Your Regulars Don't Quit Loudly. Teach an Agent to Notice.

Nobody cancels anymore. They just stop showing up.

It is a slow Tuesday afternoon at Ana's nail and spa studio in Manila. Six chairs, four staff, a client list she has built by hand over nine years. She does not think about it often, but somewhere in her booking app sit three hundred and forty names who have not walked through the door in more than two months. Some moved across the city. Some found someone cheaper two blocks over. Some are still loyal and simply forgot to rebook after their last trip abroad. Ana cannot tell which is which, not without scrolling through three hundred and forty names by hand, one at a time, guessing.

This is not a Manila problem. It is not even a salon problem. A plumber in Manchester has the same three hundred and forty names sitting in his invoicing software. A dental clinic in Lagos has them in its patient system. A specialty coffee subscription in Austin has them in Shopify. Every repeat-visit business on earth is quietly losing customers who never said goodbye.

How Do You Build an AI Agent That Wins Back Customers Who Have Gone Quiet?

You teach an AI agent your business rules one time, then hand it a standing objective instead of a one-off task. Point it at your customer or booking export, tell it what "gone quiet" means for your business (30 days for a coffee shop, 90 days for a dentist), show it two or three messages that actually got a reply in the past, and then give it a goal it keeps working on without you re-typing the prompt every week. Modern agent tools now do the checking part automatically, on a repeating schedule, inside the same session, which is the part that used to require a developer, a cron job, or a paid retention platform.

Here is how win-back usually works right now, in a business with no dedicated marketing team: someone remembers, eventually, usually after a slow month, to pull a client list and manually re-type a discount code into forty text messages. It happens in bursts, driven by anxiety about revenue, not by any actual signal from the customer. Most months, it does not happen at all, because nobody has forty spare minutes and the will to feel a little pathetic asking for business back.

The reframe: the offer was never the hard part. Ten percent off is easy to write. The hard part was noticing, on a Tuesday, in real time, that a specific person had crossed a specific threshold. That is a job for something that never gets busy, never forgets, and never feels awkward about checking.

The Two Commands That Change Everything: Teach Once, Then Hand Off

This week, Shubham Saboo, a senior AI product manager and the author of one of GitHub's most-starred open source AI repositories, posted that Hermes Agent, the open source AI agent built by Nous Research, had shipped something called Recurring Loops. "Give it a task and it repeats," he wrote. "Leave the interval off and it picks its own pace. Like cron, except it runs inside your session with full context." Cron is the decades-old tool developers use to schedule a script to run automatically, on a server, forever, in the background. Recurring Loops brings that same always-on repetition to an agent that already knows your business, without you needing a server, a developer, or a subscription to a separate scheduling tool.

To see the underlying pattern in action, it helps to watch someone actually build with it. In a widely viewed tutorial, AI automation teacher Derek Cheung demonstrates the two commands that make this possible: /learn and /goal. He starts with a Hermes agent that has never heard of a particular research API. He points /learn at the API's documentation once. A few minutes later, the agent has written itself a permanent, reusable skill. Then he hands it a /goal: enrich five blank company records in a spreadsheet, cite every source, do not stop until every field is filled. The agent searches the web, extracts the real data, and writes the completed spreadsheet back, unsupervised. Then, to prove the skill generalizes, he hands it a second spreadsheet from a completely different industry. No reteaching. Same skill, new data, same result.

Swap "enrich this spreadsheet" for "check this customer list every Monday," and you have the exact shape of a win-back agent. Teach it once what your business considers gone quiet, what tone your messages use, and what a good past win-back message looked like. Then hand it the standing goal. With Recurring Loops, it keeps checking on its own cadence, not the one time you happened to remember.

Why This Works Now, and Did Not Six Months Ago

Retention has always mattered more than most small business owners act like it does.

That number has been true since the 1990s. What changed is not the economics of loyalty. What changed is that checking for it used to require either a person with spare time, or a paid platform built for enterprise marketing teams. Now it requires an agent you already have a reason to be using for other things, taught once, and left running.

Isn't This Just an Automated Marketing Campaign?

Fair question, and worth answering honestly. A Google search for "AI agent win back customers" surfaces platforms like ChurnZero, Outreach, and Braze, real tools that do this well at enterprise scale, usually starting at a price point built for teams with a dedicated head of retention. They are built to plug into a large CRM and run predictable, templated sequences.

What Ana's salon, a Manchester plumber, or a Lagos dental clinic needs is smaller and cheaper: one honest check against one spreadsheet, once a week, with a message that actually sounds like the business, not a template with a merge field. A recurring agent loop is not a marketing automation platform. It is closer to a very reliable, very literal member of staff whose entire job is to notice one thing and draft one message, then wait for you to say yes.

Build It: The Five-Step Recurring Win-Back Loop

1. Export your customer or booking list. Most booking software, point-of-sale systems, and CRMs have a CSV export button. Columns you need at minimum: name, contact method, last visit or purchase date.
2. Teach the agent your rules, once. Tell it, in plain language, what "gone quiet" means for your business, what tone to write in, and paste in two or three messages that got a real reply in the past. In Hermes Agent this is the /learn command; in Claude, it is the same idea using a saved skill or project instructions.
3. Give it the standing goal. Something like: "Check this list every Monday. Anyone past 60 days since their last visit, draft a short, warm, personal message. Do not send anything. Hold everything for my review." That last sentence matters more than any other line in this guide.
4. Review the first batch by hand. Read every draft before anything goes out. Fix the ones that sound off. This is also how the agent, and you, calibrate what "good" looks like for your specific customers.
5. Turn on the recurring cadence. Once you trust the drafts, let the loop run weekly or monthly on its own. You are still the one who reviews and hits send; the agent is the one who never forgets to check.

What You Have After This

A week from now, you have a reviewed batch of honest, specific messages going out to people who genuinely forgot about you, not a blast to your entire list.

A month from now, you have a habit, one that used to depend entirely on your memory, now running whether you remembered Monday or not.

Six months from now, you have a client list that quietly maintains itself, and a number, real customers who came back who would otherwise have drifted, that you can actually point to.

Same client list. Same staff. A completely different relationship with the people who already chose you once.