The Challenge
Every consulting firm carries its process in three places at once: a handbook nobody reads end to end, a stack of spreadsheets, and the heads of the people who have been there longest. General-purpose AI assistants understand software delivery in the abstract, but they know nothing about how a specific firm runs a kickoff, structures a client folder, or decides when a project is ready for user acceptance testing. That gap makes AI useful for drafting and useless for doing.
There was a second problem underneath the first. AI adoption was arriving bottom up, one person and one subscription at a time, which meant useful techniques stayed trapped in individual notebooks and nothing compounded.
BRIO needed to:
- Settle on one approved platform so capability could accumulate instead of scattering.
- Encode its delivery process in a form AI could execute, not just summarize.
- Give AI live, permissioned access to the systems where work actually lives.
- Assign clear ownership so every automated capability had a human accountable for it.
The Approach
The firm made two moves in sequence, and the order mattered. First it standardized on Claude as the only approved AI tool, with access governed through corporate identity, which is covered in detail in a companion case study. That decision turned AI from a personal preference into shared infrastructure, and only then was it worth building anything durable on top.
The thing built on top was a suite of role-based agents that mirror the delivery organization itself: a Project Manager agent, a Business Analyst agent, a Salesforce Consultant agent, a Data Developer agent, a File Migration agent, an Azure Administrator agent, and an Executive Sponsor agent. Each is scoped to a role a person already holds, which keeps the boundaries legible and makes it obvious who reviews the output.
The Solution
Role-Based Agent Design: Each agent is defined by its purpose, how it processes inputs, and which other agents it is permitted to talk to. Communication paths live in a separate configuration file, so orchestration is explicit and auditable rather than emergent.
Skills as Executable Process: Agents are defined in plain-language files rather than code. A skill is a Markdown document carrying instructions, metadata, and the phrases that trigger it; the working parts underneath are scripts that call real APIs. Because the definitions are readable, the people who own a process can change the process. BRIO’s standard project task workbook now names the agent or skill that handles each step. Client folder setup, for example, is invoked in conversation, prompts for the project type and client name, and builds the full standard structure automatically.
Shared Capability Instead of Private Prompts: Routines that proved themselves individually get promoted to organization-level skills, installed once and available to everyone without manual setup. A timesheet audit routine that began as a file passed between colleagues became a firm-wide capability that simply appears for the whole team. Recurring operational work including weekly time allocation, client status reporting, staff meeting agendas, and capacity gap analysis now runs this way.
Governance and Ownership: Every agent has a named owner. Audit findings and ownership assignments were staged as structured tickets in a dedicated project with per-agent epics and a dashboard that breaks the backlog down by agent, so the program is managed like a product rather than a side experiment.
Enablement at Scale: The work was carried into the company quarterly all-hands with live demonstrations, a shared vocabulary section, and a hands-on exercise, so the first exposure for most of the firm was doing rather than watching.
Technical foundation
- Markdown skill definitions with structured metadata and natural-language triggers
- Python and Node.js execution layer
- Connector layer reaching the firm’s systems of record
- Organization-level skill distribution, centrally installed
- Dedicated project tracker with per-agent epics and a breakdown dashboard
The Impact
Encoding the process rather than describing it changed what AI could be trusted with:
- Standard project setup steps now run from a single instruction.
- Recurring operational routines moved from manual assembly to agent execution.
- Every agent action is attributable, logged under its own identity, and reviewable.
- Process changes are made by the people who own the process, in plain language.
The result is less a set of tools than an operating layer. BRIO’s delivery methodology is now something the firm can run, not just read.
