually honest.
WARNINGWeakness — Not designed at all for modern MLOps (continuous monitoring, automatic retraining). The word “production” barely appears. It must be supplemented with an MLOps layer.

Framework 2 — Microsoft TDSP

TDSP (Team Data Science Process), launched by Microsoft in 2017, is a modernized CRISP-DM with a team vision and a focus on deliverables.

The 5 phases

What changes vs CRISP-DM

TIPFor whom — Excellent for large enterprises that already have a PMO and want to standardize deliverables across teams.

Framework 3 — Google ML Lifecycle (the modern one)

Google published in 2019 an end-to-end MLOps-oriented cycle that is now the reference in modern tech (and widely used in AI startups).

The phases

NOTEStrength — Monitor is a full-fledged phase, not just a “maintenance bucket”. It forces the PM to budget for run, not only build.

Comparison table

Criterion CRISP-DM TDSP Google ML Lifecycle
Year 1999 2017 2019
Focus Data mining Team + deliverables MLOps
MLOps integrated No Partial Yes
Phases 6 5 6
Templates provided Few Many Medium
Roles specified No Yes Partial
Adapted to modern cloud No Azure ML Vertex AI / agnostic
Ideal audience Learning, agencies Large enterprises Tech startups, FAANG

Which framework to choose? Our recommendation

Tech startup (< 30 people)

Google ML Lifecycle: agility and MLOps focus. You can deploy quickly and iterate without bureaucracy.

Traditional large enterprise

TDSP: structure and deliverables, perfect for reassuring internal PMO and auditors.

Consulting / training agency

Course Presentation

NOTEObjective — Understand what makes steering an AI project so different from a classic software project, why 80% of these projects fail, and what you will concretely learn to do in this course.

Learning objectives

TIPAt the end of this module — You will be able to explain in 2 minutes what AI Product Management is, identify the 5 major differences with a software project, and have a clear vision of the skills you will develop across the 10 chapters.

The concrete problem you will solve

Imagine the following situation. You are a Product Manager in a bank or a startup. One morning, your CEO walks into your office and says:

WARNING“We need AI. The board approved an €800,000 budget. We want something that works in 6 months. You’re driving it.”

Here are the questions that will land on you during the week:

Business questions you will be asked

What you will learn to do

NOTEThe job of an AI PM in one sentence — It is the bridge between scientific uncertainty (models make mistakes) and business certainty (the sponsor wants a measurable result). All your added value lies in this translation.

Why an AI project is not a classic software project

Many beginner project managers think an AI project is just “a software project with a model inside”. That is wrong, and exactly why they fail. Here are the 5 major differences:

Aspect Classic software project AI project
Result Deterministic (same input ⇒ same output) Probabilistic (same input ⇒ result with error margin)
Specifications Detailed requirements document Hypotheses to validate experimentally
Lifecycle Dev ⇒ Test ⇒ Deploy ⇒ Maintenance Data ⇒ Model ⇒ Eval ⇒ Deploy ⇒ Monitoring ⇒ Retraining
Data dependency Marginal (config files) Vital (no data, no product)
Production drift Code does not degrade by itself The model naturally degrades (data drift)
WARNINGDirect consequence — You cannot apply standard Scrum, a fixed Gantt, or a detailed functional specification. You need an adapted model, which this course will give you.

The harsh reality: 80% of AI projects fail

According to Gartner (2024) and a McKinsey 2025 report, between 70% and 85% of enterprise AI projects never move beyond the proof of concept. Why? Here are the 3 main causes:

1. Poor scoping

The project is launched “because we need to do AI”, without a clear business problem behind it. Result: a model that works technically but delivers no value.

~35% of failures

2. Insufficient data

After 3 months of POC, the team discovers the data is incomplete, biased, or legally inaccessible. The project dies before production.

~30% of failures

3. No production deployment

The model stays in a data scientist’s Jupyter notebook. No one thought about deployment, monitoring, or SI integration.

~20% of failures

TIPThe good news — These 3 causes are entirely within the Product Manager’s remit. Properly scoping, validating data early, planning MLOps from the start: these are learnable skills, and exactly what this course covers.
go-further

This article covers the most useful excerpts — the full AI Project Management course (11 chapters, 35 lessons, corrected exercises and capstone project) takes you all the way.

./access-the-full-course free course: Claude Cowork

FAQ

How long does it take to learn AI Project Management?
With a structured progression (11 chapters, 35 short practical lessons), you reach an operational level in a few weeks at 30–60 minutes per day. The key is to practice each concept immediately.
Are there any prerequisites?
Basic computer science knowledge is enough. If you can use a terminal and read simple code, you are ready.
Where to start concretely?
Reproduce the commands in this article, then follow the full AI Project Management course: it sequences the 35 lessons in order, with exercises and a capstone project.
./read-also
→ Get started with Excel Copilot: your first concrete step today→ Essential AI Toolbox: the 9 key steps from zero to operational→ AI Career Opportunities explained simply (with diagrams and real code)

📬 Want to receive this type of guide every week? Subscribe for free — real code, zero fluff.

Stay up to date

Enter your email
Subscribe