The promise of AI is making headlines, and it makes sense to pursue its possibilities for your organization. You’ve likely played around with ChatGPT or Copilot, seeing how it can be a starting point for common workplace tasks and using it to answer basic, unnuanced questions. But you also may have noticed that generative AI doesn’t always reflect the essence of your organization, the deep expertise of your employees, or the specifics of your strategy. Is there a better way to align the promise of artificial intelligence with your organization’s goals?
In fact, there is. The missing link between your organization’s industry experience and AI’s insight- and productivity-boosting power is a set of practices known as information architecture (IA). These practices, which include ontology, data architecture, analytics architecture, and knowledge management, construct the very fabric of your data. They capture your expertise, acumen, and strategy in a way that is machine-reusable — and ready for AI to harness. This allows you to take advantage of increasingly complex use cases for AI, such as inference insights and retrieval augmented generation (RAG).
What’s preventing AI from being a completely reliable problem-solver?
In the immediate, AI can be incredibly powerful at answering questions and synthesizing multiple sources of information, but there are aspects of the technology that can lead to pitfalls. These issues may negatively impact decision-making, so it’s important to be aware of them and find ways to mitigate their impact. Here are some things that could keep you from getting the maximum value from AI:
- Hallucinations: AI that’s based on large language models (LLMs) can’t always recall facts accurately. This is because they only memorize “pieces” of facts from the body of knowledge they are trained on, causing them to give answers that seem plausible but are in fact completely erroneous. For example, an LLM could be asked “When did Einstein discover gravity?” and respond with “Einstein discovered gravity in 1687”. Our human cognition can determine that the date is correct while recognizing that the discoverer was actually Isaac Newton. The AI, on the other hand, can’t figure that out without guidance.
- Observability: Most AIs are based on black-box models, meaning that while we may know what our most recent input is into an AI, we don’t have a trail for how it arrived at the answer or output that we receive back. Some models do provide chain-of-thought explanations of their results, but these chains also suffer from the hallucination effect. On its own, AI doesn’t provide a way to audit the results.
- Replicability: AI based on LLMs is non-deterministic. Essentially, the model’s behavior and the output or answer you receive can be different each time, even with exactly the same prompt. Sometimes, this is a simple formatting difference that doesn’t materially affect the value of the AI’s output, but there is often a difference in precision or detail that could be meaningful to your organization. AI alone can’t determine what is critical and what needs to be included in results every time.
- Generality: AI needs to be trained to memorize snippets of information, and most readily available AI tools are trained on a very broad knowledge base. This means that the answers you receive without adding additional training input are general knowledge and likely won’t reflect the detail or nuance of your specific organization. You can train models with your specific knowledge, but AI can’t structure those inputs in the best way on its own.
While these issues could be potential stumbling blocks, information architecture holds the key to making AI work reliably to serve your organization.
Information architecture’s role in rescuing the user from common AI pitfalls
As we discussed above, IA is a set of practices that can transform your organization’s specific knowledge into data that machines, including AI, can use. Investing in these practices sets you and your team up to get maximum value out of AI while avoiding the pitfalls discussed above. Let’s look at each element of IA individually with a focus on how it can bridge the gap between your organization’s needs and AI’s capabilities.
- Ontology: This practice is focused on representing the human understanding of your organization in structured data — translating what (and how) you and your employees decide, determine, and know into a format that can be used by the machines supporting your work. By formalizing this representation, we can support AI by telling it what pieces of facts are needed to answer a question, thereby mitigating hallucinations. For example, we encode that a “discovery” should have a “date” and an “agent,” which helps your AI return: “Gravity was discovered by Isaac Newton in 1687, not Albert Einstein.” This encoding also supports observability and replicability and is particularly good at supporting privacy and regulatory use cases, turning laws into actionable data.
- Data architecture: This practice is focused on designing and implementing consistent methods of structuring, organizing, describing, and governing data, leading to data assets that are ready-to-go for your AI use cases. Good data architecture provides the semantics around your data that AI needs to use to produce replicable, consistent, and organization-specific outputs. It is the underpinning to implementing any data-focused AI use case, and it mitigates observability, replicability, and generality.
- Analytics architecture: This practice is focused on designing an analytical environment, including the collection, storage, analysis, and availability of data. Mature analytics architecture ensures the right data is available for your AI to use at the right time, and that you can use those AI-created insights where and when you need them. It supports data-focused AI use cases that utilize multiple data sources and addresses issues with replicability and generality.
- Knowledge management: This practice is focused on the people, processes, and technology needed to create, organize, and share the collective knowledge of your organization. This forms the backbone of your organization’s depth of information by taking it outside of someone’s head and putting it into a location where it can be reused either by other humans or AI. It is a critical practice to address replicability and generality issues in AI application.
A data-focused solution for getting the competitive advantage of AI right away
Each of these practices helps support parts of your AI implementation; together, they provide a robust data layer that mitigates any pitfalls your organization may encounter in using AI. Best of all, you don’t have to wait to reap the immediate benefits of AI while implementing your IA foundation — you can run both in parallel and strategically scale each in tandem. You get the immediate competitive advantage of AI, while placing yourself and your organization ahead of the curve to take full advantage of AI’s future promise.