We recently published a full case study on lululemon's AI strategy - the three pillars, the use case prioritization matrix, the phased investment roadmap. One finding keeps coming up in client conversations, because it is quietly one of the most practical things in the entire study: lululemon tested putting its own knowledge base online and making it available to in-store employees on iPads, powered by AI. Associates use it on the job to get answers in real time. It helps with onboarding, training, and day-to-day operations.

That sounds simple. It is not just a search box with a nicer front end. Here is what makes it different, why it matters for any company with living documentation, and what we are seeing clients do with it.

Search Returns Documents. This Returns Answers.

A traditional knowledge base search works like a library catalog. You type a question, you get a list of documents, you open the right one, and you read. The burden is on the employee: figure out which document is relevant, then extract the answer from a long SOP.

The architecture that changes this is RAG (retrieval-augmented generation). As IBM describes it, RAG is an architecture that connects an AI model to external knowledge bases, and AWS adds that the model then references that authoritative knowledge base rather than relying on its training data. The AI searches your documents behind the scenes, pulls the relevant sections, and generates a direct answer to the question that was asked. The employee does not read a 40-page manual to find the one paragraph they need. They get the answer, with the source right there to verify it.

And it is interactive. Because it is a chat interface, an employee can ask a follow-up: "What if the customer wants a return past 30 days?" "Which approval do I need for that?" The system holds the context of the conversation. That is the difference between a document search and a conversation.

The core distinction: a search function tells you where the answer lives. A RAG knowledge base gives you the answer itself, grounded in your own documents, and lets you keep asking until you actually understand it.

What the Pilot Targeted

The program is real and publicly documented. In 2024, a University of Washington iSchool capstone team built lululemon's AI tutor - a web-based internal platform for onboarding and training newly hired store associates, built on GPT-4, using Socratic-style questioning to gauge product familiarity and close knowledge gaps. The project brief notes the scale of the problem it attacked: lululemon employs thousands of retail associates across more than 700 stores, and new employees historically spent significant time learning what to recommend to shoppers.

The mechanics matter less than the outcomes it targets:

The key detail: lululemon did not build a separate training system. The knowledge base is the same one the company already maintains. They put AI in front of what they already had. The timing is no accident either - lululemon appointed its first Chief AI & Technology Officer in August 2025, after a stretch where Americas comparable sales declined 4% in Q2 FY2025. Operational efficiency is no longer a nice-to-have; it is the plan.

A Pattern, Not a Pilot

lululemon is the cleanest example we have studied, but it is far from alone. In the last two years, some of the largest companies in the world have put AI in front of their internal knowledge - for the same three reasons: faster onboarding, better training, and answers at the moment of need.

Two of these deserve a closer look, because they show what the pattern looks like at very different scales.

Morgan Stanley: answers over 100,000 documents

Morgan Stanley built AI @ Morgan Stanley Assistant to give financial advisors access to the firm's intellectual capital - a database of about 100,000 research reports and documents. Per OpenAI's case study, over 98% of advisor teams actively use it for internal information retrieval. The firm followed with AskResearchGPT, which can search more than 70,000 proprietary reports published annually. It is the same architecture as the lululemon tutor: retrieve from the firm's own corpus, generate a grounded answer, keep the source traceable.

Walmart: a knowledge base at retail scale

Walmart's associate-facing conversational AI has been running for years - more than 900,000 weekly users and over 3 million queries per day. In June 2025 the company announced a GenAI upgrade that turns lengthy process guides into clear, step-by-step instructions - answering questions like "How can I process a return without a receipt?" - alongside AI task management that cut shift planning from 90 minutes to 30. That is a knowledge base doing real operational work, at a scale most companies cannot even imagine.

And the construction signals are already here. Constructora Las Galias, a Colombia-based builder, runs a centralized internal knowledge platform on Gemini Enterprise for sales, legal, and company-wide information access, and HCSS now lets firms upload internal guides into a company knowledge base that its Copilot indexes and retrieves. The tooling exists. Most firms just have not adopted it yet - which is exactly the opening we are seeing in client conversations.

This Is Not Just a Retail Story

I sat down with a commercial construction firm this week that builds large retail and restaurant developments across New England and Upstate New York - an industry-leading, full-service contractor with more than four decades in business, 20 million+ square feet built, and a client list of national retailers and developers. They have a veteran team, and they run on documentation: SOPs for creating bids, reviewing estimates, managing change orders, contracting with subcontractors, and the rules and methodologies that define how the company works. Those documents are living - updated regularly, with new ones created to standardize practices.

They loved the lululemon example. Not because they want to be like an athleisure company, but because the pattern maps directly: a growing body of procedural knowledge, a workforce that needs the right answer at the right moment, and documents that are always being revised. Their version of the use case is an estimator asking "What is the current change order approval threshold?" from a job site, or a new project engineer getting a correct, sourced answer about subcontractor documentation instead of interrupting a superintendent to ask.

Any company with SOPs - construction, healthcare, manufacturing, professional services, logistics - has this problem. The documents exist. The knowledge base exists. What is missing is a way to get answers out of them without making people read.

Three Features That Make It Grow

What made the concept click with this client was not the RAG piece alone. It was the three features working together:

1. An interactive knowledge base that answers, not just indexes

The foundation is a chat interface over your existing documentation, with answers grounded in your own SOPs and cited back to the source. Every answer carries the document reference, so a seasoned employee can sanity-check it, and a new hire can trust it. Because the underlying documents keep changing, the system always reflects the current version - no more "I think that changed last quarter."

2. A place where employees contribute what they know

This is where it stops being a document library and becomes a knowledge base. Employees add tips, tricks, and ideas that have worked for them on the job, and those contributions are logged and searchable for everyone. A project manager logs a prompt that saves them an hour in their estimating tool. A field superintendent posts the checklist they use for subcontractor walkthroughs. A veteran shares the shortcut for a software package the whole office uses.

For a company with a veteran team, this is quiet gold. The knowledge that currently lives in people's heads - the stuff that walks out the door at retirement - starts accumulating in a form the whole organization can use. Construction is a stark example of the risk: about one in five construction workers is 55 or older, and the industry's experienced-worker exodus is a documented concern, with losing decades of institutional knowledge among the top risks. The same logic applies to any field where expertise is concentrated in long-tenured people.

3. Visibility into what people actually look up

The third feature is the one nobody expects to be valuable, and it is often the most valuable of all: logging what people search for and whether they found the answer. That data is a management instrument, not an IT artifact.

The Compound Effect

These three features reinforce each other. The RAG layer makes existing documentation useful immediately, which builds adoption. Employee contributions keep the system current and make it feel like the team's own, which builds more usage. And the analytics loop shows you exactly where the gaps are, so the knowledge base gets better in the places people actually work - not the places someone assumed they work.

None of this requires a massive AI program. It starts with documentation you already have, an AI layer that can answer questions against it, a contribution workflow, and a dashboard. The hard part is not the technology. It is the discipline to keep the documents current and the willingness to act on what the usage data tells you.

FutureInSites helps companies design and build AI knowledge bases over their existing documentation - RAG architecture, employee contribution workflows, and the analytics to know what your people actually need. We built the lululemon case study in the case studies library, and we have seen the same pattern land for commercial construction firms, professional services, and manufacturers. Get in touch if you want to put AI in front of the knowledge you already have.

References

  1. University of Washington Information School. "lululemon AI Tutor: Reducing Onboarding Time for Newly-Hired Store Associates." Capstone project, 2024.
  2. IBM. "What is RAG (Retrieval Augmented Generation)?" IBM Think.
  3. Amazon Web Services. "What is RAG? - Retrieval-Augmented Generation AI Explained." AWS.
  4. lululemon athletica inc. "lululemon Appoints Ranju Das as the Company's First Chief AI & Technology Officer." Press release, August 26, 2025.
  5. CNBC. "Lululemon (LULU) Q2 2025 earnings." September 4, 2025.
  6. Construction Dive. "Construction's age problem: A foreboding exodus of experience." May 25, 2023.
  7. Skillit. "Challenges Facing an Aging Construction Workforce."
  8. Gitnux. "150+ Onboarding Statistics: 2026 Verified Report." (SHRM data on structured onboarding and time to productivity.)
  9. CNBC. "Morgan Stanley kicks off generative AI era on Wall Street with assistant for financial advisors." September 18, 2023.
  10. OpenAI. "Morgan Stanley uses AI evals to shape the future of financial services." OpenAI case study.
  11. Morgan Stanley. "Morgan Stanley Research Announces AskResearchGPT." Press release, October 23, 2024.
  12. Walmart. "Walmart Unveils New AI-Powered Tools To Empower 1.5 Million Associates." Press release, June 24, 2025.
  13. The Home Depot. "The Home Depot Introduces Magic Apron, a Suite of Advanced AI Tools." Press release, March 6, 2025.
  14. Google Cloud. "How Gemini Enterprise is helping SMBs jumpstart their AI transformations." Google Cloud Blog, April 22, 2026.
  15. HCSS. "Construction AI Tool | HCSS Copilot." HCSS product page.

← Back to Blog