How Brand Guardrails Fail When AI Agents Operate Without Shared Context
Shared context prevents agents from making unauthorized commitments in your brand's name.

To spot when AI output drifts off-brand, teams must check it against the same current, versioned brand context used to generate it. Tone checks alone cannot catch an agent that sounds right while making a promise it has no authority to make.
Why AI agents are now making brand-facing decisions — not just drafting copy
AI agents speak and act for brands when they explain products, handle complaints, promise refunds, and complete transactions. These activities expose different parts of the customer relationship. An explanation puts trust in the brand at stake, while a promise of a refund can bind the brand to pay it. When an agent completes a transaction, it goes further by directly shaping the service a customer gets. The escalation is from speaking to committing to acting.
Yet the word “agent” does not establish where that authority ends. OpenAI defines agents as “systems that independently accomplish tasks on your behalf,” while Anthropic describes “systems where LLMs dynamically direct their own processes and tool usage.” Neither definition tells a brand whether a particular deployment may quote a price or authorize a refund. That boundary needs its own explicit definition.
Defining AI Agents: A Compendium of Criteria, Metrics, and Benchmarks describes substantial mismatches between agent behavior and users’ mental models, including uncertainty about capabilities and the user’s role. That uncertainty complicates customer-facing interactions because customers may interpret an agent’s statement as the brand’s commitment. Governing what agents may say, promise, and do is therefore both a brand management function and an engineering responsibility.
What "guardrails" actually means in an agentic workflow, and where the term breaks down
Guardrails are constraints on what an agent can say or do. They include tone rules, prohibited claims, topic limits, and conditions that require escalation to a human. Crocs designed its brand agent Rivet with guardrails against inappropriate tone shifts and monitors both what it says and how it says it. This is deliberate governance of brand representation. It treats behavior and wording as part of identity.
A prohibition on unauthorized offers depends on knowing which figures are approved offers, who approved them, and what authority the agent has. Guardrails set when an agent is deployed often assume that the context will stay the same. When several agents operate on separate copies, each may apply a valid rule to a different picture of the brand.
The compendium identifies temporal coherence and environmental interaction as dimensions of agenticness, meaning consistency over time and interaction with the surrounding environment. Both are central to this problem. A brand rule cannot work reliably if the state it reads is out of date or has parts missing. Regulating AI Agents also identifies significant obstacles for the EU AI Act in addressing autonomous task performance failures, underscoring the limits of governance designed for conventional AI systems.
The BMW Toronto incident: what context propagation failure looks like at the brand touchpoint
In June 2026, customer Zack Giacomelli exchanged messages with BMW Toronto’s AI agent Quinn about a car buy-back. Quinn offered the exact outstanding balance on his auto loan as the purchase price. A human salesperson later said the actual offer was substantially lower. The original offer was subsequently reinstated after public reporting. The dealership absorbed the gap to end the reputational dispute.
The causal sequence began with human miscommunication when an employee passed the loan balance into the agent’s input, and Quinn relayed it as a buy-back offer. BMW Toronto sales manager Scott Shadbolt explained that Quinn was meant to pass on offers made by humans and had no authority to negotiate on its own. The agent lacked shared context distinguishing a loan balance from an approved commercial offer.
That distinction changes the diagnosis and the remedy. Better prose generation would not have corrected the missing commercial meaning. In a separate Canadian bereavement-fare case, a tribunal rejected an airline’s argument that its chatbot was a separate legal entity responsible for its own responses, finding the company responsible for information published on its website. A completed transaction was not necessary to expose the brand to consequences. At BMW Toronto, the mechanism was absent shared context between the human data-entry step and the agent’s execution layer.
Why isolated context causes brand drift — and why it compounds across agents
Generative systems introduce variability even in similar situations. When several generation steps depend on one another, small deviations can accumulate. An agent may pass its interpretation to another agent that has a different copy of the guidelines. Without a common reference, neither has a dependable basis for correcting the divergence. More steps create more opportunities for drift.
Uploading a style guide once does not give agents that reference for all time. It may support strong initial output, but it is not a live connection to subsequent changes in voice, claims, or campaign direction. In a multi-agent workflow, the problem expands beyond stale instructions. Agents can work from incompatible assumptions or ignore one another’s outputs because they lack shared ground truth.
Even shared state can fail if it becomes overloaded. When an agent fills the available context with raw data, downstream agents can lose access to the brand rules needed to interpret it. The resulting loss of coherence cascades through the workflow. For agencies, another boundary is essential because two clients using the same generation surface can contaminate one another’s brand context. The structural requirement is shared context within a brand and governed separation between brands.
What shared context architecture actually fixes — and what it does not
Shared context gives agents one canonical set of brand facts in place of several separate readings of a guideline document. That source includes palette, typography, logos, voice, approved claims, and tone register. It must be live, versioned, and scoped. “Live” means agents can retrieve current rules, while versioning identifies which rules governed a task. Scope limits each agent to the context relevant to its work.
The orchestrator pattern implements this division of responsibility. One agent maintains the canonical brief, and specialized agents pull the portions they need for their deliverables. Those agents can then compare outputs against the shared brief. For agencies, the same principle requires a separate governed brand kit for each client. A shared pool of guidelines is not equivalent.
This architecture removes a cause of drift while leaving generative variability intact. It also cannot substitute for decisions about an agent’s authority. The Ritz-Carlton sets clear limits on what employees can decide for themselves, giving brands a useful model for deciding how much freedom to give agents. In Quinn’s case, accurate context needed to be paired with rules governing when a commercial figure could reach a customer. Shared context must therefore define behavioral scope as well as brand identity.
How MCP enables a live shared brand context — and the limits brands must plan around
MCP is a protocol that lets agents connect to tools and information outside the agent, including a brand’s current guidelines and approved assets. Instead of relying on information supplied at setup, an agent can query a style guide, consult previous campaigns, or retrieve approved assets at generation time. The source becomes current rather than merely remembered. This works only if the brand has made that context available in a usable form.
Anthropic donated the MCP standard to the Agentic AI Foundation under the Linux Foundation in December 2025, while its existing governance model and maintainer structure remained unchanged. SAP Commerce Cloud and Adobe Commerce have shipped MCP server implementations that expose catalog and brand data to agents. These implementations reduce the need for custom connection infrastructure. They do not decide which information is authoritative or which agents should access it.
Connections also take up context and bring security risks. Anthropic engineering reported in November 2025 that using code execution in place of upfront tool definitions cut token use from hundreds of thousands to a few thousand. Tool poisoning and command injection vulnerabilities pose another risk, especially when auto-approval is turned on. A live connection is not necessarily a trustworthy one.
For brands, these limits make MCP implementation a governance decision. What context is exposed, to which agents, and at what granularity determines how the brand can be represented. Leaving those choices entirely to engineers lets what is easiest to build drive brand strategy. A marketing decision is then being made by default in engineering.
Making brand identity machine-readable: what agents can actually use
A brand guidelines PDF is written for human interpretation. Agents need a structured representation of the same decisions that they can retrieve and apply consistently. Building that representation is translation. The human guidelines remain useful for nuance, rationale, and judgment, while a separate machine-readable artifact expresses the rules needed for execution. Both must come from the same identity and strategy.
For visual identity, that means design tokens, color values, typography specifications, and logo rules. For language, it means register, approved and prohibited wording, brand story, product claims, and approved sources. An agent can check a hex value, but it cannot reliably turn “use warm, approachable colors” into a consistent visual choice. Structured content models reduce that interpretive gap.
The translation must also cover experience patterns, who has pricing authority, what legal language to use, and the conditions that require escalation. Otherwise, a system may apply the right voice to an action the brand never authorized. Versioning connects these rules to change over time. When a campaign palette, tone, or product claim changes, connected agents should draw from the updated context. A brand that is difficult to query leaves more of its identity and authority open to approximation.
Detecting when AI output drifts off-brand before it ships
An off-brand output is one that violates identity rules, uses unsupported claims, exceeds delegated authority, or conflicts with brand priorities. The review must check visual identity, voice and tone, factual support, behavioral scope, and strategic alignment. These are distinct criteria for a reason.
Shared context makes those checks possible, but the checks do not happen on their own. Agents still produce output that varies, and they can produce more than editorial teams can check by hand. The review mechanism must therefore be as systematic as generation. Crocs demonstrates deliberate attention to this problem by monitoring Rivet’s personality and behavior as part of its brand identity, rather than treating correct information as sufficient.
Evidence-based review produces a record of the rules applied, which rules the output satisfied or failed, and why. For an agency, it also shows that the client’s own context governed the assessment. In the BMW Toronto incident, the wrong commercial commitment reached the customer without an effective review intercepting the authority failure. Checking outputs against the same versioned context used for generation closes that loop.
What the full architecture looks like for a team running multiple agents
The complete architecture has a context layer, a connection layer, and a review layer. The context layer holds machine-readable identity, approved claims, and behavioral scope in a live, versioned source. The connection layer gives agents access at task time through MCP or an API. The review layer checks outputs against that source before release and records the evidence for each assessment. Each layer has a different responsibility.
For a team producing campaign copy, web pages, launch decks, and customer-service responses, specialized agents draw only the relevant portions of the canonical brief. An orchestrator maintains and distributes that brief, while a review agent evaluates the resulting work. Updates reach connected agents without someone having to brief each one again. Agencies add a separation layer so that each client’s tasks and reviews stay within the context governed by that client’s rules.
MCP-compatible agents and products built around a REST API can draw from the same source, so the architecture need not depend on one generation surface. The governing reference remains the brand’s context. Its scope must cover both expression and permission, because consistency includes what an agent is allowed to promise. The operational test is whether each output can be traced to the relevant brand version and assessed against it before a customer sees it.


