The 7 data-steps for succeeding with an AI project

AI is one of those ubiquitous terms – it’s everywhere but what it means changes based on context.

So, if you want to roll out AI in your organization, you have to get specific. Most commonly, when people refer to AI in their everyday life, they’re talking about LLMs (Large Language models). LLMs are great with unstructured text (and other media) which is why we see a lot of use cases like

• Content Ideation & Creation

• Summarizing

• Editing & rewriting

• Coding assisting

• Data interpretation

• Personalization at scale

• Simulation & role-playing

• ...And more

So, if LLMs are great for unstructured data, what can you do with AI on your structured data?

AI and Structured Data

LLMs are not a natural fit for structured data like finance or project data, for a couple of reasons. First of all, LLMs are non-deterministic and base their answers on statistics and indexes, not the actual data like a report would be. That means the answer they provide can vary, and in the worst-case scenario, the model has made up part of the answers. You have no way of knowing if the model has made up data or identified wrong data. Even a 90% accuracy rate can be catastrophic when dealing with, for example, financial data.

Secondly, it’s difficult and expensive to train a model on large amounts of structured data. The alternative is to teach the model to query the data and interpret the answer, but this isn’t always easy if the data is broad with a lot of different tables, relationships, data sources, business logic, historical data, etc.

7 steps for making the most of AI

Data is the foundation for AI – but the kind of data makes a difference. More data means more options and more complexity. To get value out of structured data and AI, it takes an investment, but in my experience, it’s worth it.

Instead of hard-coding success criteria, you can create models that learn from your organization’s data and monitor progress and risks. It can monitor your portfolio, flag what you should pay attention to, interpret complex datasets, and translate the technical output into practical, human insights. In other words, it tells you what it means and what to do next. With Agentic AI, you can have a proactive assistant that activates the right tools for the next move.

So, here are my 7 steps for succeeding with AI

1. Start with a specific use case that creates value: because it’s a big investment, you should be sure up front that the investment matches the potential value. Plus, a well-defined scope and result increase the chance of success.

2. Create a dataset containing only the data needed to fulfill the use case: When creating custom AI, you can control the data so it doesn’t learn from irrelevant or bad data – and the result is simpler to evaluate.

3. Enrich the data with metadata: Add descriptions, meaning, relationships, query examples, and any other relevant metadata to help the model navigate the data, know what to look for so it can answer in a meaningful way

4. Provide specific definitions: Leave no room for the model to create its own definitions, so you can trust the output.

5. Instruct the model to never hallucinate! It’s better to not get an answer than to get a wrong answer.

6. Have the end-user validate the results: the end-user is the best person to evaluate whether the AI project fulfils the aim of the use case. It goes without saying that this is the iterative step; if adjustments are needed, go back to step 2 and ensure your dataset is complete.

7. Expand the solution: If the solution is successful, see if it can be expanded or scaled. Maybe there are similar use cases that follow the same guidelines.

Always ask how AI can contribute

When looking at your business needs, you should always be critical and evaluate if AI is the best tool for the job. It’s about activating AI where it has the most potential. Don’t waste resources on AI when there already exist tools and solutions that provide value. For example, “who are our top 10 customers based on revenue” is easily answered with a Power BI report – or the answer might even be in an existing report! Older tools like machine learning, for example, might be better for customer segmentation or forecasting. Especially if accuracy and consistency are key. In this case, AI could be the orchestrator bringing in the relevant tools for the job .

AI is an amazing tool. Make sure it’s solving the right problem.