Engineering, scaling enterprise AI to 8,000 employees, of a global consumer brand.

Microsoft Azure · OpenAI · Databricks · Model Context Protocol

Engagement at a glance

8,000+ employees served by a single conversational AI surface
5,000+ monthly active users within weeks of launch
87% employee satisfaction score post-deployment
3,000 users on MCP-connected workflows beyond simple Q&A

The business context

For a global consumer goods enterprise of 8,000 employees, the cost of knowledge friction is not an IT problem; it is a productivity tax that compounds every quarter. Sales operations checking SAP, support teams searching Confluence, project leads filtering Jira boards, business analysts pulling BI reports. Every minute spent context-switching across disconnected enterprise systems is a minute not spent on the customer, the channel, or the next launch.

The leadership team did not want a smarter search box. They wanted a single conversational surface where every employee could ask the company a question and receive a defensible answer in their own language within seconds. Generative AI was the means. Workforce productivity at enterprise scale was the brief.

The challenge

The brief carried four constraints that ruled out most off-the-shelf approaches. Time-to-value: the platform had to prove organization-wide value fast, not after a twelve-month roadmap. Architecture: the assistant had to be designed from day one as an extensible AI platform, not a standalone chatbot retrofitted later. Scale: the launch had to reach the entire workforce across geographies without breaking under the load.

Governance: enterprise data, intellectual property, and regulated employee information had to remain inside the company's Azure tenant under existing identity and access controls. A standalone AI assistant would have delivered Q&A. A connected enterprise platform would deliver workflows. The leadership team chose the connected platform.

The Sundew approach

Sundew joined a multi-vendor delivery team and contributed to the design and engineering of an enterprise AI platform built on the Open WebUI framework, hosted on Microsoft Azure, and powered by OpenAI's large language models. Databricks was selected as the vector store, enabling semantically rich retrieval-augmented generation (RAG) across large volumes of structured and unstructured enterprise content.

The architectural choice that defined the engagement was adopting the Model Context Protocol (MCP) as the integration layer, allowing the assistant to progressively connect to the enterprise systems that mattered most without re-engineering the core platform for each new system.

The launch deployment gave every employee a single conversational interface to query internal knowledge, with no technical expertise required. The connected platform then extended the assistant into four enterprise systems in sequence:

Intranet and Confluence, surfacing policies, procedures, and documentation through natural language
Jira, bringing project tracking and ticket context into the assistant workflow
BI database, allowing business users to query metrics and reports conversationally
SAP, extending the assistant into core business processes and operational data

The MCP-based architecture meant each new system was added as a connector, not as a re-platforming exercise. This is the architectural decision that separates AI pilots that stall at proof of concept from AI in production that scales to the workforce.

The outcome

The first release achieved organization-wide traction within weeks of launch, with over 5,000 employees using the platform every month. As MCP-powered integrations rolled out, 3,000 users moved into the connected platform, working not only with Q&A but with tool-assisted workflows where the assistant retrieved, summarised, and acted across multiple systems in a single conversation.

An employee satisfaction score of 87% confirmed the platform was delivering measurable day-to-day value at scale, not pilot enthusiasm. What began as a standalone chatbot is now the foundation of a growing enterprise AI ecosystem, with the assistant continuing to expand its reach across the organization's most critical systems.

What this means for your enterprise

If your enterprise is evaluating a generative AI assistant strategy and the conversation has been stuck between pilot something small and wait for the platform to mature, this engagement is the counter-pattern.

With the right architectural choice at the start, vector store inside the enterprise tenant, MCP as the integration layer, governance baseline established from day one, an enterprise AI platform can ship to thousands of users in weeks, then scale into connected workflows in months. The risk is not building too early. The risk is building without an architecture that compounds.

"AI in production is not a model choice. It is an architectural choice."

Talk to our Experts

Technologies and capabilities

Ready to start?

See how Sundew's Managed Engineering Services fit your industry roadmap.

Schedule a Call

Frequently Asked Questions

How long does it take to deploy an enterprise AI assistant for thousands of employees?

Deployment speed depends heavily on architectural decisions made at the outset. In this engagement, Sundew and the delivery team launched a conversational AI platform to the full 8,000-person workforce and reached 5,000+ monthly active users within weeks of launch, not months. The key enabler was building on an extensible platform from day one (Open WebUI on Azure) that could be deployed immediately without the overhead of custom front-end development.
How does the AI platform integrate with systems like SAP, Jira, and Confluence?

The platform was built around the Model Context Protocol (MCP) as the integration layer. MCP allows each enterprise system, Confluence, Jira, a BI database, and SAP, to be added as a discrete connector without rebuilding the core platform. This allows employees to retrieve information, access workflows, and interact with multiple systems through natural language without leaving the AI interface.
How is sensitive business and employee data protected when employees use the platform?

All enterprise data, intellectual property, and regulated employee information remains inside the client's Microsoft Azure tenant at all times. The platform uses the organization's existing identity and access controls; no new permissions infrastructure is required. The Databricks vector store used for retrieval-augmented generation (RAG) is hosted within the same tenant boundary. Employees only surface content they are already authorized to access. This governance baseline was a prerequisite of the design, not an afterthought added post-deployment.
What is the difference between a standalone AI chatbot and the enterprise AI platform Sundew delivered?

A chatbot primarily answers questions based on a limited knowledge source. An enterprise AI platform uses MCP-connected workflows; the assistant can retrieve live data from SAP, surface a Jira ticket, query a BI report, and summarize a Confluence policy, all within a single conversation, without the employee switching context. This makes the platform more valuable for enterprise-wide adoption and long-term scalability.
What role did Sundew play in this multi-vendor delivery?

Sundew owned the design and engineering of the platform layer; the extensible architecture, MCP integration, and RAG pipeline that all other vendors and integrations built on top of. This engagement highlights Sundew's ability to operate within complex enterprise delivery programs, drive architectural decisions with long-term business impact, and collaborate effectively across multiple vendors without serving as the primary program owner.
Can the platform expand to additional systems after the initial launch?

Yes, and that is the point of the architecture. The MCP integration layer is designed to grow. What launched as a knowledge assistant is now an expanding enterprise AI ecosystem, with each new system connected as a discrete connector. The platform compounds its capabilities without requiring a new program budget for each addition.