Leea Agency
Every company with recurring revenue knows how many customers cancelled last month. Almost none knows in which week each one decided. The monthly number lumps together someone who left on the 2nd with someone who left on the 29th, and someone who complained to support with someone who just went quiet. Leea rebuilds that timeline, customer by customer, and delivers the rule that triggers recovery before the cancellation.
Built with SQL over CRM history, AI agents orchestrated in n8n, and the intervention triggered in RD Station. The method is called Sulco.
The format
This is how Leea stores a customer's history. Each row is a period with a start and an end, plus the state the customer was in. There is no theory here, there are dates.
| cliente_id | estado | inicio | fim |
|---|---|---|---|
| 1001 | trial | 2024-01-05 | 2024-02-04 |
| 1001 | ativo_basic | 2024-02-05 | 2024-07-31 |
| 1001 | ativo_pro | 2024-08-01 | 2024-11-15 |
| 1001 | churned | 2024-11-16 | 2024-12-31 |
Read out loud: customer 1001 trialled for a month, became a paying basic user, upgraded to pro in August, and cancelled on November 16. The monthly report says one thing: churn in November. The table says the customer upgraded 107 days before leaving, and that is where the question lives.
All example data on this page is synthetic, created for demonstration.
The problem
Every month, one number summarizes customer loss. A 3% churn, for instance. That number does not say when inside the month the loss happened, or what changed before it.
One number per month. No time inside the month, no state, no language. The same rate can hide a customer who canceled yesterday and another who had already gone quiet weeks earlier.
Illustrative demonstration data, not a client result.The same aggregate rate can come from very different trajectories. See three illustrative clients in the same window.
The turn
The Trajectory Architecture reconstructs the customer's relationship with the brand over time. The result is a State Map: where the customer is now, and where it can go. Click each state.
Customer with recent behavior consistent with the journey's healthy pattern: frequent interactions, response to stimuli, language of continuity.
Signs of behavioral and language decline appear before formal cancellation: lower frequency, shorter responses, absence of continuity mentions.
No interaction within the window defined by the problem contract. Not confirmed churn yet. It is a state that can reverse or consolidate as a loss.
Customer that resumed the healthy behavior pattern after a period of risk or dormancy. Recovery also has a state. It is not a single point.
Drag the control. Above, the customer changes state across twelve weeks. Below, the agent's rule recalculates every week and shows when it triggers recovery. In week 7 all three criteria are met, and recovery goes out three weeks before the cancellation that was coming.
The rule runs every Monday over the CRM history. Three criteria, all of them must be met.
SE dias_sem_uso >= 14
E mencoes_continuidade_30d = 0
E queda_frequencia >= 40%
ENTAO disparar('recuperacao_pre_churn')The 14 day cutoff is not a guess or an industry default. It comes out of the history itself: it is the point where, in this base, the odds of coming back unprompted fall by half. Each client calibrates their own.
The method
Sulco is the methodology. These are the terms it uses to name each part of the work.
Meaning Reading→ Journey Time Reading→ CRM Decision→ Intervention→ Measurement
Methodology for reading meaning, trajectory and CRM intervention.
Data infrastructure: states, events, exposures, transitions.
Logic for the time based reconstruction of the customer's relationship with the brand.
How the customer moves through the stages of the journey.
Actionable hypothesis about the moment a change in language precedes a change in behavior.
Recurring cycle of testing, measurement, assumption review and model updates.
The trajectory reading draws on dynamic, time based journey theory. The language reading draws on semiotics, with textual analysis.
The operating framework
Each step produces a verifiable artifact. Open each one to see what comes out of it.
Defines the exact retention problem: population, time window, constraints, success criteria.
Gathers the brand corpus (campaigns, emails, surveys, support) and the data contract. Maps voice rules and brand protection limits: what the brand does not allow an AI to generate.
Time based reconstruction of customer data. Structures behavior into continuous episodes, with start date, end date and a censoring indicator: what separates a customer still active at the end of the window from real churn.
Crosses behavioral variation with language change. Formulates the Turning Points.
Specifies the CRM intervention: treatment versus control group, eligibility rules, channels, message. Human approval and controlled execution.
In practice
One slice of the chain, running every Monday morning. The AI agents do not send email. They read, classify and write the specification. The CRM sends, and a person approves before anything goes live.
One query sweeps the CRM history and turns loose events into episode rows, with a start, an end and a state. It is the four row table from above, run over the whole base.
in eventos_crm out episodios.csv
A time trigger runs every Monday at 7am, executes the query, and hands the result to the agent. No intelligence here, just plumbing.
in cron 0 7 * * 1 out payload json
This is where the agents come in, and there are two. leea_sentido reads the text of the customer's last contact and decides whether there is a continuity mention, which is what code does not do well. leea_retem takes that classification, joins it with the numbers and applies the cutoff. The arithmetic itself stays in code, not in the model. The one calling both in the right order is leea_maestro.
in episodios + texto out intervention_spec.json
The specification arrives for review with the recipient list, the proposed copy and the control group already separated. Without that approval, nothing goes out. It is the only block that is not automatic, and that is on purpose.
in intervention_spec out approved or returned
The journey goes live in the CRM with the control group preserved. Four weeks later the same query runs again and compares those who got it with those who did not. If there was no difference, the rule was wrong and the cutoff changes.
in approved journey out measurement_record
Notice where the agents belong and where they do not. They do not decide the 14 day cutoff, that comes from the history. They do not send email, that is the CRM. They do not approve anything on their own. They read text and write specifications, which is where a language model actually beats a spreadsheet. This is the "Act" slice of the chain. The full chain is right below.
How the work gets done
It is not one agent doing everything. It is eight narrow agents, plus an orchestrator that decides the order and holds the human gate. Each one has a differentiator in its configuration, and it is the configuration that makes the language reading worth something in operation.
Leea does not sell an AI agent. It sells the resolution of a marketing and retention problem, with the Sulco method operationalized by specialized agents already built.
Defines analytical intelligence, interpretation of meaning and trajectory reading.
Executes the workflow, organizes evidence and prepares specifications.
Runs the calculations, consolidates history and processes metrics.
Defines what the brand can test, protect and put live.
Decides which agent runs, in what order, and on what data. It holds the cycle state, stops two agents from writing to the same artifact, and halts everything at the human gate before anything goes live.
configures: execution order, approval gates and reprocessing limitsThe planner. Takes a vague request and returns a scoped problem: which population, which transition, which window, and what counts as success.
Differentiator: sets the success criterion before seeing the data, so the metric that looks good afterwards is not the one chosen
The one that cleans. Removes duplicates, fixes broken dates, flags invalid contacts and separates who was still active at the end of the window from who actually left.
Differentiator: censoring rule, which is what stops counting as churn someone whose term simply had not ended yet
The builder. Turns loose events into episodes with a start, an end and a state. It builds the table you saw above, for the whole base.
Differentiator: a state grammar specific to each brand, because "active" does not mean the same thing in subscription and in education
The language and semiotics reading. It reads what the customer wrote in support, in surveys and in campaign replies, and classifies what is a continuity signal and what is an exit signal.
Differentiator: brand lexicon and protection limits, what the brand will not let an AI write in its name
The pattern catcher. Looks at the whole base and finds repeating trajectories, groups that behave alike and the point in time where behavior turns.
Differentiator: pattern capture crossing behavior with language, not just with usage counts
Retention. It is the agent from the panel you dragged. It evaluates the criteria every week and decides when recovery goes out, before the cancellation and not after.
Differentiator: threshold set from the base's own history, not from an industry benchmark
The one that grows. It works the other side of the same reading: where the customer is ready to upgrade, buy again or refer, and when to speak without getting in the way.
Differentiator: opportunity window with a frequency cap, so growth does not cannibalize retention
The one that improves. Compares who received it against the control group, says whether the rule held up, and returns the corrected cutoff to the earlier agents. It is what closes the loop.
Differentiator: recalibrates against a control group and also records what was discarded, not only what worked
What leea_melhora learns goes back into leea_plano and leea_retem. That is why it is a cycle and not a one off delivery: today's cutoff ages, and the base changes behavior on its own.
From orchestration to retention, it is a single chain, and that is what Leea guarantees end to end. A single agent solves one piece and hands over a report. The full chain is what gets the delivery live and keeps it standing afterwards.
The LLM interprets and coordinates. The code calculates. The connector prepares. The human decides and approves irreversible action.
The offer
Each tier deepens the previous one. The pilot is where most brands start.
Manual diagnosis of language and journey, without data instrumentation.
Retention pilot with a delimited scope.
Recurring retainer: a continuous cycle of testing, measurement, assumption review and model updates.
The limits
Traditional semiotic consulting produces deep cultural diagnosis, but stops at the report, with no execution in a revenue channel. The automation and AI agency generates volume of copy and operational flow, but ignores identity, brand coherence and context. Leea sits in between: a reading that becomes measured execution.
Leea distinguishes correlation from causation. Language and behavior signal generates an intervention hypothesis. That hypothesis needs to be validated in a CRM test. There is no deterministic promise.
Next step
The Sulco no Tempo pilot is the entry point. It tests the method on your real retention problem, with a defined scope and a short timeline.