The Paradox Hiding in Plain Sight
Mass General Brigham operates 90,000 employees and is the largest employer in Massachusetts.
Last year, their leadership made a deliberate strategic decision: don't build your own AI infrastructure.
Let the vendors handle it. Microsoft, Epic, Workday, ServiceNow — these companies have resources and are making AI a top priority.
"Why are we building it ourselves?" their CTO asked publicly.
Six months later, MGB had built a custom platform layer around Microsoft Copilot.
Because Copilot couldn't handle protected health information (PHI) the way a 90,000-person hospital system requires.
The new platform supports up to 30,000 users and exists entirely to work around the gaps their primary AI vendor left.
Then MGB discovered that every vendor — Microsoft, Epic, Workday, ServiceNow — is now building its own AI agents.
All operating differently. All with different data models. All with different APIs.
Now MGB is investing in building "a control plane that coordinates and orchestrates all of these agents."
The company that decided not to build AI infrastructure is building AI infrastructure.
The 72% Problem
VentureBeat's Q1 2026 research surveyed 40 enterprise organizations.
Their finding: 72% of enterprises claim to have adequate AI governance — two or more primary AI platforms with clear accountability, guardrails, and security oversight.
Most don't.
What they have is sprawl: a collection of overlapping SaaS platforms, disconnected proof of concepts, and policies that were named governance after the fact.
The MGB CTO described the vendor landscape with a useful analogy.
"When you ask six blind men to touch an elephant and say, what does this elephant look like? You're going to get six different answers."
Each vendor is describing the same enterprise AI transformation.
They're each building a different part of the solution.
And the enterprise is left assembling the elephant.
Why This Happens
The root cause isn't vendor failure. It's architectural.
Most enterprises adopted AI by selecting the best tool for each use case.
Microsoft Copilot for productivity. Salesforce Agentforce for CRM. Workday AI for HR. AWS Bedrock for infrastructure.
Each vendor deployed well. Each tool delivered value in its domain.
But these systems don't share a common data model. They don't enforce consistent access controls. They don't have a unified audit trail.
When agents start taking actions — not just answering questions — this fragmentation becomes a security and compliance liability.
VentureCrowd, a startup fundraising platform, experienced the exact same pattern with AI coding agents.
They cut front-end development cycles by 90% with agentic workflows.
Their CPO, Diego Mogollon, identified the failure mode clearly: "The challenges are rarely about the coding agents themselves. It's a context problem disguised as an AI problem — and it is the number one failure mode I see across agentic implementations."
Agents reason against whatever data they can access at runtime.
When that data is fragmented, inconsistent, or uncontrolled, agents become confidently wrong at scale.
What the Architecture Needs to Look Like
The organizations getting this right aren't using fewer AI tools.
They're building a governance layer that sits above the tools.
The requirements are consistent across sectors:
A unified data layer. Every agent needs to query institutional data through a common interface — not directly hitting disparate systems. The Model Context Protocol (MCP) has emerged as the open standard for this: a structured way to give agents governed access to SIS, LMS, CRM, ERP, HRIS, and any other system, without custom integrations for each.
Role-based access control at the agent level. The same permissions logic that governs human users must apply to agents. An agent serving a student shouldn't access faculty data. An agent serving one department shouldn't query another's records. Deny-by-default, with explicit grants.
An audit trail that agents can't bypass. Every tool call, every data access, every action taken — logged with the requester identity, parameters, timestamp, and outcome. This is a regulatory requirement in healthcare (HIPAA), education (FERPA), and government (NIST 800-53). It's also how you diagnose agent failures before they become incidents.
LLM agnosticism. Locking your agent infrastructure to a single model vendor means your governance architecture inherits that vendor's roadmap. The organizations with sustainable AI programs can swap models — from GPT to Claude to Gemini to open-weight — without re-engineering the governance layer.
The Pattern Across Sectors
Adobe launched CX Enterprise at Adobe Summit this week.
An end-to-end agentic AI system for enterprise marketing, built with partnerships spanning AWS, Microsoft, NVIDIA, Anthropic, OpenAI, Google Cloud, and IBM.
The "AI kills Adobe" bear thesis answered: we're the control plane, not the target.
This is the move every mature enterprise software vendor is making.
The question isn't whether agents will run your organization's workflows.
The question is whether you own the infrastructure they run on.
MGB learned this the hard way — building around Copilot, then building around their vendors, then building the orchestration layer they should have scoped from the start.
The organizations that define the control plane first — unified data access, governed agent permissions, complete audit trails, LLM flexibility — will spend less time retrofitting governance onto agent systems that already have momentum.
What ibl.ai Builds For Enterprise
The ibl.ai platform ships as a complete AI operating system: full source code, perpetual license, deployed on your infrastructure.
The architecture is built around the exact requirements MGB and organizations like them are discovering.
An MCP-based interoperability layer connects Workday, Salesforce, SharePoint, and any other system — giving agents governed access without fragile point-to-point integrations.
Role-based access control operates at the agent level: each agent sees exactly the data its role permits, nothing more.
160+ pre-built agent templates cover the functions enterprises need most: IT help desk, HR onboarding, compliance training, sales enablement, knowledge management, operations support.
At 1,000 users, the total cost is approximately $31,000 per year — compared to $300,000 for ChatGPT Team or $360,000 for Microsoft Copilot at the same scale.
The difference: you own the code, own the data, and choose any LLM.
The governance conversation is happening in every enterprise boardroom this year.
The organizations that treat it as an architecture decision — rather than a procurement decision — are the ones building something durable.
Related: AI Budgets Are Growing 40% a Year. Deployment Isn't. · Fortune 500 AI Agents and the Data Sovereignty Question