Your AI strategy is only as strong as your knowledge architecture

Key takeaways

• Build the foundation before scaling AI
• Treat business definitions and workflows as governance
• Measure AI maturity by business value, not experimentation

In my first article, AI starts before the prompt, I wrote that trust in AI starts with reliable data, clear processes and people who know when to step in. In this article, I’d like to expand on this and talk about knowledge architecture.

Most conversations about AI today focus on productivity, prompt engineering or job replacement. While these are valid and important topics, they start too far downstream. 

When used well, AI can also help you build that foundation by sifting through messy information and creating a pathway to defining the architecture. In this manner, AI becomes not only a user of knowledge architecture, but also a practical tool for helping to build it.

When it comes to using AI effectively, I believe the next step for leaders is management discipline and organisational legibility. It’s thinking about how you can make your business understandable to machines while keeping it useful and empowering for people. 

AI tools are no longer the bottleneck 

Should we use one provider or another? Should we build directly into our existing tools or should we use an integrator that brings different AI models together?

It’s easy to assume that progress depends on choosing the right model but they’re not the first decisions I would make. The first thing I would ask is whether there is a reliable foundation for AI to work from.

For example, if information is spread across disconnected systems, or if business definitions vary among teams with multiple documents containing competing versions of truth, this becomes a knowledge architecture problem. 

Because even when taking an action, AI has to work out which source matters, which version is current and whether the most recent information is actually correct.

That’s why the challenge is bigger than choosing the right AI model. Before AI can act reliably, your knowledge needs to be coherent, clearly owned, maintained and connected, with agreed rules for resolving conflicts.

The next step then, is to design work so it is understandable to both people and machines.

What does machine-readable work mean?

Machine-readable work is work designed so both machines and people can use it effectively. At a practical level, it looks something like this: 

• Knowledge that is current, structured and connected.
• Policies and procedures are easy to retrieve and understand.
• Important decisions are not trapped in private inboxes or undocumented conversations.
• Systems use agreed definitions rather than competing versions of the same metric.
• Documents contain information in formats AI can reliably interpret.

In the past, documentation quality or spreadsheet structure could feel like an administrative concern. An employee could usually work around a confusing layout, infer what a merged cell meant or ask someone to explain the context behind a chart.

However, AI is less forgiving. 

A document with complex formatting, merged cells in a text file/image, or important information embedded in images may be difficult for a model to interpret consistently. What looks obvious to a human may not be obvious to an AI system and this changes the weight of everyday work practices. That’s why the way your organisation captures information affects what AI can automate, what it can retrieve and what it can trust.

Knowledge design is a leadership challenge 

While it may sound basic and boring, this is something I stress a lot because when information is fragmented, the cost isn’t just operational; it also affects decision-making.

Consider an organisation where the Customer Relationship Management system contains one version of a customer relationship, the sales team has another version from their call-recording tool and finance has a different view of the customer numbers. 

Whilst an AI system may be able to identify the disagreements, it can’t always determine which source should win. If that logic is not yet established, then AI is left to decide, turning it into a governance decision issue. 

Leaders need to think about knowledge architecture in the same way they think about organisational design:

• Who owns the information? 
• Which definitions are authoritative? 
• How are changes approved? 
• What should happen when two sources disagree? 
• Which information can AI use, and which information requires additional controls?

Being AI-first or AI-forward does not mean having AI do everything. It means creating systems, processes and policies in a way that allows AI to contribute safely and effectively. It includes technology and it also includes how people work and how information moves through the organisation.

The journey from AI assistance to orchestration

When it comes to AI maturity, I like to break it down as a five-level progression journey. 

Stage	What AI is doing	What organisations need
Stage 1Assistance	Helping with isolated tasks such as drafting, summarising or analysing	Clear, usable information
Stage 2Connection	Working across conversations, systems and longer contexts	Connected sources and consistent definitions
Stage 3Action	Completing tasks and running repeatable workflows	Permissions, guardrails and reliable processes
Stage 4Judgement	Comparing options, applying best practice and surfacing recommendations	High-quality data, feedback and human oversight
Stage 5Orchestration	Coordinating the way a role or organisation operates	A coherent knowledge architecture and operating model

I want to highlight that the later stages depend on the earlier ones. This means we can’t take ‘shortcuts’ and jump straight to orchestration by adding more sophisticated prompts. If our foundations are weak, the system will simply operate with greater speed and scale, but with the same ambiguity.

This is why even though many organisations are able to demonstrate useful AI experiments, they may still struggle to achieve consistent value because their operating environment is not ready. 

So what does an AI-first operating model look like? 

What an AI-first operating model looks like

Systemising AI is not just about choosing an AI model or rolling out a chatbot. It’s deciding what information is authoritative, who owns it, how it is maintained and where human approval is required. 

For example, in my world of business revenue and growth, this includes how we:

• define ownership and connect HR, CRM, finance, customer and operational systems
• structure documents, capture information and maintain knowledge bases
• choose tools, protect private information and give people a clear way to review and approve outputs

The same principle applies to organisational knowledge. 

At ELMO, we’re building a more reliable knowledge base so people and AI can find current policies, processes, product information and workforce knowledge. The goal is to create a stronger system around the tools we use. 

Here’s how we’re doing it: 

1. Establishing a connected knowledge layer 

Information doesn’t have to sit in one physical system, but it needs to be connected, governed and maintainable.

2. Agreeing on the same definitions

If teams use different meanings for the same customer, revenue or performance metric, no AI model can reliably reconcile the difference without additional instruction.

3. Workflows designed for both human and machine

This means considering how information is captured, structured, updated and handed from one system to another.

4. Setting clear boundaries for human judgment

AI can surface patterns and recommendations, but people still need to own decisions where context, sensitivity or consequence matter.

5. Bring your people along the journey 

People need to understand why AI is being introduced, how it will be used and what skills they need to develop. A technically strong system will not create value if people do not trust it or know how to work with it.

Envision and build your own Mission Control 

Today, people move between Slack, email, CRM, HR systems and other platforms to piece together information they need to make decisions. The information may exist, but it is fragmented across channels, systems and teams.

A more mature AI environment would work like Mission Control: bringing together signals from across the organisation, surfacing what matters and highlighting the risks or decisions that need attention. Instead of asking people to monitor every channel and connect every signal themselves, it would give them a clearer view of the work in front of them.

AI could handle more of the searching, sorting and routine interaction. People could spend more time on what cannot be automated: applying judgement, building relationships and making decisions in context.

That’s the long-term vision I keep coming back to. AI is not replacing the people doing the work, but it is a trusted operating layer that helps them see what matters and focus on the decisions that require human judgement.