04 — Celeti — Diagnosis

Gabriel Rodrigues
English

Operations · Supercomm

04 / 04

The wrong customer in the right queue.

Celeti’s chatbot was built for questions and feedback from the B2B customer. The conversation data showed a different operation: 37% of interactions were end subscribers trying to cancel or activate Globoplay and Premiere.

Role
Product Manager and UX/UI
Context
Celeti · support and AI cost
Result
−90.9% misdirected tickets

01 — Problem

Almost one in three was not a customer.

In the Globoplay flow, someone trying to cancel an internet plan with streaming landed on Celeti’s support page. It was the platform’s second-largest traffic source, close to 10,000 visits in the period. The Celeti brand also appeared in the activation email and on error pages.

B2B support started answering connection questions and cancellations from end subscribers. The cost was double: operations time, and language-model tokens spent out of scope.

02 — Decision

Acceptance criteria before the screen.

Three conditions closed the problem, before any interface. No mention of Celeti on the end subscriber’s journey. Every one of those people redirected to their own provider. No direct Celeti channel within their reach.

The bet

Deterministic triage before the model. If the message contains “globoplay” or “globo”, a qualification reply sends the person to their provider. The AI does not enter a conversation that is not Celeti’s.

Out of scope

Teaching the chatbot to handle streaming cancellation. That would keep the wrong person in the wrong channel, only with a better answer.

03 — Execution

Take the brand off the path of someone who is not a customer.

  • 01

    AI cost

    Decision trees with fixed conditions in front of the model. Tokens only after qualification. Long, out-of-scope conversations stop being processed.

  • 02

    Support

    The old page received 96.9% of traffic coming from Globo and offered only an FAQ. It became a neutral search: phone, email, and WhatsApp of the ISP on the other side.

  • 03

    Transactional

    The CeletiHub logo left the activation emails. Subject and header now use the name and identity of the contracted provider.

The Celeti assistant flow in Typebot: contact, platform, and integration blocks.
The assistant, with the flow open in Typebot.

04 — Impact

The queue became customers again.

  • 33 → 3monthly tickets from end subscribers on Celeti’s WhatsApp, −90.9% right after deploy
  • 19.5% → 37.6%of people who resolved their need on the support page

End subscribers who gave a CPF on WhatsApp

Monthly series from February to August. The July point marks the 30 July update. 40 30 20 10 0 Feb Mar Apr May Jun Jul Aug 30 Jul

The platform update shipped on 30 July.

End subscribers who gave a CPF on WhatsApp, by month
MonthSubscribers
February35
March27
April39
May38
June24
July25
August4

The same contacts, by service status

Grouped bars from February to August, by Active, Checkout, and Canceled status. 30 20 10 0 Feb Mar Apr May Jun Jul Aug
End subscribers by service status and month
MonthActiveCheckoutCanceled
February17117
March12312
April2658
May21512
June10410
July1537
August103