The Unstructured World: Why the CRM Vision of AI Won't Scale

A response to the "AI just needs better structured data" thesis

---

There's a persistent fantasy in enterprise software: if we could just get everyone to put everything into the right fields, in the right systems, with the right schemas, then AI would work.

This is the CRM vision of AI. Build better systems of record. Force better data hygiene. Structure everything. Then unleash AI on your beautifully organized data lake and watch productivity soar.

I understand the appeal. Structured data is tractable. You can query it, join it, verify it, audit it. Twenty years of enterprise software have been built on the premise that business value comes from structuring previously unstructured processes.

But this vision fundamentally misunderstands both how work actually happens and what AI is actually good at.

---

The Myth of the Structured Enterprise

Let's be honest about what percentage of actual work product lives in systems of record.

At any company, the most consequential decisions, the key institutional knowledge, the actual reasoning behind outcomes—where does this live?

• Not in Salesforce. The CRM has the deal amount and close date. It doesn't have the strategy session where you decided to discount, the competitive intel that shaped positioning, or the relationship dynamics that actually closed the deal.

• Not in Jira. The ticket system has the task and its status. It doesn't have the architectural discussion in Slack, the whiteboard diagram that never got photographed, or the context from the customer call that changed the priority.

• Not in the knowledge base. The wiki has the official process document from 2022. It doesn't reflect how the team actually does the work in 2025, the workarounds everyone knows, or the institutional knowledge that nobody thought to write down.

The valuable context is in Slack threads, email chains, meeting transcripts, shared documents with inline comments, Loom videos, Figma annotations, and—critically—the knowledge in people's heads that never gets externalized at all.

Systems of record capture outcomes. Work happens in unstructured channels.

---

Why "Just Structure It" Doesn't Work

The response is always: "So structure it! Better tooling, better processes, better data governance."

This argument has been made for forty years. Every generation of enterprise software promises that this time, adoption will be different. This time, people will actually fill in the fields. This time, data quality will be high.

It never happens. And it's not because people are lazy or undisciplined. It's because the cost of structuring exceeds the value to the person doing the structuring.

When a salesperson documents their deal strategy in detail in Salesforce, who benefits? Future salespeople who might learn from it. The AI system that might use it for training. The analytics team that wants cleaner data. The salesperson themselves gets almost no immediate value—they already know what they did and why.

So they don't do it. They do the minimum required to close the deal in the system and move on to the next opportunity. This isn't a discipline problem. It's an incentive alignment problem, and no amount of process mandates fixes it.

---

The AI Paradigm Shift

Here's what changes with AI: AI is natively good at unstructured data.

This is the most underappreciated capability of large language models. They can read a Slack thread and extract the decision that was made. They can watch a meeting transcript and identify the key commitments. They can parse an email chain and understand the negotiation dynamics.

You don't need to structure the data first. The AI can do the structuring on-demand, in context, for the specific question being asked.

This inverts the enterprise software paradigm. Instead of:

Old model: Force humans to structure data → Store structured data → Query structured data

New model: Capture unstructured data → AI structures on-demand → AI produces structured outputs when needed

The second model is dramatically more practical because it doesn't require changing how humans work. People can keep using Slack, email, documents, and meetings the way they naturally do. The AI layer handles the structuring.

---

What This Means for Enterprise AI

If unstructured data is the input and AI does the structuring, then the bottleneck shifts from "data quality" to "data capture."

Most AI context is currently lost because:

1. Conversations aren't recorded: The Zoom call where the decision was made wasn't transcribed
2. Ephemeral channels dominate: The Slack thread where the rationale was discussed ages out and becomes unfindable
3. Cross-system context is invisible: The connection between the email thread and the Jira ticket and the Figma file exists only in people's heads
4. Personal context isn't captured: The ChatGPT conversation where someone worked through the analysis vanishes when the browser tab closes

The opportunity isn't better systems of record. It's universal capture of the unstructured context where work actually happens.

---

The Context Engine Approach

This is why we built Context Engine as a log drain, not a knowledge graph.

A log drain captures everything:

• Workspace sessions (the AI conversations where work happens)
• Meeting transcriptions (where decisions are made)
• Communication traces (where rationale is discussed)
• Document interactions (where analysis is performed)

We're not asking people to structure their work product. We're capturing the unstructured trace of work as it happens, and providing AI-navigable access to that trace.

When someone asks "how did we handle the pricing objection from Acme Corp?", the answer isn't in a structured field somewhere. It's distributed across:

• The call transcript where the objection came up
• The Slack thread where the team discussed response strategies
• The email where the revised proposal was sent
• The Workspace session where the pricing analysis was done

Context Engine captures all of this. AI structures it on-demand to answer the specific question.

---

The Future Is Unstructured-First

The companies that will win in enterprise AI are not building better structured databases. They're building better capture for unstructured context.

This requires:

1. Presence in workflows: You can't capture context you're not present for
2. Universal ingestion: Every source where work happens needs to flow in
3. AI-native organization: Structure emerges from AI understanding, not human data entry
4. Permission-aware access: Unstructured data often contains sensitive information

The CRM vision assumes humans will adapt to what computers need. The AI vision recognizes that computers can finally adapt to how humans actually work.

The unstructured world isn't a problem to be solved through better data governance. It's the natural state of human collaboration, and AI is finally capable of working with it directly.

Stop trying to structure the unstructured. Start capturing it.

---

This is part of a series on rethinking enterprise AI at context.inc