Seventy-six percent of executives in a recent MIT Sloan and Boston Consulting Group study describe agentic AI as more like a coworker than a tool. That framing shift has significant implications for how organizations evaluate AI platforms — and most enterprise evaluation processes haven't caught up to it.
The gap between adoption intent and structured evaluation is widening. BCG's research found that agentic AI has reached 35% enterprise adoption, with 44% of organizations planning deployment. Yet the same research identified a recurring structural barrier: technology and strategy leadership are evaluating AI platforms using frameworks built for a previous generation of software.
The result is a category error. But the more interesting question is which category they're reaching for — and why neither IT nor HR has the full picture alone.
Two Governance Frameworks. One New Problem.
When a new employee joins an organization, two functions activate simultaneously.
IT provisions access — account creation, permissions, system integrations, security posture. The question IT asks is: what does this person need access to, and how do we ensure they can only access what they're authorized to access?
HR establishes role definition — job title, scope of responsibility, reporting relationships, performance expectations. The question HR asks is: what is this person actually accountable for, and how does their work connect to organizational outcomes?
For human workers, this handoff is routine. For AI agents, it almost never happens. Organizations either treat the agent as pure software — IT provisions it, nobody defines its role — or they treat it as a productivity tool — someone picks it up, nobody governs it. The two frameworks that exist to govern organizational capability never intersect around the AI question.
That's the category error. Not just that organizations are applying the wrong instrument. It's that they're applying half the instruments available to them, and always the same half.
The MIT Sloan/BCG researchers noticed this from the outside. Organizations that treat AI agents as coworkers are, implicitly, reaching for the HR governance frame — defining roles, establishing accountability, setting scope. The research found that this group has meaningfully higher deployment success rates. When you govern something like a workforce member, you ask different questions before deployment, and you catch different failure modes before they compound.
The Structural Mismatch in Enterprise AI Evaluation
Consumer AI platforms — GrokBot, Muse, Dots, and the wave of products that have followed — are designed around the individual as the unit of value. The account is personal. The configuration is personal. The admin is the user themselves. This is not a product deficiency; it's a deliberate design decision optimized for the consumer use case.
Enterprise software operates from a different design assumption. The organization is the unit of value. Provisioning, governance, and accountability are centralized. The administrator's visibility spans users and systems, not a single account.
When enterprise IT evaluates a consumer AI product against enterprise criteria, the product fails — correctly — because it was never designed to pass. The evaluation process is applying the wrong instrument to the wrong category.
But even when IT's evaluation is right, it's incomplete. An IT evaluation tells you whether the agent can be governed like software. It doesn't tell you whether the agent has a defined role, a clear scope of authority, or any organizational accountability for what it does. A human employee who passes an IT access review can still be ungovernable if nobody defined what they're supposed to do. The same is true of an AI agent.
The Two Dimensions That Establish Category
Organizational tenancy. Does the platform provision a dedicated organizational environment, or do users operate within shared consumer infrastructure? Enterprise-grade platforms treat the organization as the tenant — one environment, one administrative layer, one contractual relationship. Consumer platforms treat the individual as the tenant. This distinction determines whether enterprise governance is structurally possible, regardless of any individual feature set.
Administrative scope. Can the organization's IT function provision users, configure permissions, audit activity, and revoke access — centrally, without touching individual accounts? In enterprise software, the administrator's authority is organizational. In consumer software, the user is their own administrator. No feature addition bridges this structural gap; it requires a different product architecture.
If either dimension is absent, the product is a consumer platform. The relevant organizational question is whether it belongs in the enterprise technology stack or the personal productivity stack. Conflating the two categories produces predictable failure: the formal rejection cycle where IT correctly declines a consumer product and the business leader concludes IT is the obstacle; or the informal adoption cycle where 40 individual subscriptions accumulate across three departments, invisible to the organization's governance function.
A Note on Products That Span Categories
Some platforms are actively building enterprise capabilities onto consumer foundations. GrokBot is an example: xAI has introduced organizational tier structures with emerging administrative controls. The relevant question for these platforms is not whether enterprise features exist, but whether the organization retains meaningful control over its AI infrastructure as the product evolves. For platforms tied to a single model provider: when the provider makes model decisions — deprecations, capability changes, pricing adjustments — does the organization have the ability to adapt independently? Or does the platform's evolution dictate the organization's options?
Organizations evaluating single-provider platforms should assess not just current feature parity, but the degree to which the platform's roadmap is a dependency they're accepting. For organizations with mature AI governance requirements, that dependency is a structural risk that feature lists don't resolve.
The Evaluation Dimensions That Actually Matter
For platforms that clear the initial category test, evaluation should proceed across five dimensions — three from the IT governance frame, two from the HR governance frame. The organizations that apply both are the ones that don't get surprised at scale.
Governance architecture (IT). How does the platform implement organizational controls at the workflow level — not just user access, but the scope and authority of individual agent actions? As agentic AI adoption scales, governance becomes less about who can access the system and more about what the system is authorized to do.
Attribution and audit (IT). Can the organization produce a record of what the system did, when, and on whose behalf? This addresses workflow-level accountability across automated sequences that may execute with minimal human intervention — not just user login logs.
Integration depth (IT). Does the platform integrate with the organization's existing systems of record, or does it require workflows to migrate into the platform itself? Deep integration is a design assumption, not a feature addition.
Role definition and scope (HR). Does the platform give the organization a structured way to define what each agent is accountable for — specific workflows, bounded domains, explicit scope of authority — before deployment? An agent without a defined role is the AI equivalent of an employee without a job description. The work happens. The accountability doesn't.
Model flexibility (IT + HR). Platforms built to accommodate multiple model sources — allowing organizations to route tasks to the appropriate model based on cost, capability, and risk profile — provide a structural advantage as the provider landscape evolves. If the capabilities of your AI agents are determined entirely by a vendor's model decisions, the organization doesn't actually control the composition of its AI workforce.
The Convergence Your Governance Framework Isn't Ready For
Most organizations have mature frameworks for governing software and mature frameworks for governing people. They have almost no framework for governing something that operates with the authority of software and the functional role of a person.
That gap is where most enterprise AI deployments fail — not during evaluation, but after deployment, when an agent is running workflows that nobody defined a scope for, producing outputs that nobody has accountability for, and doing it in ways that IT can audit but HR never thought to govern.
The organizations building durable AI capability right now are treating this as a convergence problem. They're asking IT to provision AI agents the same way they provision any enterprise software. And they're asking HR — or someone performing that function — to define what those agents are actually supposed to do, what they're not authorized to do, and who in the organization is accountable for their outcomes. That's not a technology question. It's an organizational design question.
What This Means for Technology and Strategy Leaders
The rigor begins with categorization — consumer platform versus enterprise platform — and it extends through two governance lenses, not one. The IT question is whether the agent can be governed. The HR question is whether it has been defined.
Organizations that answer both questions before deployment will spend significantly less time managing the failure modes that are currently consuming enterprise AI evaluation bandwidth across the market. The tools that make both questions easy to answer are the ones worth building on.
Dakota is the Marketing Strategy & HubSpot Specialist at W(AI)kforce, an agentic AI platform built for enterprise workflows at the intersection of IT governance and organizational design.