An AI agent inside your CRM can only run the retention strategy you hand it: a map of the customer journey, a clear job for every message, a control group that shows what works, and limits on what it may change. We build that groundwork for clients. With Digital3x, we mapped 80+ automated journeys for a multi-asset broker, and conversion from registration to completed know-your-customer (KYC) verification rose 50%.
At the prop-trading firm QT Funded, every automation in the program is measured against a permanent 10% control group. That is the groundwork an agent needs before it can add anything.
CRM vendors are building agents into their platforms. On May 11, 2026, Salesforce announced its Summer ’26 release, available from June 15, and listed Multi-Agent Orchestration in Agentforce first among its top updates. Salesforce says the feature lets agents “work together as a unified team to solve complex, end-to-end workflows,” with one point of contact for the customer. It is framed around customer questions, not post-purchase marketing, but agents that share context still act on journeys, messages, and success measures that a person has to define.
What does an AI agent need from your retention program?
We look for four things before any automation, AI or not, touches a customer list.
1. A journey map with a goal for every stage
An agent needs to know where each customer is and what the next step should be. The multi-asset broker’s 18-month program, delivered with Digital3x, had journeys from registration and KYC verification through first deposit and long-term retention.
That architecture is what an agent would have to be given: named stages, the trigger that moves a customer from one stage to the next, and the outcome each journey exists to change.
2. A clear job for every post-purchase message
We learned the post-purchase side of this work at a consumer wellness brand. There, we built Klaviyo flows that gave each message after checkout one job: teach customers to use the product well, confirm their decision with reviews and testimonials, remind them to reorder when they were likely to run out, or ask for a referral only once the value was clear.
One in two first-time buyers went on to buy at least one more product, nearly one in three bought the same product again, and customer lifetime value (LTV) rose almost 300% year over year. We cover how those post-purchase flows were built and timed in a separate post.
An agent connected to the same platform should not be left to work out those jobs on its own. Someone has to write them down as rules it can follow, including when each message is due.
3. A permanent control group
An agent that can launch and adjust messages can also report on them, and open and click rates do not show whether a message changed what customers did. In our QT Funded case study, the firm’s program grew from 2 to 12 automated journeys, and the permanent 10% control group receives none of them, so each journey’s incremental impact is visible.
Give an agent the same rule: no journey counts as a win until it beats the holdout. Keep the holdout out of marketing journeys only; service, account, and legally required messages still go to everyone.
4. Permissions that match the job
Agents can now connect to marketing platforms directly. Klaviyo, for example, documents an MCP (Model Context Protocol) server that lets AI clients such as Claude or Cursor work with account data, from flow performance reports to campaign creation.
Its setup guide lists a read-only setting that disables every tool able to write to the account, and a setting that turns off specific tools that read user-generated content, with a warning to review each tool call for malicious instructions. Both are off by default, and Klaviyo notes that they require a custom connector rather than its listed Claude connector or ChatGPT app.
We would start any agent in read-only mode, where it can analyze performance and draft copy outside the account for a person to review. When it later gets write access, a person should still approve anything that sends.
The same rules apply to agents that talk to customers directly. We cover scope, handoff, and disclosure in customer-facing AI agents: guardrails before go-live.
Where should you start before connecting an agent to your CRM?
- List the journeys you already run, the stage each one serves, and the outcome it is meant to change.
- Write one sentence on the job of each post-purchase message (teach, confirm, remind, or ask for a referral) and when it should go out.
- Set aside a permanent holdout group that receives no marketing automation before you add new journeys, so later results have something to beat.
- Connect any agent read-only first, then widen its permissions one job at a time.
Right after a sale, your CRM has a lot to work with: a customer who has just made a decision and wants it to be the right one. An agent can help you use that moment, but only inside a journey you have mapped and a test you can trust. If you are planning that work, here is how we approach AI-driven CRM and workflow automation.
Alchemy Avenue is a marketing agency, not a broker, and does not provide investment advice.
Sources: Salesforce, “Summer ’26 Release: 10 Innovations Bringing the Agentic Enterprise to Life — Plus a Few Extras,” May 11, 2026. Klaviyo Developers, “Klaviyo MCP server” and “Connect to the MCP server,” accessed October 23, 2026. Our QT Funded case study, linked above, accessed October 23, 2026. The multi-asset broker’s figures come from our own client reporting and are shared anonymously. The wellness-brand figures come from a Klaviyo program our team ran.
Updated October 23, 2026: we rewrote this post, re-checked its facts, and added sources.
