Over the last 9 months, we’ve been working with dozens of teams who are challenging what’s possible with AI. They always reach the same conclusion: AI models have become so powerful that they’re no longer a bottleneck. AI context, on the other hand, has become what determines success: the relevant data, guidelines, domain expertise, decision tracks, actions, systems, and transactions specific to an enterprise. AI cannot complete complex tasks without broad, accurate, and detailed context. When it doesn’t know, it hallucinates, providing a wrong answer with the same confidence as a completely accurate one.
The main challenge is that AI context is almost exclusively unstructured.
No platform born in the “age of analytics” was designed to systematically process it. The emails that explain why an insurance claim was denied, the call notes that capture what a client actually needs, the tribal knowledge that lives in your best people’s heads and disappears when they leave – all these are a goldmine of context for agents. And for years, this context was impossible to operationalize at scale. That’s no longer true.
Unstructured context is what determines how AI connects the dots.
It determines whether AI will reach the right conclusion and make consistent decisions. It determines if two different agents will agree on an outcome. Organizations that will harness their full data, their experts’ tacit knowledge, their historical decision traces, and then overlay this knowledge with contextual guardrails such as business strategy, risk appetite, and market intelligence – are creating a moat, an unfair advantage over their competitors.
In organizations designed around context, agents will connect the dots better, faster, and more consistently.
AI with context means your fate will be determined by your strategy and operational excellence, not by the competency of a single AI model. The moat is the context you provide, and how well you’ve built the systems to make that context work at scale. Let’s dive into this idea.
The constraint is no longer the model
AI chatbots were primarily dependent on data quality, and their impact was limited. Nowadays, every single organization that we work with builds “expert AI” agentic systems that strive to complete complex tasks, autonomously.
Surprisingly, in the lab environment, modern models knock most complex tasks out of the park: with 100% information available, no noisy data to interrupt, all relevant guidelines provided, edge cases handled, extra relevant context meticulously crafted.
But when you reach production, the story changes completely. Real enterprise data is partial, noisy, messy. Models perform poorly in identifying what context to rely on without guidance. In healthcare for example, critical clinical signals like symptom progression, comorbidities, and treatment responses are locked inside free-text physician notes. Interpretation depends on the task at hand, and is never completely accurate or consistent at scale.
AI context is almost exclusively unstructured, and untapped
The same dynamic plays out across virtually every enterprise. Over 80% of organizational knowledge remains trapped in unstructured formats: PDFs with implicit hierarchies, fragmented email threads, call transcripts with decision information, documents containing cross-references that are never resolved.
Existing approaches fail here:
- ETL tools, semantic layers, and data catalogs cannot handle unstructured data – they never read through the unstructured content and miss the substance. They cannot get it AI-ready because they don’t know what’s inside.
- PDF extraction tools (old or modern, IDP or agentic) will get the content, but don’t provide AI context – simply because they operate at the datapoint level, not the dataset level. They cannot tell an AI agent that the amount in the invoice it’s reading was corrected by an email that arrived 5 minutes later.
- Embeddings and RAG erase the nuance and fail at data completeness – Embedded data limits the AI agents’ ability to dive into the details. Search is good for finding the top K results, but cannot guarantee retrieval of 100% of the results. Imagine running due-diligence on a Private Equity deal and reviewing just 10% of the documents.
MCP (Model Context Protocol) moves in the right direction by standardizing how models interact with external systems. But it operates on the assumption that the underlying data is already coherent, connected, and semantically aligned. That assumption rarely holds.
The result is AI systems operating on degraded inputs, producing outputs that are confidently wrong.
The architectural bet: agentic-first context infrastructure
The architecture for making enterprise data truly AI-ready is agentic-first, not bolted on. It’s not a prompt pipeline. It’s infrastructure built to continuously produce, maintain, and serve the context that agents need to act. LLMs and VLMs are integrated into a single execution layer, so the system natively handles textual correspondences, documents, tables, and visuals as part of one reasoning flow.
An agentic-first context architecture involves at least five distinct components:
- Data unification – translating different context sources into a unified data structure makes it easy to build an ever-growing context map: AI applications shouldn’t care whether a call transcript was generated by Zoom or by Gong, or whether a document was originally a docx or a scanned PDF. Unified data belongs in tables, not vectors.
- Data preparation – when data is being consumed thousands of times as context, it should be pre-processed once, and optimized for AI: cleaned, deduplicated, structured. For example, in an email thread that repeats the same content twenty times, eliminating the noise saves tokens, removes biases, and improves accuracy.
- Domain-specific grounding – autonomously discovering, defining, and mapping your organization’s unique jargon and domain-specific terminology into an agent-friendly AI knowledge hub has huge impact: consistency. Just like the structured semantic layer, enterprise AI needs to agree on business terminology. Is the “Project Phoenix” that appears in a Slack message, a PDF and a budget spreadsheet the same project?
- Data contextualization – Linking related information across sources and resolving missing context reduces errors dramatically. For example, connecting a CRM note with the corresponding call transcript, an insurance claim with the extra details provided over email, or the table on PDF page 13 with its description on page 18.
- Data curation – AI agents shouldn’t have access to all the data, all at once. Using LLMs to turn huge datasets into smaller, accurate and relevant datasets improves accuracy, reduces token usage, and improves latency. A procurement AI agent pointed to the relevant 15% of all contracts classified as “vendor contracts” will rarely make up information.
The architectural implication: delivering context directly to the existing enterprise data environment
The AI context engine belongs as a part of your existing ecosystem, where your data and AI teams already work. It sits between raw enterprise data and the AI systems consuming it. It connects to SaaS platforms, document stores, and data platform, ingests unstructured data, automates data preparation and context engineering and maintains a continuously updated semantic graph of entities, relationships, and events that AI systems and teams can query directly.
The legacy System of Record was a filing cabinet designed for human retrieval. The context layer is a living system designed for both AI and humans: it gives humans full lineage to trace the why, and gives models the high-fidelity structure to execute the how.
When AI confidence becomes the enterprise’s competitive advantage
Within a few years, access to capable models will be fully undifferentiated. The question that will separate winners from everyone else is: what are those models given to work with?
Your operational context – how your business runs, what your teams know, how your systems connect, what your history reveals – is the one advantage competitors cannot replicate by upgrading an API tier. AI is only as good as the context it operates on. That context is uniquely yours.
