Stop Being the Follow-Up. Build an AI Loop Instead.
To get an AI agent to follow up automatically without babysitting it, do not write it a better prompt. Build it a loop: give it one goal you can measure, let it act and grade its own work against that goal, tell it exactly which decisions get handed to you, and write down before you press run when it should stop. That combination, goal, evaluation, escalation, and a stopping condition, is what separates an agent you can leave running overnight from a chatbot you have to keep checking on.
You did not hire an assistant. You hired yourself, and gave it a chat window.
It is 9 p.m. In Manila. Mia runs a six-person events and party rental business, tents, linens, folding chairs, the works. Three weeks ago a supplier shipped 40 tablecloths for a wedding with a wine stain already baked into the fabric. The supplier promised a refund. That was three weeks and four unanswered emails ago.
Mia is not bad at her job. She is just the one who has to remember. Remember to check if they replied. Remember what day she last followed up. Remember to sound polite the fourth time she asks for the same thing. Remember to escalate to a phone call if email keeps going nowhere. Nobody put that job on her contract. She just ended up doing it, the way every small business owner ends up doing it, because someone has to hold the thread.
You Already Run a Loop. It Just Runs In Your Head.
Here is the reframe that matters: chasing that refund is not a task. It is a loop. Send, wait, check, escalate if needed, stop when it is resolved. You already do every step of it. You are just doing it manually, in your head, at 9 p.m., instead of handing it to a system that does not get tired of asking.
This is the distinction a lot of AI advice skips. A chatbot answers one question and forgets everything. A single prompt gets you one draft, and you are the one who reads it, catches the problem, and runs it again. That correction loop, judge the output, remember what went wrong, know where the line is, decide when it is done, is work. It just happens invisibly, inside your own head, one email at a time.
An AI loop is what happens when you write those four jobs down and hand them to a system: a way to judge the output (an eval), a way to remember what already happened (memory), a line it is not allowed to cross on its own (a guardrail), and a rule for when to stop. Do that, and the same follow-up that ate your evening runs itself while you are asleep.
What Is an AI Follow-Up Agent, and How Is It Different From a Chatbot?
An AI follow-up agent is software that tracks an open item on your behalf, an unanswered email, a late invoice, a claim, a booking, and keeps nudging the other side on a schedule, without you starting a new chat each time. A chatbot answers what you type into it right now and forgets everything the moment you close the tab. A loop remembers, acts on its own schedule, checks its own output, and knows when to stop or hand the decision to you. That third part, the checking and the stopping, is the part almost everyone skips when they first try to automate a follow-up.
| What you need | A chatbot | A loop |
|---|---|---|
| Remembers past attempts | No, starts fresh each time | Yes, that is its memory |
| Runs on a schedule without you | No, waits for you to type | Yes, wakes itself up |
| Grades its own output | No | Yes, against a goal you set |
| Knows when to stop | Never stops, never starts either | Yes, on a rule you wrote first |
| Knows when to ask you | Always asks you, it is a chat | Only for the calls that matter |
Isn't This Just a Fancier Chatbot With Extra Steps?
Fair question, and the honest answer is no, because of two things a chatbot does not have: an evaluator and a stopping condition. Without an evaluator, a loop is just automated guessing, it produces output and has no idea if the output is any good. Without a stopping condition, it is a billing machine, happy to keep running and keep costing you money without ever finishing anything. Add both, and you have something closer to a very patient, very literal junior employee: infinite patience, because it will happily try the same follow-up 40 times without getting tired or annoyed, but zero judgment, because it has no idea what "good" looks like unless you tell it.
The Five Things Every Follow-Up Loop Needs
A full production loop has nine working parts, but a small business does not need all nine on day one. Here are the five that matter for a first build, adapted from the loop-engineering framework AI product lead Shubham Saboo laid out this week.
- A goal you can measure. Not "chase the refund politely." Try: "the refund is confirmed in writing, or 21 days pass and the case is escalated to me." A number or a clear yes/no beats a wish every time.
- Context. What does the loop know going in? The original order, the promise the supplier made, the date they promised it by, your tone guide. Thin context gets you confident nonsense at scale, not just once.
- Actions and tools. What is it actually allowed to touch? Draft a reply, check the inbox, log the date, update a spreadsheet. Small, defined moves, not "do whatever it takes."
- An evaluator. Something that checks each attempt against the goal. Did the reply cite the claim number? Did it name the promised date? Is it under 150 words? This is the step most first attempts skip, and it is the one that keeps the loop honest.
- An escalation rule and a stopping condition. Anything involving money, a customer relationship, or a decision you cannot undo goes to you, always. And before you press run, decide on all three exits: a target (refund confirmed), a budget (21 days or a dollar cap on paid tools), and a stall rule (three follow-ups with zero reply means something is stuck and a human should look at it).

A Worked Example: The Refund Nobody Was Chasing
Back to Mia. Here is what her loop spec looks like once she writes it down instead of carrying it in her head.
Goal: refund of 8,400 pesos confirmed in writing, or the case is flagged to Mia after 21 days.
Context: the order number, the photo evidence of the stained linens, the supplier's original promise ("processing, allow 15 days"), and Mia's usual tone, direct, friendly, no exclamation points.
Actions and tools: read the shared inbox thread, draft the next follow-up, log the date of each attempt in a spreadsheet.
Evaluator: does the draft reference the order number and the original promise date, does it avoid sounding desperate or aggressive, is it three sentences or fewer.
Guardrail: it drafts, it never sends without Mia's one-tap approval, because this touches a supplier relationship she still needs next season.
Stop: refund arrives (target), 21 days pass (budget), or the supplier goes silent for two follow-ups in a row (stall, escalate to a phone call).
Notice what changed. Mia is not doing the remembering anymore. She is doing the approving, which takes ten seconds a day instead of an evening of hunting through old emails. The loop does not need her to be smart about tone every single time. It needs her to have been smart about it once, when she wrote the spec.
This is the same shift behind the AI loop that finds you customers while you sleep, a lead-generation loop we broke down here last week. Different job, same architecture: a goal, a check, and a rule for when to stop.
Why Most Follow-Up Bots Quietly Fail
Five failure modes show up again and again, and every one of them is a design gap, not a smarter-model problem.
- Drift. Left ungraded, a loop optimizing for "shorter replies pass the check slightly more often" will quietly write shorter and shorter messages until they say almost nothing. Restate the actual goal in every attempt, not just the last instruction.
- A weak evaluator. "Rate this reply 1 to 10" is not an evaluator, it is a guess with a number attached. Use concrete yes/no checks instead: did it name the order number, did it name the date.
- Runaway cost. "I'll keep an eye on it" is not a budget. A dollar cap or a day limit, decided before you press run, is a budget.
- Watching it live. A loop that takes 40 minutes and you sit there watching is the worst of both worlds, your time and the machine's time. Kick it off, close the laptop, read the log in the morning.
- No real stop. "Stop when it's perfect" on a task where perfect is undefined is the same as no stop at all. If you remember one rule, remember the stall: three attempts with no progress means something is wrong and a human should look.
What You Have After This
A week in, you have one loop running, probably the most annoying repeated task on your plate, the one you have done three or more times this month without ever writing it down. A month in, you have stopped being the person who remembers to follow up, and started being the person who approves the follow-up, which is a very different job. Three months in, you have a small library of loop specs, one for refunds, one for quotes gone quiet, one for late invoices, each one a page long, each one something you could hand to a new hire on their first day, because that is really what a loop spec is.
Same business. Same relationships with your suppliers and customers. A completely different amount of your evening spent remembering things a system could remember for you.
Frequently Asked Questions
What is loop engineering in AI?
Loop engineering is the practice of designing an AI system that acts, checks its own result against a measurable goal, adjusts, and repeats, then stops or escalates on rules you set in advance, rather than running a single prompt and manually judging the output yourself each time.
How is an AI loop different from a Zapier or Make automation?
A workflow tool like Zapier, Make, or n8n is excellent at moving data between apps on a trigger. A loop adds judgment on top of that: it grades its own attempts against a goal, decides whether to try again or stop, and knows which situations need a human. You can build a loop using workflow tools as the plumbing and an AI model as the judgment layer.
How do I know what to let an AI loop do without watching it?
Ask three questions. Can you write down how you would judge the output? Then a loop can check it. Does it involve money, a customer relationship, or a decision you cannot undo? Then it needs a human approval step, always. Have you decided the stopping point before you press run? If not, do that first.
What is a stopping condition and why does every loop need one?
A stopping condition is a rule that ends the loop even if the task is not perfectly finished. The three worth having on every loop are a target (the goal is met), a budget (a time or cost cap), and a stall rule (a set number of attempts with no progress). Without one, a loop will keep running and keep costing you money without ever truly finishing.
Do I need to know how to code to build one of these?
No. You need a written spec, goal, context, evaluator, escalation rule, stopping condition, and a way to run it, which can be a no-code automation tool paired with an AI model, or a simple scheduled task using a tool like Claude with a defined skill. The hard part is the thinking, not the code.
What is loop engineering in AI?
Loop engineering is the practice of designing an AI system that acts, checks its own result against a measurable goal, adjusts, and repeats, then stops or escalates on rules you set in advance, rather than running a single prompt and manually judging the output yourself each time.
How is an AI loop different from a Zapier or Make automation?
A workflow tool like Zapier, Make, or n8n is excellent at moving data between apps on a trigger. A loop adds judgment on top of that: it grades its own attempts against a goal, decides whether to try again or stop, and knows which situations need a human. You can build a loop using workflow tools as the plumbing and an AI model as the judgment layer.
How do I know what to let an AI loop do without watching it?
Ask three questions. Can you write down how you would judge the output? Then a loop can check it. Does it involve money, a customer relationship, or a decision you cannot undo? Then it needs a human approval step, always. Have you decided the stopping point before you press run? If not, do that first.
What is a stopping condition and why does every loop need one?
A stopping condition is a rule that ends the loop even if the task is not perfectly finished. The three worth having on every loop are a target (the goal is met), a budget (a time or cost cap), and a stall rule (a set number of attempts with no progress). Without one, a loop will keep running and keep costing you money without ever truly finishing.
Do I need to know how to code to build one of these?
No. You need a written spec, goal, context, evaluator, escalation rule, stopping condition, and a way to run it, which can be a no-code automation tool paired with an AI model, or a simple scheduled task using a tool like Claude with a defined skill. The hard part is the thinking, not the code.
Find your first high-payback workflow.
See the SprintSources
Find your first high-payback workflow.
Book a free conversation or start with the fixed-fee Sprint.
Keep reading
Stop Telling Your AI to Just Build It. Map It First.
Before you ask an AI agent to build a big automation project, map every open decision first, then resolve them one at a time until nothing is left to guess.
Front-office AIBuild an AI Loop That Finds You Customers While You Sleep
An AI builder named Jason Zhou just published proof that a small, repeatable AI agent loop, post something genuinely useful, track what happens, do more of…
Back-office AIThe Fastest Way to Get Paid Is to Stop Sending the Invoice Yourself
Build a workflow that fires the moment a job is marked complete: it creates the invoice, sends it within seconds, and files a polite AI-drafted reminder if…