Structured Data: The Hidden Force Multiplier for AI

Aug 21

Every project manager knows the drill: you’ve been asked to prepare the slide deck for the next Program Management Review. You’re handed a chaotic dump of spreadsheets, SharePoint documents, and Jira screenshots and told to “summarize the data and provide analysis.” None of it lines up cleanly. Dates don’t match, structures are inconsistent, and key details are buried in free-form fields.

We often imagine AI as the magic fix for this problem—drop the data in, ask a question, and let ChatGPT, or another large language model (LLM), untangle the mess. Sometimes it works beautifully. But other times the AI misses key points, merges things together that don’t belong, or even invents details out of thin air.

This isn’t a bug. It’s a symptom of how AI sees the world.

Why AI Struggles with Loose Data

LLMs are probabilistic pattern‑matchers. They’re trained on oceans of human language and are brilliant at generating text that sounds right. But when you hand them semi‑structured inputs—like a table in a Word document—the model has to guess:

Where does a row end?
Which column is the “real” label?
What’s noise versus what’s signal?

The result is often incomplete or inconsistent. Sometimes an AI answer looks polished but quietly drops an entire criterion or attribute. Other times, it blurs two ideas together into something that sounds plausible but wasn’t in the source at all.

The Example: From Word Table to JSON

We saw this firsthand when working with governance criteria from the TMS Foundation Project Governance Framework (PGF). The PGF defines the core processes every project must follow to ensure scope, cost, schedule, and risk are managed in a consistent, auditable way. It’s the backbone of how we run projects — but capturing and documenting all of its criteria in a way that both humans and AI can reliably interpret is no small task.

Here’s a simple excerpt as it originally appeared in a Word table:

Word Table Input (Unstructured)

Attribute	Criteria Code	Criterion
SCM1 – Req.	SCM1.C1	A requirements list exists for each project
	SCM1.C2	Each requirement is uniquely identified
	SCM1.C3	Requirements are classified by type

When we asked ChatGPT to analyze this table, sometimes it ignored SCM1.C2, or blended SCM1.C3 into SCM1.C1. The formatting created ambiguity.

Then we transformed the same content into JSON:

JSON Input (Structured)

{
  "SCM1": {
    "criteria": {
      "SCM1.C1": "A requirements list exists for each project",
      "SCM1.C2": "Each requirement is uniquely identified",
      "SCM1.C3": "Requirements are classified by type"
    }
  }
}

With the JSON version, the AI always returned all three criteria correctly. No hallucinations. No omissions. It could even cross‑reference them with related governance processes automatically. The difference wasn’t in the model. The difference was in the structure.

Why Structure Works

Structured data gives AI a scaffold to reason with:

Constraints reduce ambiguity. Field names like SCM1.C2 can’t be mistaken or skipped.
Deterministic checks are possible. The AI can iterate keys in JSON instead of trying to parse a table grid.
Traceability is built in. Every output can be mapped directly back to a schema.

This is why modern governance frameworks insist on defining outputs—requirements, risks, budgets, work packages—in schema‑based, machine‑readable form, so that AI can operate on data that already has clarity built in.

The Broader Lesson

What happened with that governance table is the same story that plays out across industries:

McKinsey reports that AI‑driven operations forecasting—even in data‑light environments—can reduce error rates by 20–50%, slash lost‑sales and unavailability by up to 65%, and shrink warehousing costs by 5–10% and administrative costs by 25–40% (McKinsey & Company).
Techniques like Retrieval‑Augmented Generation (RAG) have emerged as powerful guards against hallucination by grounding LLM outputs in real data—significantly improving accuracy in structured outputs (arXiv).
An MIT Sloan study finds that AI initiatives often stall not for lack of compute or talent, but because organizations lack structured, accessible data—a “data problem” that stalls AI’s progress (MIT Sloan Management Review).
Warnings from TechRadar echo this: poor data quality, fragmented governance, and messy internal systems are the biggest barriers to AI success—not the AI itself (TechRadar).

In other words: it’s not just how powerful your AI is—it’s how well your data is structured.

Takeaway

Unstructured inputs make AI guess. Structured inputs let AI reason. That’s the real force multiplier.

When your systems—whether for governance, project controls, or portfolio reporting—capture data in consistent, schema-based formats, AI becomes a reliable partner instead of a creative guesser.

The Djobu case made this crystal clear: once governance criteria were transformed from Word tables to JSON, the AI stopped improvising and started delivering.

That’s a lesson worth remembering: if you want AI to work for you, give it the structure it craves.

Let me know if you’d like to adjust the tone, add visuals, or tweak any examples further!

Eric Christoph