Problem
WhatsApp groups have become one of the main channels for servicing B2B customers, and many companies report that customers prefer it that way. The problem lies in the operation behind it: in each group there are repeated questions, requests for status, and real requests mixed together, and no system separates what a person actually requires. Result: important messages disappear from the screen, when the responsible CS goes on vacation nobody takes over and the team just reacts. The cost increases in line with the number of groups, because the standard solution is to hire more analysts, and yet the team loses signs of risk, which appears as churn and uncaptured expansion. I experience this problem in my work, with [N] groups of clients. Our hypothesis, which we are validating, is that a large part of the volume (repeated doubts and status) can be handled by supervised AI, leaving the human team to do what requires judgment.
Solution
We don't ask the company to abandon the groups, because customers like them. We put an AI layer on top. In each group, the AI classifies the messages into repeated questions, requests for status, real requests, and signs of risk. For what is already answered in the company's documentation, it drafts the response and the CS approves or edits it. What requires judgment, or signals risk, goes to the responsible party with a summary context, and the manager receives a periodic report with what is happening in each account. Each group has a key to autonomy: while the CS is training and checking the AI, it only drafts. When he trusts the level of the answers, he turns on the automatic mode in that group. So the team stops reading everything and losing messages, and the cost of coverage stops growing in proportion to the number of groups. We're starting with a shadow mode test, without sending anything to customers, to measure the accepted response rate before automating.
Business model
Recurring subscription (SaaS) charged per active WhatsApp group, in message volume ranges, with an additional one per group where the customer turns on the automatic reply. We charge per group, not per user, so that our revenue doesn't fall when AI reduces the need for analysts. Price hypothesis: R$ 49 the 99 per group per month in draft mode, plus R$ 30 the 60 per group with automatic response. The expected initial ticket is R$ 1.500 the 4.000 per month (15 the 50 groups), reaching around R$ 8.000 per month when the customer brings about 100 groups. These values have not yet been validated: we are testing with paid pilots and questions about price to 5 the 8 niche companies. With variable AI cost estimated at R$ 4 the 85 per group per month, depending on volume, the gross margin projected in the average scenario is between 70% and 85% (estimate to be confirmed with real data).
Market
There is still no reliable number for our segment (CS teams that serve B2B clients through WhatsApp groups), so we used two references. From top to bottom, the market for customer success platforms is estimated at around US$ 3,4 billions in 2026, with growth above 20% per year, according to a research consultancy (estimates vary between reports). Our segment is a fraction of that market. From the bottom up, we started with B2B SaaS in Brazil: the ABStartups mapping 2025 Record 3.650 active startups, 39% of the SaaS, close to 1.400 companies in this sample alone. With a price per group and conservative adoption assumptions, we estimate an initial market of around R$ 33 millions per year in the base scenario (range of R$ 4 the 140 millions), starting point before expanding to branches, other B2B niches, and Latin America.
Competitors
The market already has tools for parts of the problem, in four groups. (1) Shared inboxes for groups, such as Periskope and TimelineSai, that centralize messages and create tickets, charge primarily per user, and have AI as an additional feature. (2) Conversation intelligence on WhatsApp, such as Lunatic Monkey (Brazil) and Eazybe, with summaries, sentiment, and alerts. (3) Community analytics, such as My Group Metrics, aimed at large communities. (4) Homemade solutions with Evolution API, Chatwoot and n8n, and traditional helpdesks, which mainly cover conversations 1:1. Our bet differs in three points that we are still validating: AI that drafts and, per group and with the release of the CS, begins to respond on its own; price per group instead of per user, in line with reducing the need for analysts; and a focus on B2B SaaS CS teams in Brazil.
Competitive differentiation
We don't see a single differential, but rather a combination that we are still validating. First, gradual and measurable autonomy: AI starts only by drafting, and each group has an automation key that CS links based on an accepted response rate that we measure, instead of relying on AI in the dark. Second, price per group instead of per user: if AI works and the team needs fewer analysts, our revenue doesn't fall together, unlike the tools charged per seat. Third, focus on B2B SaaS CS teams, with a founder who experiences the problem on a daily basis. In the long term, the history per account and the accumulated risk signals increase the exchange cost. We recognize that competitors like Periskope already offer AI agents, so our initial advantage is focus and execution, not exclusive technology.
Entry barrier
Today the entry barrier is low, and we recognize this: access to AI models and the connection to WhatsApp are available to any technical team. The barrier we are building is one of exchange and data costs. Each group accumulates the configured knowledge base, conversation history, CS rules, and memory of the corrections it makes to drafts. That memory, along with the measured acceptance rate, is what gives you confidence to turn on the automatic response in that group. Migrating to another tool means restarting this training and losing the history. With scale, the set of CS B2B conversations in Portuguese, anonymized and treated according to the LGPD, would allow us to build references for signs of risk of churn that a new entrant does not have. We don't have that advantage yet: it only develops with clients and with time.
Traction
Not yet. The product was designed in September of 2026 and it's under validation, with no revenue or paying customers. Today we have signs of demand: a founder who experiences the problem on a daily basis, entrepreneurs from our network who report the same pain, and two partners who can provide authorized groups for testing. In the coming weeks, we will be running a test in shadow mode and are looking for paid drivers.