01 Person, request, response, done — and tomorrow, it starts all over

In most businesses today, AI works in a straight line. Someone writes a request — people in the field call it a prompt —, the AI returns an answer, the person adjusts what is needed, and sends it. Done. Next time starts from the same point: a new request, a blank page, no memory of the last one.

Regina owns an aesthetics clinic and sends about 200 proposals a month over WhatsApp. For each one, she writes something like “draft a skincare package proposal for the customer who asked about the price.” The proposal comes back, she changes two sentences, and sends it. The customer replies or disappears. Proposal 201 starts with no knowledge of the previous 200 — neither which got a reply nor which led to a sale.

The AI did not do a bad job. Every proposal was good. The problem is different: the path has four stops and no way back.

02 AI remembers nothing from one request to the next — and that is not a flaw

An AI chat is like an excellent temporary assistant who arrives new for every shift: it writes very well and knows the subject, but it was not here yesterday. It did not see whether the customer replied, the lead booked a visit, or the support ticket came back. Everything that happens afterward the response stays outside the conversation.

Marcos, a real estate agent, gets about 40 leads a week through a listing portal — a lead is someone who asked for information but has not decided yet. He replies to each one with AI’s help. When a lead responds well because Marcos mentioned the balcony’s square footage, that teaches something — but only Marcos, if he notices. To the AI, the next lead is the first one it has ever seen.

That is why “using it a lot” does not become “learning.” More use changes nothing until someone records the result and carries it into the next request. The rest of this course builds that feedback loop.

03 Real improvement takes six things — and none of them is “effort”

When someone says “I want AI to improve,” they usually mean writing a better request. But real improvement does not come from the request. It comes from six ingredients connected in a loop: evidence (what happened, written down), metric (a number that says whether it worked), memory (where past attempts are kept), hypothesis (an idea for a change, with a reason), experiment (the idea tested against the current version), and evaluation (someone determines which one won, based on the number).

Paulo manages support at a small software company: 10,000 tickets a year. For his team, the six ingredients look like this. Evidence: each ticket has its topic recorded. Metric: resolved on first contact, yes or no. Memory: a place where the team records what it has already tried in its canned replies. Hypothesis: “if the password reply includes a direct link, fewer people will come back.” Experiment: half of password tickets get the new reply, half the old one. Evaluation: the count of tickets that “came back within 7 days” determines the winner, not the opinion of the person who wrote the reply.

Notice: none of the six means “ask more forcefully.” If one is missing, the loop does not close — a hypothesis without an experiment is a guess; an experiment without memory is a discovery lost the next week.

04 What a loop guarantees — and what it does not promise

Before we continue, here is the anti-promise, stated in the very first lesson so it cannot be hidden: no improvement loop guarantees that the number will rise. Anyone promising that is selling luck dressed up as a method. A well-built loop guarantees something else, more modest and more valuable:

Notice that all four guarantees are about process, not results. That may sound modest. It is exactly what no promise of “AI that improves itself” can honestly deliver — and it lets you sleep while the cycle runs: nothing enters without evidence, and anything that entered can be removed.

In the LOOP-R reference example — a fictional clinic like Regina’s, using simulated data — the proposal response rate rose from 19.3% to 29.3% in a test with 300 sends per version. Sales conversion did not move: 3.7% versus 3.0%, a difference that could be chance. The loop did not hide that. It recorded it, and the person made the decision with the numbers in front of them. For Regina, that is what matters: not a promised number, but one she can verify.

01 Ten thousand tickets, one issue in three—and nobody noticed

Paulo, support manager at a small software company, ends the year with 10,000 tickets handled. Each one was resolved well, one at a time, with AI helping write the reply. The team gets praised. And nobody knows that 37% of those tickets—3,700—are about the same issue: a password screen that confuses everyone.

Why did nobody notice? Because each ticket is a request that ends right there. From inside one ticket, you can’t see the pile. Only someone looking at the whole set—and, for that, needing a “what was it about?” column filled in for each one—can see that one in three points to the same thing.

Here’s the difference between “we handled a lot” and “we learned a lot.” The 10,000 resolved tickets left 10,000 customers served and zero product changes. A single fix to that screen would have prevented 3,700 conversations. The information was there all year; it just wasn’t written down.

02 Starting over has a cost that never appears on the invoice

Marcos, the real estate agent, replies to about 40 leads a week—around 160 a month. Each reply takes about six minutes to request, adjust, and send. That’s 16 hours a month. He doesn’t think it’s expensive because AI “does it fast.” But next month’s 16 hours will buy exactly the same thing as this month’s: 160 isolated replies, none smarter than the last.

Compare that with another scenario using the same amount of time: the same 160 replies, but each leaves a row—date, property type, opening used, whether the lead replied, whether they booked a visit. The effort per reply is the same. The difference is the trail. In one case, all that’s left is fatigue; in the other, there’s material to answer a question Marcos can’t answer today: “Which opening gets more people to book?”

The cost of starting over isn’t the time—you’d spend that time anyway. It’s the learning that passed through your hands and didn’t stay.

03 Two companies use the same AI: the difference is who learns faster

For years, a company’s advantage was “having more data than the competition.” That changed. Today, everyone has access to the same AI, with the same capabilities. If two real estate agencies use the same chat, the same listing sites, and serve the same neighborhood, what sets them apart? It isn’t a better-written prompt—you can copy one in a day.

What can’t be copied is the history. One agency records the channel, property type, reply sent, and outcome for each lead. After six months, it has about 1,000 rows to consult before changing anything. The other has 1,000 conversations lost in each agent’s WhatsApp. A competitor can copy the text; they can’t copy six months of “what we’ve already tested and how it went.”

That accumulated history is the asset. It doesn’t come from better technology—it comes from one row per execution, every day, without exception.

04 Where evidence begins: one row per execution

The whole loop depends on one humble step: recording. The place where it happens is the tracking spreadsheet —a regular spreadsheet, with one row for every proposal, reply, or ticket, and a column most people forget: what happened next. Without that column, you have a file of text, not evidence.

For Paulo’s support team, a row looks like this: date · ticket topic · which saved reply was used · resolved on first contact (yes/no) · came back within 7 days (yes/no). Five columns. The question that used to take a week of reading (“what do people complain about most?”) becomes a two-minute count.

One honest point about the method: the evidence source is the bottleneck, not the AI. If the data live in a salesperson’s head or are scattered across conversations, the loop has nothing to read—and a loop without evidence is theater. One caution from the start: 12 rows don’t make a pattern. “Tuesday’s proposals get more replies,” based on 12 proposals, is chance dressed up as a discovery. The loop’s rule is to claim a pattern only with 30 or more items per group.

01 Level 1 — Responds: you ask, it delivers, you check

Level 1 is where almost every business is, and there’s nothing wrong with that. It’s AI answering a request. You write, it delivers text, you check and use it. All the process intelligence lives in your head: what to ask, what to adjust, what to send.

Regina, at the clinic, is a pure level 1. The signs are easy to spot: each use starts with her writing; the result of each proposal lives in her memory, nowhere else; if she goes on vacation, the AI “forgets” the clinic—because it never knew anything beyond that day’s request.

Level 1 isn’t bad. Getting a good answer a thousand times already saves many hours. Just don’t confuse that with learning: a thousand good answers are still a thousand first times.

02 Level 2 — Executes a goal: the single-purpose assistant

At level 2 comes what practitioners call an agent—in this course, a single-purpose assistant. You no longer dictate each step; you give it a goal. It plans, uses a tool (a calendar, a spreadsheet, WhatsApp), and delivers the result.

Marcos, the real estate agent, has one: the assistant reads the lead that came through the listing site, finds the property in his spreadsheet, writes a reply, and suggests two times for a viewing. Marcos just checks and approves. It does more than reply—it executes a goal from start to finish.

Here’s the catch: the assistant does more things and learns exactly as much as at level 1—nothing. Nobody records whether the viewing happened. Level 2 expands reach; it doesn’t close the loop.

03 Level 3 — The loop: execute, measure, evaluate, improve, execute again

Level 3 is the first one where the company learns. Each execution becomes a row; every week, someone (or an assistant with that role) reviews the rows, proposes a hypothesis, tests the new version against the current one, and decides based on the numbers. If it wins with proof, it becomes the official version. If not, the result is recorded so no one unknowingly tries it again.

In LOOP-R’s reference example (a fictional clinic with simulated data), the loop ran four cycles between February and July. One hypothesis—“open by mentioning the review the client left”—was tested with 300 proposals on each side: response rate went from 19.3% to 29.3%, without worsening anything being monitored (complaints, discount); it was approved on the card and promoted to the official version. Another, “up to 80 words,” rose by almost as much and was discarded by rule. Both remain in memory.

Within level 3 there are three degrees of autonomy. L0the loop only suggests; you do everything. L1you approve every change on a five-line card, at most once a week. L2it promotes on its own, alerts you, and keeps the rollback button—and unlocks only after you correctly approve five changes at L1. This course sets yours up at L1.

04 Levels 4 and 5—and why your only next step is level 3

Level 4 is the system that observes its own improvement system: the cost of each cycle, how many hypotheses became versions, which assistant gets vetoed most. It sounds like magic, and honesty matters here: today, in the product, this level only reports. It doesn’t redesign anything on its own. A system that changes itself with little data multiplies every error from the levels below—which is why autonomous redesign remains research, not a product.

Level 5 is the whole company: every department has its loop, with another loop above them connecting the lessons. Paulo’s support team shows the path: tickets logged → cause found (the password screen) → new reply tested against the old one → help text updated → product change suggested. Support stops just answering and starts improving the product. But that only works because, underneath, every ticket became a row.

That leads to the rule: you don’t skip a step. Someone at level 1 who buys “level 5” is buying theater, because the loop above has no evidence to learn from. From levels 1 and 2, the only useful next step is level 3—and level 3 starts with one more column in the spreadsheet and a test with numbers.

01 “Analyze your work and keep getting better” isn’t a system—it’s a sentence

Paulo from support wrote this in the assistant that replies to tickets: “analyze your replies and keep improving.” A month later, he asked whether it had improved. The assistant confidently said yes. There was no way to know if that was true—and that’s what makes the sentence dangerous: it creates the feeling of improvement without producing any.

Three questions the sentence doesn’t answer are missing. Better than what (which version?). Measured by whom (the assistant itself, which has every reason to say it improved?). Compared how (against what, and using how many cases?). Without recorded evidence, a number, and a previous version, “improved” is just the expected answer to the question you asked.

This isn’t an AI limitation; it’s a request limitation. The course’s anchor sentence exists to replace exactly this habit:

02 A goal needs a number, a direction, and “without worsening”

The first change is to replace “improve” with “from X to Y, without worsening Z”. For Regina: “proposal response rate from 20% to 30%, without worsening complaints or average discount.” For Marcos: “viewing booking rate from 20% to 30%, without worsening no-show visits.” The “without worsening” part has a name— guardrail —and it’s not decoration.

Without a guardrail, every goal becomes a target on its own, and the assistant learns to cheat: ask for “more clicks” and it writes clickbait; ask for “less time per ticket” and Paulo’s team starts closing tickets quickly without resolving them; ask for “more replies” and messages start provoking “stop sending me this”—which counts as a reply. A goal without “without worsening” is this track’s quietest mistake, because the number goes up while the business gets worse.

Two cautions about the number. First, choose a metric customers respond to quickly—reply rate, bookings—and monitor a money metric (conversion, margin) as a guardrail. Second, goals near 3% are almost impossible to prove in a small business: proving 3% → 5% takes about 1,500 executions on each side, which at 200 a month takes 15 months. Proving 20% → 30% takes ~300 on each side. Start where proof is possible.

03 A hypothesis is a sentence with IF, THEN, and BECAUSE—and only one change

The second change is to replace “idea” with a hypothesis. A hypothesis has a fixed form: IF [one change] THEN [expected effect on the goal] BECAUSE [reason]. In LOOP-R’s reference example, the fictional clinic wrote: “IF the proposal is up to 80 words, THEN the reply rate rises from 19% to about 30%, BECAUSE long WhatsApp proposals aren’t read to the end. Change: length only.”

The hardest part is the last one: one change only. Marcos wants to “cut the text, change the tone, and send it in the morning”—three changes at once. If it works, he won’t know which one worked; if it fails, he won’t know which one hurt. Three ideas are three hypotheses, tested one at a time or recorded for later.

For Regina, the same form applies: “IF the proposal opens by mentioning the review the client left, THEN more clients reply, BECAUSE they recognize the message was written for them”—that was, in fact, the hypothesis that was promoted in the example. A good hypothesis also protects you: if it fails, you know exactly what to discard and what to record so nobody unknowingly tests it again.

Notice what happened to Paulo’s vague instruction after these two changes:

04 The person who decides whether it worked isn’t the person who came up with the idea

The third change is to separate the proposer from the judge. The judge can only answer one of three things: B won by a clear margin, A stays, or not enough data—keep testing. The third answer is the most common and must be acceptable, because the alternative is promoting chance. That “300 on each side” has a name— sample size that proves it —and it’s calculated before the test, not after.

This is where the second quiet mistake lives: stopping a test halfway because the new version is “in the lead.” Marcos, with 40 leads a week, needs about 15 weeks to reach 300 per version. In week 2, the new version has 12 replies out of 40 versus 7 out of 40 for the old one, and he wants to promote it. Twelve versus seven out of 40 is exactly the kind of difference that appears and disappears on its own. Stop early for one reason only: a guardrail is falling. Never because of a partial lead.

05 “Looks better” isn’t “is better”: the verdict comes from the number and the guardrails

Back to the reference example. In cycle two, the “up to 80 words” hypothesis was tested with 300 proposals on each side: replies rose from 19.3% to 29.7%—ten points, with a proven sample size. And it wasn’t promoted. The “complaints” guardrail showed 6 versus 1, and the agreed tolerance was zero. The rule was applied as written; the version stayed out; everything was recorded.

Then comes the second lesson, as important as the first: the rule was poorly calibrated. Six complaints versus one out of 300 is noise—it can happen by chance. Zero tolerance for a rare event rejects good ideas because of bad luck. The team adjusted the tolerance to 1% in later tests, and the rejected hypothesis stayed in memory with the reason. In cycle four, “open by mentioning the client’s review” rose from 19.3% to 29.3%, stayed within the guardrails, and became the official version. Sales conversion didn’t move—and that was recorded too.

The same rule applies to Paulo: if “resolved on first contact” rises but “reopened within 7 days” rises too, A stays. The system doesn’t promote something because it looks better—and at L1, you’re the one who approves it on the card, with the numbers in front of you.