AI Case Study

Takeaways

• 1.9PB of unsearchable content scattered globally across Azure regions.
• Graph deployed Azure AI + natural language search to index the archive.
• Within 83 days: $2.2M saved, 100% adoption, and search time cut to 30s.

As CTO of a global agency with 5,000 employees across 11 offices, he had a problem that was growing by 100,000 files a night. Every night.

Employees were spending hours searching through 1.9 petabytes of archived content. Not because they couldn't find files - but because they couldn't find the right ones. The product launch presentation? There were 47 versions across 12 folders.

Picture this: It's 4 pm in New York. A creative director needs the key shot from that Coca-Cola campaign. She knows it exists. She remembers approving it. Two hours later, she's reviewed 47 versions. None are right. The client meeting is tomorrow.

This wasn't an edge case. This was Tuesday.

"We had years of valuable content," Jason recalls. "But no one could find it when they needed it."

This wasn't just about search. It was about wasted time and effort, and recreating content that already existed - rather than being able to easily find and reuse it.

The numbers were brutal:

• 1.5 million files restored annually as content teams hunted for the right version
• Hot storage for 1.9PB globally? $1.1 million annually
• Manual tagging 60 million files? 23 work-years
• Growing by 100,000 files nightly? The problem doubles every 19 months

Traditional enterprise search was useless. File names meant nothing. Metadata was non-existent. The actual content - decades of work - was locked inside formats that search couldn't understand.

"We tried everything," Jason recalls. "Taxonomies. Naming conventions. Nothing scaled."

The system had to:

• Understand what was in the content
• Understand industry-specific file formats
• Understand groups of related files
• Work with cold storage economics
• Learn without manual training
• Live where people already worked

Most vendors were touting yet more search tools. Microsoft proposed Azure Knowledge Store and Artificial Intelligence. In 2019.

This was pre-ChatGPT. Pre-COVID. This was still 2019 when solving enterprise problems with AI required vision. For Jason's enterprise, that meant a big bet on the future.

"Most vendors showed us yet more search tools," Jason said. "Microsoft showed us how to leverage artificial intelligence."

Together, we designed:

• Document cracking pipelines to analyse complex file formats at scale
• Custom AI skills to detect brands, objects, context
• Multi-stage enrichment combining Vision, Text Analytics, Custom Models
• Cosmos graph database connecting 60M files into one intelligent system

But here's where it gets interesting. The UI? Microsoft Teams.

Not another dashboard to learn. Not another login to manage. We turned natural language into action. An employee could simply type: "Coca-Cola red truck sunset NYC"

Within seconds:

1. AI found the exact files among 60 million
2. Triggered cold storage restoration
3. Notified the team they were incoming

No training. No manual. Just... working.

Week 1: Discovery sessions in New York

The first question: "How do people actually search?" We mapped 20 years of content patterns. Not filenames - but how people remembered. "The ad that made people cry." "The Coca-Cola red truck." AI would need to understand context, not just metadata.

Week 2: Validation in London

Same patterns, different accents. Everyone searched by memory and context. This wasn't a New York problem. It was universal.

Week 3-4: Integration architecture

This is where most AI projects fail - bridging the gap between proof-of-concept and production. We built for how people actually worked. No new workflows. No behavior change required. Everything through Teams.

Week 5-8: Enterprise readiness

The technology was straightforward. Getting through security reviews, budget approvals, and stakeholder buy-in? That's where projects typically stall. Microsoft's involvement opened doors. Live demonstrations converted skeptics into advocates.

Week 9-12: Global deployment

No phases. No extended pilots. Full rollout across 11 offices from New York to London to Barcelona. Day one: everyone had access. The best adoption strategy? Build something people desperately need.

Creative directors stopped hoarding "just in case" folders. Why maintain duplicates when you can find anything in 30 seconds?

New employees discovered work from before they were hired. "I found content from years back that solved exactly the problem I was facing," one designer said. "It was like accessing the company's institutional memory."

Global teams became truly global. A brief in Sao Paulo could instantly access excellence from Tokyo. Not through complex systems or workflows. Just by asking: "Show me our best automotive work with urban settings."

"We didn't just solve search," Jason reflects. "We unlocked 20 years of creative intelligence. Suddenly, our history became our competitive advantage."

1. Your biggest constraint is your biggest opportunity
We couldn't afford $1.1M for hot storage. So we built a $160K AI system instead. Constraints force innovation. What "impossible" limitation might be hiding your breakthrough?

2. Adoption is architecture
The most sophisticated AI is worthless if people won't use it. We built inside Teams - where 5,000 employees already worked. Result: 100% adoption, zero training. Where does your workforce already operate?

3. Great AI is invisible AI
Success isn't when people discuss your AI capabilities. It's when they stop noticing the technology exists. When search becomes as natural as asking a colleague. When artificial intelligence becomes actual intelligence.

For Jason, it was creative assets.
For you, it might be:

• Product documentation with 47 conflicting versions
• Customer insights buried in shared drives
• Institutional knowledge leaving with every retirement

We solved this in 83 days. In 2019. Before AI became a LinkedIn buzzword.

Today? Most "AI consultancies" are wrapping ChatGPT and calling it transformation. They're building chatbots while your knowledge remains trapped in storage.

The question isn't whether you need intelligence.
It's whether you'll build it before your competitors do.