To set the scene, let me introduce my professional context. I work at Argusa, where we help organizations make sense of their data and extract insights to improve operational efficiency, support strategic decision-making, and monitor risks.

A typical project looks like this:

A client, let’s call him Roy, is head of analytics at an international organization and is struggling to deliver analytics on time. Dashboards are slow to create and often inconsistent. The deeper issues are threefold:

1. User requirements are often ambiguous, and little thought goes into governance; resulting in duplicated efforts and inconsistent numbers.
2. The data engineering and pipelines needed for robust and timely analytics delivery are not fully in place.
3. The organization lacks a strong data culture. Upskilling and coaching business users to become data-fluent remains limited.

As a consulting company, we are often brought in mid-project, which limits our ability to shape overall strategy. Instead, our role typically involves:

1. Clarifying requirements, often by building quick prototypes and gathering feedback.
2. Constructing the data pipelines and models needed for dashboards.
3. Coaching and training business users.

I have previously written about why self-service analytics has largely remained a dream (link). Here, I would like to argue that AI, despite the buzz, depends on solid data foundations. As things stand at the moment, I believe that nearly all AI projects are, at their core, still data projects, that fall into three categories:

I. Performing Analytics with AI

Let’s define analytics: it is the practice of creating context-rich dashboards using accurate data at the right level of granularity. A dashboard for a supermarket sales director will differ vastly from one for an inventory manager. Both, however, rely on solid data.

Modern dashboards, typically built with BI tools and deployed on analytics platforms, have come a long way from the static, PDF-style reports of the past. Today’s tools offer interactivity, better governance, and faster turnaround. Yet many vendors now claim dashboards are obsolete and replaced by “chatting with your data.”

There are three issues with this claim:

1. Underestimating Data Preparation

Transforming data to support natural language querying requires significant backend work. Vendors claim that “semantic models” as a turnkey solution, but these models often represent more than half the effort in dashboard creation. Poor data foundations limit chatbot capabilities just as they limit dashboards.

2. Lack of Contextual Validation  

Dashboards provide vetted, shared views that teams align on. Chat-based systems lack this built-in validation. To get around this limitation, some vendors are now proposing the use of predefined and pre-validated queries to add structure.

This mirrors the traditional dashboard model and does not solve the original dilemma of getting fast, reliable and context rich information about a business scenario.

3. Data Governance Complexity

Governance is inherently complex, and AI systems e.g. those using RAG (Retrieval-Augmented Generation) add new layers of risk. These concerns are magnified beyond simple pilot use cases, where governance risks can quickly become operational burdens.

II. Automating Repeatable Tasks with AI

This category is often misbranded as AI. In reality, many projects are just digital process automation: rule-based, deterministic workflows.

Platforms like Calendly, Eventbrite, Zapier, Asana, and Monday.com have existed for years:

• Calendly automates meeting scheduling.
• Event platforms automate RSVPs, reminders, and follow-ups.
• Project management tools assign tasks, send deadline reminders, and move items across workflow stages.
• Zapier integrates apps and automates cross-platform workflows.

These systems improve efficiency, but they aren’t “intelligent.” Adding LLMs to deterministic logic doesn’t add real value, it adds confusion and, eventually, disillusionment.

Another aspect is that in enterprise settings, automation tools provide only one piece of the picture. True insights come from combining data, e.g., syncing project timelines with financials, or linking Calendly events to CRM pipelines and revenue.

Automation has been around long before OpenAI, and often creates more demand for analytics and governance, not less.

Here’s why:

1. Automation increases complexity. You are doing more, many more interactions, more dependencies, more moving parts. For example, integrating Calendly introduces new needs like tracking no-shows or sending confirmations.
2. New tech means new stakeholders. Where a single person once handled the process, now IT and business teams share responsibility. Each stakeholder has different goals and perspectives, increasing coordination overhead.
3. Responsibility gets blurred. More stakeholders = more middle management. And management needs dashboards to monitor processes, performance, and accountability.

III. True AI Use Cases (LLMs, RAG and Agentic workflows)

Now we come to actual AI: Large Language Models (LLMs), RAG systems, and agentic workflows.

RAG combines an LLM with external data sources (documents, websites, structured data). Its success depends on both the quality of your data and your metadata (labels, taxonomies, access control). If your internal documentation is poor, RAG won’t help much.

Agentic systems take things further, using LLMs to both extract insights and initiate actions (like sending emails or updating systems). But these workflows are only as good as the data and governance underneath them.

As they involve multiple stakeholders and connect business processes across functions, agentic systems actually amplify the need for robust analytics and governance.

Without solid data pipelines and a culture of data quality, these systems will underperform, or create risk.

In Conclusion

If you have made it this far, thank you.

I may sound skeptical of AI, but I am not. I believe it holds immense value. However, I argue for measured, deliberate adoption, built on a solid use-cases and data foundations.

In short, I encourage you to:

1. Think data-first.Whether you’re building analytics or automation, start with high-quality, well-structured data and curated knowledge bases.
2. Make governance a priority.Both RAG and automation systems require rigorous data access controls. Start with governance, not after the fact.
3. Avoid using a nuclear solution when a Swiss army knife will do.Many use cases don’t require AI. Don’t complicate things (or increase your costs) unless AI is truly the right tool.

Since you are still reading, I hope you found this useful. I would love to hear how your organisation is approaching AI and what challenges you are facing.