Ten Questions on AI Implementation
From identifying the right problem to measuring impact, scaling beyond MVP, and the ethical guardrails: a Q&A on how AI implementation actually plays out at scale.
1. The most critical first steps
The very first step towards implementing AI is leveraging human intelligence in the company. That means having data-savvy people in the team who are able to identify the low-hanging fruits: which problems, bottlenecks, or inefficiencies yield the most impact, considering that they can be addressed by available data.
There are many flavours of AI solutions, from GenAI that creates code and marketing campaigns to machine learning algorithms that recommend the next product to be offered. Choosing which to start implementing requires human intelligence over the company’s goals.
Second is embracing experimentation: testing those solutions in small markets or in limited time periods, learning from the results. Experimentation is the process of accruing the impact each solution has brought, understanding what works and what does not.
Third is consistency in scaling up and iterating. Whatever solution works must be expanded to cover the whole business while it continues to be measured for efficiency and improved.
2. Identifying the right problem
Before starting, companies must run an assessment of the existing challenges, capabilities, investment levels, and benefits being pursued. The objective is to balance technical feasibility with business impact. Candidate problems must align with company core business strategies.
My favourite framework is an AI opportunity matrix that maps potential use cases against two axes: strategic value (revenue impact, customer experience) and implementation complexity (data readiness, technical complexity, financial ability, current stage of data development).
3. Scaling AI beyond the prototype stage
Three big barriers. First is trying to utilise the same AI algorithm or ML methodology developed for the MVP in the scaled-up version. It is the equivalent of knowing how to prepare pasta at home and then working the same way to feed an army battalion. Even though the main course is the same, things run very differently in the kitchen. Different algorithms, codes, and tech stack are needed to scale things up.
The second barrier is cost. Costs differ greatly from MVPs to scale. Investment costs have to be factored in, as well as operational costs, training, security, compliance, redundancy. P&L analysis is still king.
Third is the integration of the AI solution into the overall customer experience. One must never forget that AI is part of the customer experience, it is not all of it. Any AI-powered solution must fit perfectly into the customer journey. If AI introduces an element of oddity that undermines other aspects of the company’s offer, it might bring problems far greater than the ones it initially was designed to solve.
4. Agile versus enterprise approach
It is not one or the other, but one AND the other. The important thing is understanding where companies are in the development cycle. Fast agile experimentation is essential to identify the existing problems, costs, ROI, and interactions with established business models. Enterprise-level approach usually tackles the topics related to scalability, reliability, security: things that allow the MVPs to profoundly transform the business and unlock major returns.
5. Measuring AI’s impact
This is fundamentally linked to the company expectations prior to the AI deployment. When it comes to cost, consider total cost of ownership (data acquisition, model development, cloud and compute expenses, maintenance) and the reduction of operational costs (productivity increases, error correction, process inefficiencies).
Scalability usually involves KPIs that measure performance like inference latency, cloud costs per million requests, and so on.
Effectiveness metrics are measured by an internal metric that has to make sense for the leadership, reflecting a balance between technical results (how well a given model identifies fraud) and business results (how much money was saved by flagging fraudulent operations).
6. The role of data
In my experience, one must always start working with scalable architectures that can handle the amounts of structured and unstructured data the company needs at that particular moment. Cloud-based data lakes, modern data warehouses, and real-time processing pipelines are essential. Without these core capabilities, AI projects often become bottlenecked by slow retrieval times, fragmented datasets, and rigid legacy systems.
Moving beyond infrastructure, a strong governance to maintain accuracy, consistency, and security is the often-overlooked stepping stone that guarantees data can be trusted and put to good use.
Ultimately, AI success is as much about culture as it is about technology. A strong data infrastructure must be accompanied by a shift in mindset, prioritising cross-functional collaboration, data literacy, and ethical considerations.
7. AI and decision-making
Algorithms focused on automation, efficiency, and quick response belong to the type of AI and machine learning mankind has already put to use for the last 60 years. What is new about the current stage of AI is that it already is far more than an automation tool. It has become a powerful engine for strategic decision-making, capable of uncovering patterns, predicting trends, and guiding leaders through complexity with data-driven insights.
The challenge is to organise the flow of data so AI can simulate different scenarios. AI agents and specialised analytics help businesses anticipate market shifts, mitigate risks, and optimise resource allocation. It also enhances human judgment by surfacing hidden correlations and challenging biases.
8. Ethical challenges
Bias in AI arises from two root causes. First, when models learn from skewed or incomplete data, leading to unfair outcomes. Second, when humans interacting with AI carry their own biases into the models by the prompts, flows, or selections they provide.
To combat this, companies must implement rigorous data governance practices, regularly auditing datasets for representativeness and fairness. Techniques such as bias detection algorithms, adversarial testing, and diverse data sampling can help mitigate unintended discrimination.
But technology alone is not enough. Fairness and ethical behaviour must be part of the company culture and training. Establishing interdisciplinary review teams, bringing together ethicists, legal experts, and domain specialists to scrutinise AI systems before they are deployed at scale, must be part of the toolbox.
Transparency and accountability are equally critical. AI models should not operate as “black boxes”. AI development must prioritise explainability by using interpretable AI techniques. This approach, favoured in Europe and the UK, is the most sensible one.
9. Common pitfalls
One of the most common mistakes is treating AI as a plug-and-play solution rather than a strategic investment that requires careful planning and integration. Many organisations rush to deploy AI without first ensuring they have high-quality, well-governed data.
Another critical misstep is failing to define clear success metrics and expecting immediate results without iterative refinement. AI is not a one-time deployment. It requires continuous monitoring, retraining, and adaptation as business conditions evolve.
To avoid these pitfalls: invest in AI literacy across all levels; build a strong data foundation; foster collaboration between AI engineers, data scientists, business strategists, and frontline employees; establish feedback loops with human oversight and performance audits.
10. The future of AI implementation
AI implementation is poised to move from boardroom buzzword to an invisible, indispensable force running beneath the surface of everyday business operations. Companies will no longer ask “should we use AI?” but rather “how do we solve this?”
I expect AI to shift from clunky, one-size-fits-all models to hyper-personalised, context-aware systems that adapt to individual users, industries, and even moods. Businesses should brace for the rise of AI copilots, not just automating tasks but actively enhancing human decision-making.
Ethical AI, explainability, and regulatory compliance will become non-negotiable, not just to avoid lawsuits but to build trust in an increasingly sceptical world. The future of AI is not about replacing people; it is about augmenting them.
As the AI-powered economy takes shape, I strongly believe that the currency of the future is trust. If your company can accurately interpret the vast amount of data being collected from the world, it will be trusted and therefore valued by customers, partners, and society.