Why AI Projects Start With Content, Not Technology

AI projects often begin with the wrong question

When organisations start thinking about AI, the conversation often begins with technology. Which AI model should we use? Do we need a chatbot? How could we add AI search to the website? Should we connect an AI system to our CMS?

Those are reasonable questions, but they are rarely the best place to start.

Most useful AI applications depend on something much more familiar: information. An AI system needs something reliable to search, interpret, summarise, connect or act upon. If the information behind it is inconsistent, outdated, duplicated or poorly organised, changing the technology will only solve part of the problem.

This is why many AI projects are also content projects.

Before deciding how AI should work, it helps to understand what information it will need, where that information currently lives and whether it is in a form that can support what you want the AI to do.

AI is only as useful as the information it can work with

AI systems can do impressive things with large amounts of information, but they do not remove the need for that information to be trustworthy.

Imagine an organisation wants to introduce an AI assistant that answers questions about its services. The technology might be perfectly capable of finding and summarising information, but what happens if the website contains three slightly different descriptions of the same service? What if an old PDF contradicts a newer page? What if important eligibility information is missing altogether?

The AI still has to work with what it has been given.

Depending on how a system is designed, there are ways to select sources, retrieve relevant information and control which content an AI application can use. But technology cannot automatically decide what an organisation considers authoritative, resolve every contradiction or fill gaps where reliable information does not exist.

The quality of the application therefore depends partly on the quality of the information behind it.

This applies to far more than conversational assistants. AI-powered search, content recommendations, internal knowledge tools, automated content workflows and other AI features all rely on information being available in a form the system can use.

What do we mean by content in an AI project?

Good AI needs more than good writing

Why poorly organised content creates problems for AI

Structure gives AI more context

Structured content can make the meaning and relationships within information more explicit.

Instead of treating a service page as one large piece of content, for example, a website might separately understand its:

• Service name
• Description
• Intended audience
• Related sectors
• Locations
• FAQs
• Related services
• Case studies
• Contact information

Metadata and taxonomy can add further context by describing what content is about and how different pieces of information relate to one another.

This does not mean every sentence needs to become a database field. Too much structure can make a CMS frustrating to use and difficult to maintain. The goal is to identify information that has a meaningful reason to be structured.

A useful way to think about the progression is:

Content provides the information. Structure identifies what that information represents. Context helps systems distinguish and connect related information. The AI application can then use those foundations to provide a particular experience.

The more explicit those foundations are, the less ambiguity there is when an application retrieves and uses that information.

AI can expose existing content problems

Planning an AI project often reveals information problems that were already there.

Questions that previously seemed unimportant can suddenly become difficult to avoid.

Which version of this policy is current?

Are these two services actually different?

Why do three departments use different names for the same process?

Who is responsible for updating this information?

Which source should be treated as authoritative?

Why is an important piece of guidance only available inside a five-year-old PDF?

These are not really AI questions. They are questions about content governance, information architecture and organisational knowledge.

That distinction matters because implementing an AI tool will not resolve them by itself. Before asking a system to find and combine information automatically, an organisation needs to make some decisions about which information can be trusted and how it will be maintained.

The aim is not simply to provide an AI system with more information, but to give it access to the right information for the job.

Start with the use case, then work backwards

Starting with content does not mean reviewing every page, document and database before deciding what to build.

A more practical approach is to start with a specific user need and work backwards.

The process might look something like this:

Imagine the goal is to help prospective students find suitable courses using natural-language questions.

Rather than immediately choosing an AI search platform, you could first ask what information would be needed to answer those questions well. That might include subjects, qualifications, entry requirements, locations, study modes, course durations and application details.

You can then look at where that information currently lives.

Is it structured consistently across every course? Are important details hidden in free text? Are the same fields described using different terminology? Is anything missing? How often does the information change?

Those answers give you a clearer idea of what the project actually requires. At that point, technology choices become easier to make because you understand the job the technology needs to perform and the information it will be working with.

What should you review before choosing the technology?

The exact requirements will depend on the project, but there are several useful questions to ask before making significant technology decisions.

What information does the AI need?

Start with the intended experience. What questions should it answer, decisions should it support or tasks should it help someone complete?

This gives the content review a clear boundary and helps avoid auditing information that has little relevance to the project.

Where does that information live?

It might all exist in the website CMS, but it could equally be spread across documents, internal systems, databases or third-party platforms.

Understanding the sources can reveal integration requirements before technology choices are made.

Can the information be trusted?

Look for contradictions, duplicates, missing information and content that has not been reviewed recently.

Where several sources exist, it should be clear which one is authoritative.

How is the information structured?

Some content may already be stored in useful fields and categories. Other information might exist only inside long passages of text or documents.

The aim is not to structure everything, but to understand where additional structure would make the information easier to find, reuse or interpret.

Who owns the information?

An AI application needs more than a good set of information on launch day. Someone needs to keep that information accurate.

Ownership, review processes and publishing workflows therefore matter just as much as the initial content review.

What happens when the information changes?

If a service changes, a policy is replaced or a new product is introduced, how quickly will the AI application have access to the new information?

A system that starts with accurate information but gradually falls out of date will become less useful over time.

How will you know whether it works?

Good source information provides a stronger foundation, but it does not guarantee that an AI application will produce useful results.

Once the application is working, it needs to be tested against realistic questions and tasks. Does it find the right information? Does it use that information accurately? What happens when the answer is missing, ambiguous or spread across several sources?

Testing those situations can reveal problems with the application itself as well as gaps in the underlying content. It also gives teams something more useful to judge than whether an AI-generated response simply sounds convincing.

You don’t need to fix everything before you start

The idea of reviewing content can make an AI project sound much bigger than it needs to be.

Not every organisation needs a perfectly structured content estate before it can experiment with AI. In many cases, trying to clean up everything first would create a large project with no clear end.

The better approach is to focus on the information required for a particular use case.

An internal assistant designed to answer questions about HR policies has different content requirements from AI-powered website search. A tool that recommends related articles has different requirements from a system helping customers compare complex products.

That is why AI readiness is rarely a simple yes-or-no question. An organisation may have everything required to test one idea while needing significant content, governance or technical work before attempting another.

Starting with a contained use case also makes it easier to learn. You can identify which content problems genuinely affect the application, improve those areas and use what you discover to inform future projects.

Better content foundations have value beyond AI

One of the advantages of starting with content is that the work does not only benefit AI.

Improving the way information is structured, categorised and maintained can support many parts of a digital platform.

It can make content easier to:

• Find through website search
• Filter and browse
• Reuse across different pages and channels
• Maintain consistently
• Connect to other systems
• Present through structured data
• Manage through editorial workflows
• Adapt as the website develops

This matters because individual AI technologies will change.

The model, platform or integration that seems appropriate today may not be the one an organisation uses several years from now. Well-managed information is much less dependent on a particular technology.

That makes content architecture, taxonomy and governance useful foundations in their own right rather than work that only makes sense if a particular AI project goes ahead.

Build the information foundation before the AI layer

AI can provide new ways for people to search, explore and work with information, but it does not remove the need to understand the information itself.

Before asking which model, platform or integration to use, it is worth asking a more basic set of questions: what does the application need to know, where will that knowledge come from and can we rely on it?

Those questions often reveal that the first stage of an AI project is not particularly futuristic. It involves understanding content, resolving inconsistencies, improving structure and deciding how information will be maintained.

The technology will continue to change. Organisations that understand and manage their information well will be in a much stronger position to make useful choices as it does.

How this translates into real projects