Configuring the Feedback-Routing Agent

In a complex production, feedback arrives in a chaotic stream across email, Slack, and meeting notes. To stop this from flattening your creative idea, you use a feedback-routing agent. An agent is a specialist AI that holds a recurring job—in this case, monitoring your communications to gather and organize input.

The most critical rule for this agent is its lack of authority: it can gather, prepare, and flag, but it can never approve work.

For example, imagine you set your agent to watch a shared email chain for a new campaign. When the CEO emails the group saying the imagery needs to feel "warmer," the agent identifies this as a creative request and logs it in your tracking document. However, if the CEO later writes "this looks great," the agent does not mark the asset as "approved" in the system. Instead, it flags the comment for your review.

By removing the power of approval from the AI, you ensure that the Creative Director remains the sole gatekeeper of the work, preventing a stakeholder's preference from being treated as a final creative decision without your oversight.

To automate feedback collection, you need a proactive agent. Unlike a standard chatbot that only responds when you ask it a question, a proactive agent runs on its own cloud computer. This allows it to work in the background while you are busy, connecting directly to your apps to monitor for changes in real time.

Two primary platforms for this are Dots (by OpenAI) and Claude Duet (by Anthropic). Both act as specialists on your team, but they may differ in how they integrate with your specific toolstack.

For example, imagine your marketing and legal teams post their notes in a dedicated Slack channel. You would choose either Dots or Claude Duet to act as your monitor. You grant the agent permission to access that specific channel; it then stays "awake," scanning every message. When a legal lead posts a note about a disclaimer, the agent doesn't wait for you to prompt it—it proactively captures the text and routes it to your feedback log.

As a lead, your task is to select the platform that best connects to your current ecosystem, ensuring the agent can reach the sources where stakeholders actually speak.

Once you have chosen a platform, you must give your agent a remit. A remit is essentially a professional job description written for an AI. It defines three critical boundaries: what "wakes" the agent (the trigger), which sources it is allowed to access, and its strict operational limits. Without a precise remit, an agent may pull irrelevant data or, worse, overstep its authority.

Consider a scenario where you are managing a high-stakes brand launch. You write a remit that instructs the agent to "wake" specifically whenever a new email arrives from the "Stakeholder" email group. You then define its sources as that specific email thread and the project's shared Google Drive folder. Finally, you establish a strict limit: the agent must never contact the client or stakeholders directly.

By defining these parameters, you transform the AI from a general-purpose tool into a specialized team member. As a Creative Director, you use the remit to ensure the agent is proactive enough to catch every note, but constrained enough to never create a communication crisis by emailing a CEO without your approval.

After writing the remit, you must specify the exact data sources the agent is authorized to monitor. If you give an agent broad access to your entire workspace, it will pull in 'noise'—irrelevant chatter, unrelated project threads, or social banter—that clutters your feedback log and makes it impossible to find actual decisions. Restricting the agent to a narrow set of authorized channels ensures that every piece of data captured is project-specific and actionable.

For example, instead of letting the agent monitor all of your Slack or Gmail, you configure it to pull exclusively from two sources: the project's dedicated email thread (e.g., "Project-X-Feedback@company.com") and the specific "Feedback" folder in Google Drive where stakeholders upload marked-up PDFs. If a stakeholder mentions the project in a general company-wide chat or a random DM, the agent ignores it.

As a lead, you define these boundaries to create a 'clean' data pipeline. By explicitly excluding general company chats and unrelated folders, you ensure the agent only surfaces input that is intentionally directed toward the creative work, saving you from sorting through hundreds of irrelevant notifications.

When an agent collects feedback, there is a temptation to ask it to 'summarize' or 'clean up' the notes to make the log easier to read. However, for a Creative Director, this is a dangerous failure mode called 'flattening.' Paraphrasing removes the emotional nuance and specific language a stakeholder used, which are often the only clues to what they actually want. If you lose the original wording, you lose the ability to judge whether a comment is a hard requirement or a vague feeling.

For example, imagine a stakeholder writes: "It feels a bit too corporate." If the agent paraphrases this as "Stakeholder suggests a less formal tone," it has changed the feedback from a feeling about brand identity into a task about tone. "Too corporate" might mean the colors are too safe, the font is too rigid, or the imagery is too generic; "less formal" suggests a different problem entirely.

To prevent this, you configure your agent's remit to copy comments verbatim. You instruct the agent to capture the raw text exactly as written, then place any AI-generated categorization (like 'Visuals' or 'Legal') in a separate column. This keeps the source of truth intact while still allowing for organization.

Even with a defined remit and clean data sources, an agent still needs a specific routing trigger to know which pieces of a conversation are actionable feedback and which are just social chatter. A trigger is a logic rule that tells the agent: "When X happens, move this data to Y." Without this, your tracking sheet would be flooded with every "Thanks!" and "Got it" from the email chain, forcing you back into the role of manual sorter.

For example, you can configure a trigger based on specific keywords. You instruct your agent to scan every incoming email in the project thread for words like "Change," "Revision," "Correction," or "Update." When the agent detects the phrase "I have a small revision to the headline," the trigger fires. The agent immediately extracts the sentence, identifies the sender, and routes it into your central tracking sheet in the correct column.

As a lead, you define these triggers to create a filter that separates signal from noise. You aren't just collecting data; you are designing a system that only alerts you when a stakeholder is actually attempting to modify the creative work.

The final step is to stress-test your agent to ensure it is a helpful assistant and not an unauthorized decision-maker. A properly configured agent should act as a notification system that flags a request for your attention, rather than taking a creative action itself. If your agent attempts to modify a file or mark a task as complete without your sign-off, your remit is too broad and must be tightened.

To verify the setup, perform a live test. Send a sample email to the project thread from a test account, using a trigger word like "Revision." For example, write: "I have a quick revision for the legal disclaimer on the footer."

Observe the agent's reaction. The correct behavior is for the agent to generate a notification or a log entry saying, "New feedback received from Legal," and then stop. If the agent instead attempts to open the source file and edit the text, it has overstepped. You then revisit the remit to reinforce the limit: "Gather and flag only; never modify assets." Once the agent consistently flags without acting, you have a reliable system that protects your creative custody.