Focused on Delivery
We built central governance for Claude Code and Claude Desktop, a single MCP gateway in front of corporate systems, and shared evaluation infrastructure that runs inside existing build pipelines.
Focused on Results
Teams across the organization can build agents on one security baseline, reach corporate systems through one controlled path, and hold model and prompt changes until evaluations pass meaningful business thresholds.
Focused on Partnership
Focused engineers embedded with the GenAI platform team, and paired with product engineers who were writing their first evaluations.
Stifel is a regulated wealth-management firm with thousands of financial advisors. The company handles financial planning, institutional equity, fixed income sales, and investment banking, among other services. In its efforts to roll out AI across the organization, Stifel was looking for a partner to implement AI and AI development strategically and securely.
Focused embedded with Stifel's GenAI team to build a foundation that allowed teams to build agents without losing visibility, quality or control.
The Challenge
Stifel wanted to make AI accessible to its many teams, but, in a highly regulated industry like finance, there were concerns with letting teams build any agents without the right visibility, quality, and control measures in place.
As an example, Developers wanted Claude Code connected to multiple internal systems. Without a shared path, each team would stand up its own credentials, MCP servers, and review process. Stifel's platform team is small. Reviewing requests from each team would have consumed their workload.
Stifel's teams were building agents faster than any shared testing practice existed to support them. Evaluation was new ground for most of them, and each team working it out alone would have meant inconsistent quality and no common way to judge whether a change was safe to ship.
The Solution
Focused embedded with Stifel’s GenAI team to build shared infrastructure so that the second, tenth and fortieth teams could inherit the right controls for building agents. A key part of the solution was to grant everyone Claude, but governed centrally. Claude Code and Claude Desktop were rolled out with firm controls that included blocking dangerous commands, making available only approved connectors, and ensuring that every session emitted OpenTelemetry metrics into Prometheus and Grafana. Stifel can see how Claude is being used across teams.
Additionally, the team built a gateway between Claude and Stifel’s systems. When Claude reaches one of those systems it acts as the signed-in person, carrying that person's permissions, which keeps actions attributable and removes the need for teams to hold their own service credentials. Teams add tools by declaring them in code. Automated rules check the declaration, so the platform team is no longer in the loop on each individual tool.
Focused also worked with Stifel to build shared evaluation infrastructure for internal reuse. A shared evaluation library and Azure DevOps pipeline templates allow teams to run evaluations in CI without designing pipelines first. Most of the work was in the seams. LangSmith datasets do not come with conventions for how cases are organized, versioned, or shared between teams, so those had to be defined before the library was useful to multiple teams. Connecting LangSmith to Stifel's build and identity systems required a meaningful amount of integration code, because the two were not designed to work together.
The enablement work ran alongside the build. Focused engineers paired with product teams on their own agents, writing the first test cases and scorers with them and working through failures together, rather than handing over a library and documentation.
Compliance controls ship in the same shared code. Personal data is scrubbed before traces are recorded. Prompts that trip a guardrail are removed from conversation history instead of stored. Content guardrails are applied by policy across the organization's AWS accounts, so individual teams do not configure them.
Delivery and Adoption
Governed Claude Code and Claude Desktop have been rolling out since spring 2026 and are in use by developers and non-developers, with the population still expanding. All Claude traffic reaches corporate systems through the gateway. Developer onboarding includes triggering a blocked command and confirming that events appear in monitoring.
The gateway went live in August 2026 with the Azure DevOps, Outlook, and Atlassian connectors, serving Claude Code, Claude Desktop, and claude.ai. Self-service tool onboarding followed later that month, which is what moved new tool requests off the platform team's queue. Dashboards built from Claude's telemetry are live for adoption, approval friction, context health, skill usage, and gateway performance.
Results
Claude is available to developers and other roles under a single security baseline, with usage and cost attributable per team. An internal study of 460 matched work items found that teams working through the governed toolchain had 45% shorter lead time from request to delivery.
Routing traffic through one gateway made failure patterns visible in aggregate. Analysis of a month of gateway traffic showed that most failures came from a single cause: the model guessing repository names. A fix at the gateway removed that class of failure for every client at once, which would not have been possible if each team held its own connectors.
Evaluation now gates releases. Teams tune against their evaluation banks and see response time, inference cost, and answer correctness move together. In one case the evaluations stopped a model upgrade that scored better on correctness but made worst-case response time much worse and changed which tools got selected. The trade-off was visible before anything shipped.
As new agent teams spin up within the organization, they inherit the guardrails and per-user credentials that the shared libraries offer.
Your systems are ready. Your agent isn't. Let's fix that.
We're the experts in agents and integrations. We'll tell you if we can help.