Company Brain: How Lead with AI Is Connecting Knowledge Without Losing Control

Company knowledge sits across CRMs, learning platforms, Google Workspace, and other systems. Employees can connect AI to those services directly, but Tate argued that many integrations make it difficult to limit precisely which data an agent can access.

He gave the example of connecting an AI agent to Slack or email. Once a connection is authorized, the scope of what it can read may be broader than the employee expects.

The challenge isn't just retrieving an answer, but controlling what information reaches the model.

“This is not just a technology problem,” Tate said. “It’s people, process, and technology.”

Lead with AI’s response was to build shared controls rather than rely on individual employees to remember every privacy rule.

What does the Company Brain actually do?

The Company Brain sits between familiar AI tools and the systems where business information already lives. Employees can work in ChatGPT, Claude, or Cursor through company plugins. Those plugins connect to the Brain rather than directly to each underlying application.

Instead of copying all company information into one large database, the architecture uses indexes to locate relevant data in its source systems.

It can retrieve information when needed and keep selected information in an operational cache to improve speed. Tate noted that keeping the cache current remains a technical challenge.

The shared layer also controls access. For example, when employees authenticate through Google Workspace, permissions depend on their roles.

Similarly, a cohort engager can access only the cohorts assigned to them, while course-content staff may have additional permissions to edit lessons.

“We want the people closest to the work to build the automations,” Tate said. The shared infrastructure makes that experimentation safer.

Tate described a privacy screen that pseudonymizes personal information before passing it to an AI model. Depending on the context, a participant’s name might be replaced with a numbered identifier such as “Learner645.”

The system is designed to apply different privacy levels for different uses, including demonstrations.

He showed how the Course Operations plugin could answer questions about a cohort while limiting access to personally identifying details.

Importantly, these were descriptions and demonstrations of the team’s implementation, not an independent security audit or a guarantee that every risk has been eliminated.

“Privacy really needs to be a design decision from the start,” Tate said.

He also shared a lesson from his first implementation.

He didn’t spend enough time separating the AI-facing connection from backend data access. As such, adding a layered gateway and role-based controls later required substantial rework.

Why build a Company Brain instead of buying one?

Tate pointed to three reasons for building internally: tailoring the system to the workflows Lead with AI actually needs, keeping the freedom to switch among AI providers, and applying its own privacy and access rules.

As the company adopts new AI tools or business software, the team can connect them to LughOS without rebuilding every workflow. Tate also emphasized that building it internally lets the company learn from real usage before bringing similar approaches to clients.

That does not make building the right choice for every organization.

In our case, building internally made sense because the team had the technical expertise and wanted greater control over how the system worked.

Tate acknowledged that parts of the platform, particularly caching and the privacy layer, still need refinement.

Tate explained how employee experiments can become tools that the whole company uses:

1. Employees who do the everyday work experiment with innovative ways to use AI,
2. Teams decide which approaches are useful or reusable for wider adoption,
3. Finally, the technology team tests and integrates promising work. Approved processes, tools, and skills then become what he calls “golden paths” for wider use.

In the demo, he showed separate staging and production plugins. Staging lets colleagues test emerging capabilities, while production holds tools that have passed internal review.

He also described a possible course-operations workflow: a team member could use the available tools to create a daily briefing on intake completion, quiz results, and learner progress, rather than asking a central engineer to design that reporting process from scratch.

Tate also emphasized the importance of failures. He described a skill for organizing an unsuccessful AI conversation for colleagues to review. A failed attempt can show the team where a workflow needs a different approach or additional support.

What lessons should leaders take from the experiment?

Tate’s most concrete lessons were more organizational than technical. He mentioned the need to build privacy into the design, organize access around roles, pilot workflows before rollout, and make people accountable for checking AI outputs.

He also emphasized usefulness. “If we build stuff that is going to slow people down, they’re not going to use it,” he said.

Adoption depends not only on speed but also on whether the automated work meets the team’s expectations for quality.

More importantly, an internal AI platform should not simply make information easier to retrieve.

It should help employees redesign their work and share what they learn.

By handling access and privacy centrally, LughOS gives employees more freedom to experiment with AI and improve their own workflows.

In other words, centralized technical controls can support decentralized employee experimentation.

• Build access and privacy controls into AI workflows from the beginning, rather than relying on policies alone.
• Keep company knowledge in its source systems where practical, with clear retrieval and caching rules.
• Let teams closest to the work design workflows, then test and promote reusable approaches.
• Make sure employees can only access the AI tools and company information they need for their roles.
• Measure whether an automation actually improves speed or quality before rolling it out.

This article draws on Gary C. Tate’s demonstration and discussion during a Lead with AI PRO briefing.