From Generic AI to Deeply Personal Agents: How Enterprises Are Racing to Own User Context

The first wave of enterprise AI centered on generic copilots and basic recommendations. The next wave is about something more ambitious — and riskier: systems that build a rich, persistent understanding of individual users and use that context to act on their behalf. For IT leaders and product owners, this shift from generic AI to deeply personal agents is reshaping how tools are designed, governed, and measured.

From generic copilots to deeply personalized agents

Traditional personalization in software has largely been about pattern-matching: recommender systems correlating user behavior to suggest content or next steps. Emerging AI agents, powered by large language models (LLMs), are moving beyond that. They can analyze the user directly — their stated preferences, role, vocabulary, and constraints — and adapt behavior in near real time.

That is increasingly what users expect. They no longer want systems that infer who they are from clicks alone. As Zoom AI product head Lijuan Qin put it in a recent Beyond the Pilot podcast discussion: the goal is not to randomize or guess, but to let the user explicitly say, “This is what I care about,” and have the system respond accordingly.

For enterprises, this represents a strategic shift. Generic copilots can be rolled out uniformly across an organization, but their impact is limited. Deeply personalized agents promise more meaningful productivity gains — at the cost of new complexity in data access, security posture, and governance.

Zoom AI Companion: a case study in controlled personalization

Zoom’s AI Companion illustrates how major enterprise platforms are operationalizing this trend. The assistant began with familiar features such as meeting summaries, smart recordings, and action item extraction. It now extends personalization into how it interprets and acts on meeting content.

Users can tailor meeting summaries to their specific interests — for example, emphasizing technical risks for an engineer or commercial implications for an account executive. They can also define templates for follow-up emails to different personas. After a call, the assistant can automatically generate these materials, targeted to the intended audience, based on the same underlying transcript.

Zoom exposes deeper configuration in its AI Studio. A custom dictionary lets enterprises inject their own terminology and internal vocabulary so that outputs align with domain language — a critical requirement in industries rich with acronyms and proprietary product names. A “deep research” mode can synthesize what the organization already knows (“internal expertise”) with information from external sources to produce more comprehensive analyses.

Underpinning this is a deliberate emphasis on control. Qin stressed that humans can be “very specific” in defining what agents are allowed to do. Administrators and users can configure whether the agent may automatically send emails to particular recipients or whether a human verification step is required, especially when transcripts include sensitive information. Features can be enabled or disabled, and data access can be tightly scoped.

The design assumption, Qin emphasized, is that “we do not assume AI is smart enough to get everything right.” The system is built around the expectation that agents will sometimes be wrong or off-target — and that users need visibility into behavior and the ability to intervene. For enterprise buyers, this is a preview of what “personalization with guardrails” looks like in practice.

The land grab for user context

Zoom is not alone in this pursuit. As Sam Witteveen, co-founder of Red Dragon AI and host of Beyond the Pilot, observed on the same podcast, the current phase of AI feels like a “land grab for context.” The platforms that best understand users’ tools, tasks, and preferences will be positioned to deliver the most effective agents.

According to Witteveen, “knowing your users is the big thing.” That includes which applications they live in and what day-to-day tasks they repeatedly perform. The more structured context a system can maintain about these patterns, the better its memory and the more tailored its responses can be.

He points to tools like Claude Cowork and OpenClaw as examples of applications leaning into this model. These systems are designed to accumulate context about the user and then act on high-level instructions such as: “You know a bunch of things about me. You’ve got all this context. Go and generate the skills that are going to help me do a better job.”

In practice, that means allowing users to specify both preferences and routines. With an agent like OpenClaw, Witteveen notes, a user can instruct it conversationally — for instance, “At 4 o’clock I want you to do this” — and expect it to remember, schedule, and execute that behavior in line with what it already knows about the user’s work.

For enterprises, this type of behavior raises the stakes on who “owns” and manages user context. The more powerful the agent, the more important it becomes to define where that context resides, how it is shared across tools, and what constraints apply when it is used to automate actions rather than just generate content.

Security and governance risks as agents gain autonomy

The same capabilities that make deeply personal agents attractive also make them a potential security and governance risk. OpenClaw is a case in point. Witteveen notes that the tool has been plagued by security issues since launch, to the extent that some enterprises have banned or uninstalled it.

Those decisions have not been trivial. Because an autonomous agent can be tightly integrated into workflows and infrastructure, removing it must be done carefully. Witteveen warns that incorrect uninstall procedures could inadvertently delete portions of an enterprise’s broader stack, underscoring how entangled such agents can become in production environments.

For security and IT leaders, this highlights several practical concerns:

• Agents that operate across multiple systems can bypass traditional controls (such as endpoint or identity tools) if not correctly configured.
• Autonomous behavior based on personalized context amplifies the impact of misconfigurations or vulnerabilities.
• Unclear separation between the agent’s “memory” and core infrastructure increases the risk of unintended changes during removal or updates.

Zoom’s approach — emphasizing explicit permissioning, feature toggles, and data access controls — illustrates one possible response: treat autonomy as something to be progressively granted and continuously monitored, not enabled by default. The broader lesson is that personalization and autonomy must be paired with rigorous operational controls if they are to be viable at enterprise scale.

The cost of personalization: tokens, metrics, and trade-offs

Beyond security, there is an economic dimension to deep personalization. Rich context requires more data to be fed into LLMs, which in turn consumes more tokens — and raises costs. Every additional slice of user history, enterprise terminology, and external knowledge adds to the token budget for a given interaction.

Witteveen cautions that organizations need to think carefully about what they track and optimize for. The relevant metrics will vary from product to product, but measuring personalization purely by user delight, without accounting for underlying token usage and infrastructure load, is unlikely to be sustainable.

For AI strategists, this creates a new optimization problem: how to balance depth of context with cost and latency. Features such as Zoom’s deep research mode — which taps both internal expertise and external insights — can be highly valuable, but they require conscious decisions about when that depth is justified and how to monitor its operational impact.

Strategic implications: skills, build vs. buy, and timing

The move toward skilled, personalized agents also reframes several strategic questions for enterprises. In the podcast discussion, the hosts highlight a few emerging themes:

• Organizations that do not experiment with AI “skills” — reusable, task-specific capabilities agents can invoke — risk falling behind. The suggestion is that those who delay may find themselves “toast” as competitors embed skills into core workflows.
• The classic build-vs.-buy question for enterprise software is becoming more urgent. As platforms like Zoom, Claude Cowork, and OpenClaw race to capture user context, enterprises must decide whether to rely on third-party ecosystems for personalization or invest in their own context and skills frameworks.
• The concept of “skills” may matter more than underlying protocol abstractions such as MCP (Model Context Protocol) for practical adoption. What matters to end users is what the agent can reliably do for them, not how context is technically routed.

None of these questions have definitive answers yet, but the direction of travel is clear: enterprises are under pressure to turn generic AI capabilities into tailored, role-aware systems that feel native to their users’ work. That requires both experimentation and a willingness to rethink ownership of user context.

What IT and product leaders should do next

For decision-makers evaluating personalized AI tools, the examples discussed in the Beyond the Pilot podcast point to several immediate actions:

• Define your context strategy. Clarify what user and enterprise context you are willing to centralize, where it will live, and which tools may access it.
• Insist on explicit controls. Favor vendors that, like Zoom, expose granular permissioning, feature toggles, and visibility into agent behavior rather than opaque “magic.”
• Integrate security from the start. Treat autonomous agents as first-class components in your security model, not as add-ons, and plan for safe deprovisioning and uninstalls.
• Track both value and cost. Establish metrics that capture user impact and business outcomes alongside token usage and infrastructure spend.
• Pilot skills in real workflows. Focus early experiments on concrete, repeatable tasks where personalization offers clear leverage, and iterate based on observed behavior and feedback.

The underlying technologies will continue to evolve, but the central challenge is already visible: turning powerful, generic models into responsible, deeply personal agents that genuinely understand users — without overstepping security, governance, or economic bounds.