Consent-aware personalization: agents should respect preferences by design
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Personalization used to be limited by effort. Writing a tailored message for every contact was too expensive, so most personalization stayed shallow: a first name, a company name, a segment-level message. Agents remove much of that cost.
An agent can research a contact and write a message that reflects their role, company, and recent activity. That capability raises the stakes on a question marketers sometimes treat as an afterthought: what are we allowed to use, and what would the person on the other end consider appropriate? In our research, we treat both halves of that question as design requirements for any agent that writes to people.
This post describes practical design patterns. It is not legal advice, and the rules that apply to your data and your messages depend on where you operate and who you contact. Confirm the specifics with counsel.
Two separate questions
Consent and appropriateness are related, and they fail in different ways. We find it useful to keep them apart.
Is it permitted? Consent rules, privacy laws, platform terms, and your own policies define what data you may use and which contacts you may message. These requirements vary by region and channel. Your legal team should define them, and your systems should apply them without exception.
Is it welcome? Even when something is permitted, highly specific personalization can feel intrusive. Referencing a prospect's recent personal social media post in a cold email may be allowed and still damage trust.
Permission is a yes or no question that a system can check. Welcome is a judgment that changes with context. That difference decides where each rule should live.
How agents change the risk
Before agents, a person assembling a campaign would usually notice when a list looked wrong or a detail felt too personal. That informal check disappears when an agent pulls data and writes at volume. The agent will use whatever it can reach, in whatever way its instructions allow.
Agents also widen what can be combined. An agent with access to a CRM, web research, and product analytics can join facts that were never meant to sit in the same message.
Each source may be fine to use on its own. The combination can still surprise the recipient. Our post on identity resolution in the age of agents covers how records get joined and why that deserves care.
Enforce permission in code
Consent and opt-out status should never depend on an agent remembering an instruction. Models follow instructions most of the time, and most of the time is not good enough for consent. We recommend four checks built into the system around the agent.
- Eligible retrieval: agents can only retrieve contacts who are eligible for the intended use.
- Check at send: message-sending tools verify consent at the moment of sending, not when the draft was written.
- Upstream filtering: excluded regions, suppression lists, and opt-outs are applied before data reaches the agent.
- Logged basis: every send is recorded with the consent basis that allowed it.
The check at send matters because consent can change between drafting and sending. A contact who opts out in the morning should not receive an email the agent drafted the night before. This is the same principle as guardrails as code: rules that matter belong outside the model.
The logged basis matters later. If someone asks why a person received a message, you need to answer from records rather than memory. See audit trails that satisfy legal, brand, and finance for how to design those records.
Guide what is welcome
For appropriateness, written guidance plus human review works better than code. We suggest starting with four guidelines.
- Professional context first: use role, company news, and public work rather than personal details.
- Expected knowledge only: reference only information the recipient would expect a business to know.
- No sense of surveillance: avoid implying that you watched them, such as listing pages they visited.
- Less when unsure: when in doubt, personalize less.
Include contrasting examples in the agent's instructions so the boundary is concrete. Agents follow examples more reliably than adjectives. A pair that shows "mentioned your team's recent product launch" next to "saw you were browsing our pricing page at 11 p.m." teaches more than a paragraph on tone.
A worked example
Consider a hypothetical B2B team with an agent that drafts follow-up emails after webinars. The agent can see the attendee's company, role, questions asked during the session, and pages viewed on the website afterward.
The first drafts are accurate and a little unsettling. One opens by noting that the reader returned to the pricing page three times. Nothing in it is false, and the team's consent records may well allow the use. It still reads as if someone was watching.
The fix comes in two parts. The team removes page-level browsing data from the agent's inputs entirely, since the drafting task does not need it. It then adds examples to the instructions that reference the webinar topic and the attendee's own questions, which the reader knowingly shared. The emails stay personal, and the reader is not left wondering what else the company knows.
Minimize what agents see
The safest data is data the agent never receives. Pass agents only the fields they need for the task. A drafting agent writing to a segment does not need individual browsing histories. A research agent preparing account briefs does not need personal contact details.
Data minimization also helps quality. Fewer fields mean less noise in the agent's context and fewer chances to combine facts in ways that surprise the reader. Our post on the marketing data contract describes how to define which fields each agent can depend on.
What to measure
Have reviewers check personalized messages for anything that feels too familiar, and track these as their own rejection category. A steady trickle is expected. A rising count suggests the guidelines or the inputs need adjusting.
On the permission side, we suggest two checks. Count sends blocked by the consent check at send, which shows how often consent changed between draft and delivery. And sample sent messages monthly to confirm each has a logged consent basis.
The question we are studying next is how to test appropriateness before launch. A panel of readers can judge a sample of messages, but their sense of what is welcome varies by region, industry, and age. Whether a small, diverse panel gives stable enough judgments to set guidelines is something we plan to test.
How to apply this: confirm with counsel which consent rules apply, then check that they are enforced in your sending tools rather than only in your prompts. Write five examples of personalization that is welcome and five that is not, and add them to your agent's instructions.
Related research
Red-teaming marketing agents: find the failure modes before customers do
Deliberately trying to make your agents fail is one of the fastest ways to make them safer. Here is a practical red-teaming process for marketing teams.
Guardrails as code: the rules that matter belong outside the model
Rules written only in a prompt are requests. Rules enforced in code are controls. Important limits on marketing agents belong in the second category.
Escalation paths: agents should know when to stop and ask
Good agents know their limits. Clear escalation rules, and a fast way for people to resolve what agents raise, prevent many of the mistakes that damage trust.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
