How to Successfully Start an AI Transformation Project

The Anatomy of a Stalled AI Programme

Many enterprise AI transformation programmes follow a predictable, and often disappointing, trajectory. They launch with executive fanfare, a significant budget, and a mandate to innovate. The first three months are a whirlwind of vendor meetings, platform evaluations, and brainstorming sessions. A promising pilot project is chosen—often something technically interesting but perhaps disconnected from a core business value driver.

Then, around month four, momentum grinds to a halt. The pilot project, which seemed straightforward as a proof of concept (PoC) on a clean dataset, hits the reality of production data. It is messy, incomplete, and siloed. The business unit that was meant to "own" the solution is unclear on how it will integrate into their daily workflows. The central AI team, operating in isolation, struggles to get the business context they need. Questions about data privacy, model explainability, and compliance with regulations like the EU AI Act suddenly move from theoretical to blocking.

This is "pilot purgatory," and it is the primary reason AI transformations fail. They stall because the initial focus is on technology, not on the integrated system of value, people, process, and governance required to move from a Jupyter notebook to a production service that generates measurable return.

A successful programme anticipates these month-four hurdles from day one. It begins not with a platform, but with a portfolio. It designs an operating model for scale, not just for a single project. It treats governance and change management as prerequisites for success, not as afterthoughts. In our engagements, we have found that the organisations that succeed are those that approach AI as a fundamental change to the business, requiring the same rigour as a major M&A integration or a new market entry.

Portfolio Selection: Balancing Value, Feasibility, and Data

The single most important factor in a successful AI transformation is selecting the right initial projects. The goal is not to find the most technologically advanced application, but to deliver demonstrable business value quickly, creating momentum and earning the right to tackle more complex challenges. We advise clients to evaluate potential use cases across three dimensions: business value, technical feasibility, and data readiness.

A common mistake is to pursue dozens of ideas at once, spreading resources too thin. A better approach is to identify a long list of 10-20 potential use cases and then rigorously score them to down-select to a portfolio of two or three initial projects for the first six months.

The Scoring Framework

• Business Value (40% weight): How significant is the impact? This can be measured through increased revenue (e.g., cross-sell models), cost savings (e.g., process automation), or risk reduction (e.g., fraud detection). The value should be quantifiable in EUR and directly attributable to a specific P&L owner.
• Technical Feasibility (30% weight): Can we build this with a high degree of confidence? This considers the maturity of the required AI techniques, the complexity of integration with existing systems, and the availability of skills. A demand forecasting model using established time-series algorithms is more feasible than attempting to build a novel multi-modal foundation model from scratch.
• Data Readiness (30% weight): Is the necessary data available, accessible, and of sufficient quality? This is often the biggest obstacle. A project is "data ready" if the core datasets exist, are reasonably clean, and can be accessed by an engineering team without a six-month procurement and legal battle. Strong data engineering capabilities are a prerequisite here.

Below is a sample scoring matrix we might use with a client in the insurance sector:

Based on this analysis, Automated Claims Triage and Predictive Agent Churn are the clear winners for an initial portfolio. They offer a strong combination of high value, proven technical approaches, and ready data. Personalised Pricing is a valuable future goal, but the data readiness issues mean it should be deferred until a foundational data platform is more mature. The generative AI chatbot, while exciting, carries too much technical and reputational risk for a first project.

Worked Example: The Business Case for Claims Triage

Let's quantify the value for the "Automated Claims Triage" project for a mid-sized European insurer.

• Annual Claims Volume: 500,000 claims
• Average Manual Triage Time: 15 minutes (0.25 hours) per claim
• Fully-Loaded Cost of Claims Handler: €35 per hour
• Current Annual Triage Cost: 500,000 * 0.25 * €35 = €4,375,000

The AI project aims to automate 80% of simple claims triage, routing them directly to the correct department, and flagging the 20% of complex claims for expert human review.

• Projected Automation Rate: 80%
• Projected Annual Savings: €4,375,000 * 80% = €3,500,000
• Estimated Project Cost (Year 1): This includes a team of 4-5 senior engineers for two quarters (€400k), plus software/infra costs (€50k). Total ≈ €450,000.
• Year 1 Net Value: €3,500,000 - €450,000 = €3,050,000

This is the kind of clear, compelling business case that secures executive sponsorship and justifies the investment in high-quality AI engineering.

Operating Model: Central Enablement, Federated Delivery

How you organise your teams to deliver AI projects is as important as which projects you choose. We see organisations struggle with two extremes: a fully centralised "AI Center of Excellence" that becomes a bottleneck, and a fully decentralised "free for all" that leads to duplicated effort, inconsistent quality, and spiralling costs.

The most effective model we have implemented with clients is a hybrid: a central enablement team (the "hub") that supports federated delivery teams embedded within business units (the "spokes").

                  +--------------------------------+
                  |       AI Steering Committee    |
                  | (Exec Sponsor, Business Leads) |
                  +--------------------------------+
                                  |
                                  | Governs & Prioritises
                                  |
                  +--------------------------------+
                  |    Central AI Enablement Hub   |
                  |--------------------------------|
                  | - Platform & MLOps             |
                  | - Governance & Best Practices  |
                  | - Reusable Components          |
                  | - Advanced R&D, Talent Dev.    |
                  +--------------------------------+
                     |            |            |
           Provides Tools &     |       Supports & Guides
             Frameworks         |
           +--------------------+--------------------+
           |                    |                    |
+---------------------+ +---------------------+ +---------------------+
|   Spoke: Marketing  | |   Spoke: Operations | |    Spoke: Finance   |
|---------------------| |---------------------| |---------------------|
| - Product Manager   | | - Product Manager   | | - Product Manager   |
| - AI/ML Engineer    | | - AI/ML Engineer    | | - AI/ML Engineer    |
| - Data Scientist    | | - Data Scientist    | | - Data Analyst      |
| - Business Analyst  | | - Process Expert    | | - Domain Expert     |
| (Delivers Use Case) | | (Delivers Use Case) | | (Delivers Use Case) |
+---------------------+ +---------------------+ +---------------------+

The Central Hub's Role

The central team does not build all the AI applications. Instead, it acts as a force multiplier. Its responsibilities include:

• Platform: Building and maintaining the core MLOps infrastructure, feature stores, and model monitoring tools that all teams use.
• Governance: Setting standards for model development, validation, ethics, and compliance (e.g., creating templates for EU AI Act documentation).
• Best Practices: Curating and disseminating best practices for software engineering, model development, and project management.
• Talent: Acting as a hub for hiring, training, and career development for AI talent across the organisation.

The Federated Spokes' Role

The "spokes" are cross-functional teams embedded in the business units. They are closest to the problems and the data. Their responsibilities include:

• Delivery: Owning the end-to-end lifecycle of a specific AI solution, from ideation to production deployment and maintenance.
• Domain Expertise: Bringing deep knowledge of the business process, the data's meaning, and the end-user's needs.
• Value Realisation: Ensuring the solution is adopted and tracking the business metrics that prove its value.

This model balances central efficiency with local context and ownership. The hub prevents each team from reinventing the wheel, while the spokes ensure that what is built actually solves a real business problem.

Governance, Risk, and the EU AI Act Reality Check

In Europe, AI governance is not optional. The EU AI Act introduces a risk-based framework that will have significant implications for how AI systems are designed, deployed, and monitored. Ignoring this is not a viable strategy; it introduces legal, financial, and reputational risk.

The Act categorises AI systems into four risk levels:

1. Unacceptable Risk: Systems that are banned outright (e.g., social scoring by governments).
2. High Risk: Systems used in critical areas like medical devices, critical infrastructure management, employment, or law enforcement. These face stringent requirements for data quality, documentation, human oversight, and transparency before they can be placed on the market. Many common enterprise use cases (e.g., CV-scanning, credit scoring) will fall into this category.
3. Limited Risk: Systems that require transparency obligations. For example, users must be aware they are interacting with a chatbot.
4. Minimal Risk: The vast majority of AI systems (e.g., spam filters, recommender systems in video games).

For an enterprise starting its AI transformation, this means:

• Risk Assessment is Part of Project Selection: The risk categorisation under the AI Act must be a factor in your portfolio scoring. A "High Risk" project is not impossible, but it carries a significant overhead in terms of compliance, documentation, and a longer time to market. Your first projects should ideally be in the "Minimal" or "Limited" risk categories to build momentum.
• Governance Is an Engineering Discipline: Compliance is not just a checklist for the legal team. Requirements like data governance, model traceability, and robustness must be built into the development lifecycle from the start. This is a core function of the central enablement hub—providing the MLOps tooling and templates that make compliance the path of least resistance for development teams.
• Human-in-the-Loop is a Key Pattern: For many high-consequence applications, a fully autonomous AI is both risky and often non-compliant. Designing systems where the AI provides recommendations or flags exceptions for a human expert to review is a powerful pattern. It mitigates risk, improves user acceptance, and often leads to better outcomes.

What we would NOT do is create a large, bureaucratic "AI ethics committee" that reviews projects late in the cycle. Governance must be lightweight, automated where possible, and embedded into the daily workflow of the delivery teams.

Talent: The Hire, Partner, or Upskill Decision

An AI transformation is fundamentally a talent transformation. You cannot succeed without the right skills. Organisations face a classic "build, buy, or borrow" decision for talent: hire in-house, partner with a specialist firm, or upskill your existing workforce. The right strategy is usually a blend of all three, evolving over time.

When to Hire

Hiring is the long-term goal for core capabilities. You want to own the strategic intellectual property and have deep domain expertise fused with AI skills. However, hiring a full senior team from scratch is slow and expensive. In the current market, it can take over six months to find, hire, and onboard a single experienced ML Engineer. Use hiring for key leadership roles first (e.g., Head of AI, Principal ML Engineer) and to augment existing teams.

When to Partner

Partnering with a specialist consultancy is the fastest way to achieve initial results and de-risk your programme. An experienced partner brings not just engineers, but a proven methodology, reusable solution architectures, and experience from dozens of similar engagements. This is ideal for your first 1-2 waves of projects. The goal of a good partnership is not just to deliver a project, but to transfer knowledge and help you build your own capabilities. When assessing partners, it's crucial to understand how to evaluate providers properly based on their engineering depth, not just their sales pitches. The right partner works as an extension of your team, as outlined in how we work.

Worked Example: Cost of a Pilot Project Team (3 Months)

Let's compare the cost of getting a 3-month pilot project delivered by hiring versus partnering.

Option A: Hiring a team (4 people)

• Roles: 1x Product Manager, 1x Principal ML Engineer, 2x Senior ML/Data Engineer
• Recruitment Costs: Assume 20% of first-year salary. For senior roles averaging €120k, this is ~€24k per hire. Total: €96,000.
• Time to Hire: Optimistically, 4 months.
• Salary Costs (for 3-month project): (4 * €120,000 / 12) * 3 = €120,000
• Fully-Loaded Cost (incl. benefits, NI, office etc. @ 1.4x salary): €120,000 * 1.4 = €168,000
• Total Effective Cost for First Project: €96,000 (recruitment) + €168,000 (loaded salary) = €264,000. And this assumes you can even find and land these four people in four months, which is unlikely. The project starts in month 5.

Option B: Partnering with a senior-only firm like Golux

• Team: Similar senior-only pod of 4.
• Time to Start: 2-4 weeks.
• Cost: Based on a blended day rate of ~€950/person. 4 people * 60 days * €950/day = €228,000.
• Total Cost for First Project: €228,000. The project starts this month.

The partner route is not only cheaper for the initial engagement but delivers value at least four months faster. A good strategy is to use a partner to deliver the first win, while your internal team hires and learns alongside them. This blend of external expertise and internal ownership is a powerful accelerator. This is a key consideration within the broader build vs buy AI software decision framework.

When to Upskill

Upskilling your existing analysts, software engineers, and business experts is crucial for long-term scale. Your people have invaluable domain knowledge. Training them in data literacy, basic data science techniques, and how to use AI tools makes them more effective and creates a culture of data-driven decision making. Upskilling is not about turning every analyst into a PhD-level research scientist; it's about creating "AI translators" who can bridge the gap between business and technology.

Change Management and Workforce Communication

An AI model that is technically perfect but never used by the business delivers zero value. AI transformation is, above all, a human change programme. Resistance to change is natural, often rooted in fear of job displacement, disruption to established workflows, or a lack of understanding. Proactive and thoughtful change management is non-negotiable.

What Not To Do:

• Announce the new "AI tool" in a single all-hands meeting a week before launch.
• Focus communication solely on the technology and its features.
• Ignore the "What's in it for me?" question from the perspective of the end-user.
• Assume training is a one-off event.

A Better Approach:

1. Involve Users from Day One: The end-users of the AI system should be involved in the design process. The federated "spoke" teams must include representatives from the business process being changed. Their insights are critical for designing a usable tool that fits the workflow.
2. Identify and Empower Champions: Within each affected business unit, identify influential individuals who are enthusiastic or at least open-minded about the change. Make them part of an advisory group, give them early access, and let them become advocates for the new system among their peers.
3. Communicate Honestly and Often: Be transparent about the goals of the project. If the goal is efficiency, frame it as freeing up employees from repetitive, low-value tasks to focus on more strategic, high-value work (e.g., handling complex customer cases instead of data entry). Acknowledge that roles will evolve and provide a clear vision for what the future roles look like, along with the training to get there.
4. Focus on Augmentation, Not Replacement: For most enterprise AI, the most powerful narrative is one of augmentation. The AI is a "co-pilot" that helps the human expert make faster, better decisions. The claims handler still makes the final call on a complex claim, but the AI surfaces the relevant information and flags potential fraud, saving them hours of manual work.
5. Measure Adoption as a Key Metric: Track not just whether the system is live, but who is using it, how often, and whether they are achieving the intended benefits. Link adoption metrics to business outcomes.

Metrics the Board Should See Quarterly

The executive board and steering committee need a concise, business-focused view of the programme's health and impact. Reporting should go beyond technical jargon and focus on value. A good quarterly report can be structured around three areas: Programme Velocity, Business Impact, and Capability Growth.

Sample Quarterly AI Programme Dashboard

This kind of dashboard tells a clear story. In this example, the programme is successfully delivering cost savings and building capability. However, it's lagging on revenue-generating projects and the cycle time needs improvement, which might trigger a discussion about streamlining the governance process or providing more support to the revenue-focused "spokes."

A 12-Month Plan with Quarterly Checkpoints

A successful AI transformation is a marathon, not a sprint. A realistic 12-month roadmap provides structure and focus, with clear goals for each quarter.

Quarter 1 (Months 1-3): Foundation & First Win

• Focus: Establish the core team and deliver a single, high-impact project.
• Actions:
  • Form a steering committee and a lean central enablement hub.
  • Conduct a portfolio selection workshop to identify and approve the first 2-3 projects.
  • Launch Project #1 with a dedicated team (internal or partner-led).
  • Define a "Version 1" of the MLOps platform and governance framework.
  • Engage an expert partner for initial AI consulting to accelerate this phase.
• Checkpoint Goal: Project #1 live in production, delivering measurable value. Business case for Q2 projects approved.

Quarter 2 (Months 4-6): Standardise & Scale

• Focus: Learn from the first win, standardise processes, and begin parallel execution.
• Actions:
  • Conduct a post-mortem on Project #1. What worked? What didn't?
  • Refine the MLOps platform and governance based on learnings. Build the first reusable components (e.g., a standard model deployment pipeline).
  • Launch Project #2 and #3.
  • Begin a formal upskilling programme for business analysts and software engineers.
• Checkpoint Goal: Three projects in flight. First version of the central "AI Playbook" is published.

Quarter 3 (Months 7-9): Embed & Expand

• Focus: Drive adoption of existing solutions and expand the portfolio into new business domains.
• Actions:
  • Focus on change management and driving adoption for the first wave of projects.
  • Establish the federated "spoke" teams as a formal part of the organisational structure.
  • Launch the second wave of projects (Projects #4, #5, #6), potentially tackling slightly more complex problems.
  • Mature data governance practices and the shared feature store.
• Checkpoint Goal: At least one business unit is operating self-sufficiently within the hub-and-spoke model. Tangible ROI from the first projects is confirmed.

Quarter 4 (Months 10-12): Industrialise & Plan

• Focus: Optimise for efficiency and plan the roadmap for Year 2 based on ROI.
• Actions:
  • Consolidate MLOps tooling and automate compliance/monitoring processes.
  • Measure the cumulative ROI of the year's portfolio.
  • Build the strategic roadmap and budget for Year 2, focusing on scaling successful patterns and tackling higher-value, more complex initiatives.
• Checkpoint Goal: Year 1 ROI report delivered to the board. Year 2 plan and budget approved.

Frequently asked questions

How should a company start an AI transformation?

A company should begin by resisting the urge to boil the ocean with a large, speculative platform build. Instead, the focus should be on demonstrating tangible value quickly. This means identifying two or three high-value use cases where the necessary data is accessible and a clear business owner is committed to the outcome. These initial projects should be delivered end-to-end within a single quarter. In parallel, a lightweight central team should be formed to establish the initial governance guardrails and technical standards, creating a foundation that can be built upon as the programme scales and proves its worth.

Key takeaways

• AI transformations stall due to a premature focus on technology over business value, poor data readiness, and a lack of change management.
• Start with a portfolio of 2-3 projects rigorously selected for high value, technical feasibility, and data availability. Quantify the expected ROI in EUR.
• Adopt a "hub-and-spoke" operating model: a central team enables platform and governance, while federated teams in business units deliver the use cases.
• Integrate governance, risk, and compliance with the EU AI Act into your engineering lifecycle from day one. It is a design constraint, not a legal afterthought.
• Use a blended talent strategy. Partner with experts like Golux to get fast initial wins and transfer knowledge, while simultaneously hiring for key long-term roles and upskilling your existing workforce.
• Success should be measured with business-focused metrics (cost savings, revenue uplift, adoption rates) that are reported to the board quarterly, not just technical metrics like model accuracy.

Successfully embedding AI into your organisation is a multi-year journey of strategic change. By starting with focused, high-value projects and deliberately building the operating model, governance, and talent around them, you can create the momentum needed for a true and lasting transformation. To discuss how to tailor this blueprint to your organisation's specific needs, we invite you to book a strategy session with our senior team.