No Data, No AI by Sandro Saitta

You could have written a book purely about AI. Why did you choose to lead with data?

I initially considered AI as the sole focus, but every discussion kept circling back to data quality, data visualisation, and the fundamental importance of data. Those topics became core chapters in their own right. The result is a book about both data and AI because fundamentally, no data means no AI. I also wanted to challenge readers who might be entirely focused on AI to invest more attention in data. That was the driving objective.

Leaders are being told they need to become AI-first. What’s the risk of skipping straight to AI before data fundamentals are in place?

The biggest risk is a false sense of completion – i.e. believing the transformation is done once a chatbot is deployed and some training has taken place. That approach only captures the low-hanging fruit. The real competitive advantage comes from leveraging analytics and AI on your own proprietary data. If you focus exclusively on AI as a generic tool, you get what everyone else gets. Data – and specifically its quality and governance – is what differentiates one organisation from another.

How important is data quality, and how should organisations approach it?

“Garbage in, garbage out” is a truism, but it shouldn’t be a showstopper. Every company I work with believes it has particularly poor data quality, so this is universal. The reality is that data quality is rarely good anywhere; the task is to improve it continuously.

The key mistake I see is treating data quality as a tool problem and simply deploying a data catalogue, expecting it to solve everything. A tool is only as useful as the processes and people supporting it.

Organisations need to define data ownership, document datasets, establish stewardship roles, and maintain this on an ongoing basis. A one-time effort is never enough.

In terms of approach, I recommend combining top-down and bottom-up methods. Top-down means identifying the KPIs and decisions that depend on data, and defining the quality threshold required for those purposes. Bottom-up means examining datasets directly, measuring missing values, duplicates, and outliers and then cleaning with a clear business purpose in mind. The goal is not perfect data; it is data that is fit for its intended use.

For companies early in their data and AI journey, particularly in traditional sectors, where should they start?

Customer-focused use cases are the natural starting point. Targeting leads, upselling, and cross-selling have been used effectively for decades. Most companies have reasonable customer data, and the return on investment is relatively straightforward to measure in revenue terms.

For companies in a optimisation phase, rather than a growth phase, supply chain and data-driven forecasting make more sense. The focus shifts from generating more revenue to doing more with less, and use cases in that space are well-aligned with that objective.

Can you give an example from your own experience in forecasting?

A good example is a project I worked on at a large consumer coffee brand. The goal was to use historical sales data, alongside other relevant variables, to predict coffee consumption and ensure the right quantity of capsules would ship to the right markets globally.

Given the costs of warehousing and distribution, precision in timing and volume was critical.

The solution was designed to be centralised but collaborative. Forecasts were generated centrally from HQ and shared with individual markets, which used them as inputs to their own demand planning. Crucially, the human demand planner retained final decision-making authority and could adjust the forecast when local knowledge warranted it. The model informed the decision but it did not replace it.

Many companies still lack reliable C-level reporting. Should they be thinking about AI at all before that is in place?

Ideally, the progression should be sequential: data foundations first, then reporting and self-service BI, then machine learning, and finally generative AI. In practice, some parallel activity is unavoidable given how quickly the field is moving.

That said, C-level reporting matters both directly and indirectly. Directly, it gives leaders the data they need to run the business. Indirectly, and this is often underestimated, it builds a data-driven mindset at the top of the organisation. Without executive buy-in and sponsorship, data and AI initiatives will remain isolated projects with limited organisational impact.

Getting leaders to experience the value of data first-hand is one of the most effective ways to drive broader transformation.

How long does it typically take to establish robust C-level reporting in a traditional sector company?

I have been having this conversation for approximately 15 years and it remains unfinished in many organisations. Technology is advancing rapidly, but organisational adoption is not keeping pace, particularly at the executive level.

Trust, culture, and the human dimension of change are as important as the technical infrastructure.

The most effective approach is to build incrementally by sharing insights, demonstrating small wins, and earning trust over time. Attempts to deliver a comprehensive, transformative change all at once rarely succeed. It is a continuous journey.

For companies with a more established data team and some standard analytics capability in place, what is the logical next step?

The priority should be scalability. Rather than simply hiring more data specialists, the greater opportunity is enabling the wider workforce to work more independently with data using self-service BI tools, automating routine tasks, and developing a data-driven approach to their own work.

In many organisations, highly capable data teams spend a disproportionate amount of time distributing reports because broader employees won’t engage with the tools themselves. Breaking down that barrier through targeted training, change management, and communicating the personal benefit to individuals would free data teams to focus on genuinely advanced work: machine learning, predictive analytics, and higher-value decision support.

I would go further and argue that data scientists should invest more time coaching colleagues to leverage data effectively, rather than focusing exclusively on building increasingly complex models.

How should non-technical business leaders approach selecting data and AI use cases?

A combined top-down and bottom-up approach is most effective. Top-down ensures that any use case is anchored to a real business objective or KPI. Bottom-up ensures the data actually available supports it. Working only top-down produces compelling ideas with nothing to execute them. Working only bottom-up produces technically sound solutions that fail to address genuine business needs.

Equally important is rigour in defining use cases. A simple one-page canvas agreed by all relevant stakeholders and covering context, data sources, expected outputs, and how the solution will be integrated prevents the ambiguity that so often derails projects.

Let’s do sales forecasting can mean entirely different things to different people. Getting specific, then prioritising across use cases using a straightforward feasibility-impact matrix, significantly improves outcomes.

You cite a high failure rate for data and AI projects. What are the most common causes?

Misalignment between the technical team and the business is the most frequent root cause. The one-page canvas helps, as does having an analytics translator, i.e. someone who can bridge data and business teams and maintain alignment throughout a project.

In most failure cases I have seen, the problem is not a technical one. The solution may work; it simply fails to answer the right question, or it is not delivered at the frequency, granularity, or speed that makes it useful in practice.

A second, increasingly common failure pattern, particularly with generative AI, is the proof-of-concept that never reaches production. It is straightforward to build a convincing POC with generative AI. Moving it into a reliable production environment is significantly more challenging, and the gap between those two states is considerably wider than for traditional machine learning projects.

What are the main misconceptions you see around generative AI in organisations?

The most pervasive is over-trust. These tools are remarkably convincing and they present outputs confidently, which leads people to accept them without question. Related to this is the misconception that generative AI possesses genuine intelligence or logical reasoning. It does not. These models generate statistically probable sequences of words; there is no underlying logic or reasoning process.

This becomes apparent when you push them into genuinely complex analytical territory.

Leaders need to set realistic expectations: generative AI is excellent at many tasks, but it lacks common sense and true reasoning capability. Treating it otherwise leads to poor decisions and misplaced confidence.

Where is generative AI genuinely adding value for businesses today?

The most consistent value I see comes from using it as a sounding board, not as a starting point. My recommendation is always to begin your own work first: draft the document, develop the strategy, build the presentation. Then use generative AI to challenge your thinking, identify gaps, and pressure-test your approach.

Starting with the AI tends to produce generic outputs that reflect the average of existing web content.

The more specific and contextual the challenge or company strategy, organisational capabilities, data maturity, and business context, the less useful a pure AI-generated response becomes. As a tool for refining and stress-testing your own thinking, however, it is genuinely valuable.

On the other side, where are claims about generative AI’s benefits more noise than substance?

Two areas stand out.

The first is large-scale content generation such as producing emails, newsletters, and communications at volume. If an organisation is generating significant content with AI while recipients are using AI tools to summarise it, it is worth asking whether any real value is being created in that cycle.

The second is the argument that AI can replace junior employees. I regularly hear the claim that generative AI can perform the work that previously required early-career hires. I disagree strongly.

These tools cannot replicate the fresh thinking, genuine challenge, or unique perspective that junior colleagues bring; they produce an aggregate of existing content.

Beyond the immediate quality concern, there is a strategic problem: if organisations stop bringing in junior talent, they undermine their own pipeline of future senior expertise.

Are you seeing LLMs genuinely embedded into enterprise workflows anywhere?

Legal departments are a particularly interesting case.

One might expect them to be among the last to adopt, given the precision their work demands. In practice, however, a significant part of legal work involves searching large document repositories, finding relevant references, and summarising information across lengthy materials, which are tasks at which LLMs excel.

The use case is not about replacing legal reasoning or decision-making; it is about navigating a 500-page contract efficiently and surfacing the information that matters. That frees legal professionals to spend their time on work that genuinely requires human expertise and judgement.

Is that adoption coming from individuals using consumer AI tools, or from specialist legal AI vendors?

Both, but data privacy is the critical constraint.

Most commercial AI solutions depend on large technology platforms, and for many organisations, particularly public sector bodies and regulated industries, routing sensitive data through those platforms raises significant compliance concerns, especially around cross-border data access.

The opportunity for specialist startups lies precisely here: developing solutions where data is hosted locally or within a specific jurisdiction, satisfying the regulatory requirements that major platforms currently cannot.

In regulated and public sector contexts, that capability is becoming a genuine differentiator.

How exposed are businesses to legal risk around copyright, patents, and privacy when using AI?

This is an area of rapid legal development, but one important misconception is worth addressing directly.

AI-generated content is not straightforwardly protected as intellectual property. Current frameworks in most jurisdictions require meaningful human involvement for copyright or patent protection to apply.

If you generate a logo entirely with AI, another company could generate the same output, and your legal recourse would be limited.

I would encourage leaders to think carefully about how they use AI in any context where they intend to protect the resulting work. The tool can support the creative process; it should not be the sole author of something you plan to own.

Any final advice for data leaders?

Stay curious. Treat AI as a tool, and treat data as the differentiator.

The organisations that will create lasting competitive advantage are those that invest seriously in their own data, not those that simply adopt the most visible AI applications.

Make sure that everything you do with data and AI is genuinely creating value, not just demonstrating activity.

Trust, culture, and the human dimension of change are as important as the technical infrastructure.