Embedded Brand AI System

This is where an employee encounters Brand AI while doing real work: developing a product narrative, naming a capability, preparing an executive communication, reviewing a claim or translating technical information for an audience.

The goal is not to make employees leave their work, find a brand portal and interpret the right document. Brand intelligence should be accessible inside the environment where the work already happens.

• CoWork is the likely primary conversational surface for business users.
• Teams or Microsoft 365 Copilot can bring the agent into an existing collaboration environment.
• A custom application can provide a purpose-built Brand AI experience.
• A connected workflow can invoke Brand AI through a button, sidebar, review gate or authorized event.

CoCo belongs primarily in the build-and-operate environment, where technical teams create, test and manage agents and skills.

Experience decision: For each use case, determine where employees already perform the task and what form of support creates the least friction.

A skill is an applied brand playbook an agent can execute. It turns general guidance into a repeatable process: what questions to ask, what sources to consult, what rules to apply, what tools to use, what output to create and when to involve a person.

• Git repository: The development workshop for version history, peer review, testing and rollback.
• Snowflake named stage: A secure, named file location governed through Snowflake roles. It can distribute approved skill files inside the organization.
• Cortex Extension: A formal, versioned Snowflake package containing a skill or plugin, its files and metadata. It supports certification, access control and reuse across agents.

A likely progression is to develop in Git, release approved versions into Snowflake and package coordinated capabilities as Cortex Extensions when reuse and scale warrant it. A Cortex Extension packages capability; a Cortex Knowledge Extension packages searchable knowledge.

Operating decision: Brand leadership owns strategic accuracy, use-case priority and approval criteria. The Brand AI Architect owns technical implementation, testing and release mechanics.

Skills define how work should be performed. Brand intelligence supplies what the agent must know while performing it.

The architecture should separate the authoritative system of record from Snowflake’s governed knowledge foundation and from the mechanisms that retrieve it. Existing content, research, product or asset systems may remain authoritative; Snowflake can ingest or synchronize what the Brand AI System needs.

• Structured brand intelligence belongs in governed tables: approved claims, product names, audiences, message hierarchies, owners, status and effective dates. Semantic Views explain the relationships so Cortex Analyst can interpret them.
• Unstructured brand intelligence includes strategy documents, research, narrative guidance and approved examples. Files can live in a Snowflake stage or synchronize from another system; Cortex Search indexes them for retrieval.

Cortex Search is the retrieval layer, not the source of truth. A Semantic View is the business-meaning layer over structured data, not the data store.

Recommended authority order

1. Active, explicitly approved structured rules
2. Current approved strategy and guidance
3. Approved examples and applications
4. Current research and contextual evidence
5. Historical material, clearly labeled
6. Draft or superseded material, normally excluded

Brand intelligence decision: Define what is authoritative, who owns it, how current it is, and what happens when sources conflict. Relevance alone should never outrank authority.

Skills provide instructions. Knowledge provides context. Tools give the agent the ability to retrieve, compute, create or take an approved action.

The agent configuration establishes what the agent is permitted to use. The skill contract explains which tools a particular workflow requires, when to call them, what inputs and outputs are expected, and what to do if a tool is unavailable.

For example, a Claims Review skill might retrieve a claim, call an approved-claims registry, search for supporting evidence, compare wording and scope, and route any unapproved claim to a human owner.

Not every skill needs a tool. When tools are required, their use should be explicit, testable and governed through the employee’s permissions.

Governance decision: Give each agent the smallest set of tools required for its defined work. More capability is not automatically better capability.

The model is the generative reasoning engine, but it is not the Brand AI System. It does not inherently know Snowflake’s positioning, approved claims, decision rights or brand processes. Those come from the surrounding agent, skills, knowledge, tools and policies.

The default path is an approved foundation model accessed through Snowflake Cortex, selected for task quality, reliability, cost, speed and regional availability.

A compatible Hugging Face model can be imported into the Snowflake Model Registry, versioned and deployed as a Snowflake inference service. It then operates inside Snowflake’s governed environment, but it does not automatically become a native Cortex foundation model or the Cortex Agent’s central orchestration model.

The Brand AI Agent would typically call that specialized model as a tool for a narrow task, then use the result in its broader reasoning.

Model decision: Prove the use case with strong skills, trusted retrieval and evaluations first. Add a specialized model only when measured evidence reveals a capability gap.

Value is realized in the employee’s work. The system may recommend a direction, create a draft, review existing work or take an approved downstream action, but the surrounding experience determines how that assistance appears.

• Conversation: The employee asks the agent in CoWork, Teams or Microsoft 365 Copilot.
• Embedded assistant: A sidebar, add-in or menu offers actions such as Ask Brand AI, Review this content or Check this claim.
• Workflow checkpoint: Support appears when work reaches a meaningful stage, such as launch submission or executive review.
• Defined event: A narrow, authorized trigger initiates a review and returns a comment, risk flag, recommendation or request for approval.

The future-state aspiration is not constant pop-ups. It is guidance that is helpful by default and proportionate by design: light-touch when stakes are low, more active when brand consequences are meaningful, and human-led when judgment or approval is required.

Intervention decision: Progress from on-demand help to opt-in contextual support, then natural workflow checkpoints, with automatic intervention reserved for clearly defined risks.