Building tomorrow's intelligence on yesterday's mess

Building the invisible foundation

We spent eight weeks doing the work nobody sees but AI desperately needs.

Step 1: The content reality check

We evaluated thousands of pieces of content. Not skimmed. Actually evaluated. Using classification with AI to start to chip away at the deep dives to spot patterns. Then scrubbed them again against criteria that would matter to machines:

The hard part? The people part.

Their advisors were brilliant people drowning in garbage information architecture. One advisor told me: "Sometimes you scroll endlessly to find a form." Another just gave up: "I keep my own folder of the five things I actually use."

After interviewing 15 advisors, the issue wasn't what anyone expected.

The problem wasn't that the content was badly organized. The problem was that three completely different types of advisors were trying to use the same system.

The patterns were striking. Legacy advisors had admins to look up things, they needed notifications, not navigation. Overwhelmed users were teams of one who needed filtering, not more features. Pioneers knew the tools (and were the meatspace version of the search engine for others) and needed guidance, not access.

Step 2: Creating AI training data (without calling it that)

Here's the secret: AI doesn't need perfect content. It needs coherent patterns. So we built them:

The taxonomy was the first place to establish context and structure:

Simple, right? But now AI can understand:

• "Show me disability claim forms"

• "What do advisors need for new accounts?"

• "Find everything about critical illness for clients"

We didn't just organize content. We taught it how to be found.

Step 3: The prompt library nobody knew they needed

Everyone thinks AI just magically knows how to write like your brand or your audience. Spoiler: it kind of, well, doesn't.

We built an initial 9+ prompt templates for different scenarios:

• Regulatory updates (lawyer-approved language)

• Advisor communications (conversational but compliant)

• Client materials (simple but not patronizing)

Each prompt had:

• Context instructions

• Tone guidelines

• Compliance boundaries

• Quality checkpoints

This isn't sexy work. But it's the difference between AI that sounds like a robot reading regulations and AI that actually helps humans. It the difference between guessing and learning.

The framework that made it scalable

We didn't just fix their content. We built systems that could fix themselves:

The content scorecard

Every piece of content was graded on three dimensions:

• Content quality (45%) - Is it actually good?

• Technical health (30%) - Can systems use it?

• Business impact (25%) - Does it matter?

But here's the clever part: we built it as an algorithm. Now they can:

• Auto-score new content on upload

• Flag problems before they publish

• Track quality over time

• Train AI on what "good" looks like

The governance that doesn't suck

Most governance is where good ideas go to die in committee. We made it simple:

Three questions for every piece of content:

1. Can AI categorize this automatically? (If no, fix the taxonomy)

2. Would you trust AI to summarize this? (If no, fix the structure)

3. Could AI find this based on user intent? (If no, fix the metadata)

That's it. No 47-page governance document. Just three questions that the governance committee could use as starters to keep AI honest. Partnering with legal and compliance teams, I established:

Governance Principles:

• Transparency: All AI-generated content clearly labeled

• Accountability: Human review required for customer-facing content

• Fairness: Bias detection and mitigation processes

• Privacy: No PII in training data or prompts

• Security: Encrypted prompt storage and access controls

Then to move decisions along, we defined some benchmarks on what “good” is: