Navari: Enterprise Agentic AI
From a pilot group to 1,000+ users while I ran it, and about 10,000 after I handed it off
An agentic platform for R&D scientists: describe the work and the agents do it, through MCP servers wired to literature, trial registries, regulatory data and enterprise data. It launched in spring 2025, grew by word of mouth past 1,000 users, and is reported at about 10,000 users as of late 2026, after the handoff.
The problem
The Answer Problem
Enterprise AI tools were glorified search engines. A scientist would ask a question, get an answer, then spend the rest of the week turning it into an analysis, a document or a piece of code by hand. The insight was cheap. The work was still the work.
- Answers with no deliverable attached: the scientist still had to build the thing
- No path from a question to the internal data that could answer it
- Every query started from zero, with no tools, no memory and no sources to show
- Governance built for chatbots, which meant no governed route for an agent that acts
What I built
Agents That Finish the Job
Navari completes tasks. Through the Model Context Protocol it connects a reasoning model to enterprise tools and data, runs multi-step workflows, and hands back a finished deliverable with its sources. Model-agnostic by design: Claude first, with swappable backends.
- MCP-based tool execution with 10+ deployed servers: literature, trial registries, regulatory data, document and slide generation, data analysis, enterprise data
- Governed, traceable outputs so an R&D decision can show where it came from
- Complete deliverables: applications, analyses, reports, automations
- Sessions isolated per user, enterprise auth, audit logging
How it fits together
Architecture
A layered architecture separates reasoning from execution, so a new tool or data source ships as an MCP server without touching the core.
Timeline
How it went
- Early 2025Framework built in about six weeks; first agents and workflows shared internally
- Spring 2025Launch to a pilot group
- Sep 2025Hundreds of users, MCP servers past ten, adoption spreading function to function
- Dec 2025Production approval; 600+ users
- Mar 20261,000+ users
- Jun 2026Handed off
- Late 2026About 10,000 users across the organization, as reported to me after the handoff
What it taught me
Lessons
- Scientists do not want answers to implement. They want the thing done.
- MCP kept the tools separate from the reasoning, so adding a data source never put the core at risk.
- Word of mouth beat every rollout plan. A scientist who saved a week is the marketing.
- The platform scaled faster than the organization around it. By 2026 the bottleneck was governance, support and change management, never the code.
- It kept growing after I left. That is the test of a platform.
Stack
Connected work