I'm Kozo Moriguchi, CEO of Kawamura International.

We started in translation and localization, moved into MT SaaS, and now I'm focused on building out the LDX business. I'm not an engineer — but I test and rebuild business processes with AI, and I write about what actually happens.

This is Part 5 of that series. The question this time: can you build a reliable management dashboard by just handing your meeting minutes to an AI?

The context: most companies can only use Copilot

When you try to bring AI into enterprise workflows, you hit a wall quickly.

Many organizations — due to security policies or vendor agreements — restrict which AI tools employees can use. Claude, ChatGPT: blocked. Copilot: allowed. I've seen this pattern repeatedly across companies we work with.

That's why I ran this test. If Copilot is all you have, what can it actually do? And how does it compare to using LDX hub StructFlow to extract structured data first?

I used 20 departments worth of real meeting minutes and ran both approaches on the same data.

What we tested

Copilot approach
Paste meeting minutes into Microsoft Copilot. Ask it to organize the data into an HTML management dashboard. No schema defined. Let Copilot decide what's important.

StructFlow approach
Define a schema upfront (tasks, risks, cross-department requests as structured JSON). Run StructFlow via LDX hub API. Use Power Automate to generate an HTML dashboard with Chart.js charts. Save to SharePoint. Return the URL to Copilot Studio.

Same input data. Same goal. Different method.

See the dashboards yourself

Rather than just describing the difference, here are both dashboards built from the same 20-department dataset.

Copilot-generated dashboard (sample)
👉 Open dashboard

Copilot + LDX hub StructFlow dashboard (sample)
👉 Open dashboard

These pages are unlisted (no-index) and contain no real company data. URL access only.

Looking at them side by side, the difference in information density becomes immediately visible — before you even look at the numbers below.

The numbers

From the same 20 departments of meeting notes:

• 5.5× more tasks extracted with StructFlow
• 2.6× more risks surfaced
• 4.9× more cross-department dependencies identified

Copilot didn't fail to read the documents. It read them, summarized them, and produced a polished HTML output. But in the process of summarizing, an enormous amount of information was compressed away.

What specifically got dropped

The quantitative gap is significant. But what matters more is what fell through.

Here's what Copilot's dashboard was missing that StructFlow caught:

① Operations team overload — 3 simultaneous cases, high attrition risk
StructFlow: extracted as 3 separate structured task items with named assignees and risk flags.
Copilot: compressed to "resource reallocation needed."

② Finance team: M&A accounting and audit response
StructFlow: specific tasks with deadlines.
Copilot: not mentioned.

③ Partner team: top 3 partners = 70% of revenue (concentration risk)
A significant business continuity risk.
Copilot: not mentioned.

④ QA team: 2 security vulnerabilities
StructFlow: extracted as independent risk items.
Copilot: buried in another department's section.

⑤ Facilities: monthly office rent at ¥32M
A concrete cost figure.
Copilot: omitted entirely.

The dangerous part isn't that these were summarized — it's that you wouldn't know they were missing. Looking at a clean Copilot-generated dashboard, there's no indication that any of this exists. The gap only became visible when we had the StructFlow version to compare against.

Why "just ask the AI" fails structurally

There are four structural reasons why unstructured AI extraction breaks down for management reporting:

1. Summarization pressure
LLMs are optimized to produce readable output. Given a long document, they compress it using their own judgment about what matters. That judgment is based on training data — not your company's priorities.

2. No consistent format = no trend analysis
Copilot's output format changes every run. If last month's dashboard has different columns than this month's, you can't track trends. A management dashboard needs consistent structure by definition.

3. You can't see what's missing
This is the most dangerous failure mode. A well-formatted summary looks complete. You only discover the gaps when you have a structured comparison. In practice, most teams would never run both approaches side by side — so the gaps stay invisible.

4. Doesn't scale
StructFlow processes 20 departments at the same cost as 5. Ask Copilot to handle 30 departments in one prompt and output quality degrades. The approaches scale in opposite directions.

Where Copilot genuinely wins

To be fair: Copilot has real strengths.

• KPI snapshots — organizing revenue, churn rate, and utilization into a summary view is something Copilot does quickly and well.
• Design quality — the HTML output looks polished. Good for executive presentations.
• Narrative context — Copilot understands "A company deal → technical proposal → executive meeting" as a sequence. It writes it as a coherent story.

The honest summary:

• For "quickly get a sense of things before a meeting" → Copilot
• For "make sure nothing critical is missed" → StructFlow

These aren't competing tools. They serve different purposes.

The pipeline that's running now

Here's the StructFlow setup we built:

Meeting minutes (SharePoint files)
    ↓
Power Automate (batch processing flow)
    ↓
LDX hub StructFlow API (JSON structured output)
    ↓ polling every 3 seconds (Do Until)
HTML dashboard generation (with Chart.js)
    ↓
Saved to SharePoint → URL returned to Copilot Studio
    ↓
User views in browser

The key piece: Power Automate's "Create file" action writes the full HTML (including <script> tags) to SharePoint as a static file. Chart.js runs fine in that context. Copilot Studio returns the URL to the user.

Total setup time from zero: about 3 hours, including all the wrong turns.

What's next: multilingual dashboards

LDX hub has another API alongside StructFlow: RefineLoop.

RefineLoop is designed for translation quality refinement. But combined with StructFlow, it opens up an interesting use case: companies with global operations often have meeting notes in multiple languages — English, Japanese, Spanish, Mandarin. StructFlow extracts structured data from each. RefineLoop normalizes and quality-checks the translation into a unified language. The result: a single dashboard aggregating data from every regional office, regardless of what language their notes were written in.

Given that Kawamura International's roots are in translation and localization, this is a direction we're particularly interested in building out.

Coming up: n8n and Dify

I will run the same test on n8n and Dify.

Power Automate is a Microsoft-native tool. n8n is an open-source workflow automation platform you can self-host. Dify is an LLM app development framework. Same StructFlow API underneath — but the surrounding infrastructure changes. I want to know what that changes in practice.

If you're locked into a Microsoft environment, the setup from this post should work for you. If you have more flexibility, stay tuned for Part 6.

Kawamura International is actively testing AI integration across translation and localization workflows. We publish findings through LDX Lab as we go.