[ Case study ]Financial Services

    TTT · AI WhatsApp Tax Agent

    How an AI assistant collects a client's tax documents, returns forms to complete, answers general questions and tax queries, and is grounded in TTT's own knowledge base and live client records.

    • WhatsApp + AI
    • Live in production

    What 10 days in production looked like

    Results
    879
    Cases handled
    72.7%
    Auto-resolved, no human
    42
    Qualified leads surfaced
    ~$0.25
    Cost per case

    When inbound never stops, every reply costs a person's time

    The problem

    Every client-facing business runs the same loop: repetitive questions, after-hours messages, and a handful of conversations that genuinely need a human. Doing that well at scale either burns the team out or quietly loses leads. The status quo is to throw more headcount at it - or to accept the leak.

    An AI assistant that lives where customers already are

    What we built

    A purpose-built assistant inside WhatsApp that handles the bulk of inbound traffic on its own, captures qualified opportunities with full attribution, and hands off cleanly when a person is the right answer. It runs 24/7, scales with volume instead of headcount, and leaves a complete audit trail behind every conversation.

    Inbound
    Quick question…
    Hi, need help
    Quote please
    Tax cert?
    VAT refund?
    Booking time?
    Pricing on returns
    Can you call me?
    Address update
    Need invoice
    WhatsApp here?
    Provisional tax?
    Quick question…
    Hi, need help
    Quote please
    Tax cert?
    VAT refund?
    Booking time?
    Pricing on returns
    Can you call me?
    Address update
    Need invoice
    WhatsApp here?
    Provisional tax?
    24/7 agent enriches CRM & hands off
    CRM
    enriching
    NameS. Mokoena
    Phone+27 •• ••• 4421
    IntentVAT refund
    SourceWhatsApp
    StatusQualified
    Team
    Receives qualified lead
    Inbound · enrich CRM · handoff

    Real conversations, in real time

    Solution in action

    A few real WhatsApp threads, lightly trimmed and with names anonymised. Use the arrows or dots to cycle through.

    IRP5 question
    General tax · self-serve

    How Tina was built

    Behind the build

    Two models, split by task

    Claude Opus handles nuanced client conversations and complex tax guidance. Claude Haiku handles fast admin work - scanning documents, sorting routine requests.

    Built on Claude Code, hosted on Azure

    The platform was built with Claude Code and runs on Microsoft Azure, with incoming WhatsApp messages queued in the cloud so nothing is delayed.

    Grounded, bounded, synced

    Answers are grounded in TTT's approved tax manuals within a safelist of 41 permitted actions. Every conversation, document, and qualified lead syncs to Microsoft Dynamics 365.

    A game-changer for TTT clients

    In the client's words

    "Tina has been a game-changer for our clients at TTT Financial Group. We built her to solve a real problem - clients who had small but important questions were hesitant to reach out. Since launching Tina as our WhatsApp assistant, our clients can get instant answers around the clock, no appointment needed. The technology has allowed our team to focus on the high-value, human work - strategy, growth, and financial planning - while Tina handles the everyday queries. The response has been overwhelmingly positive, and we're proud to offer this kind of accessible, always-on support to every TTT client."

    - TTT Financial Group

    From scoped to live

    Implementation
    Step 1

    Discovery

    Scoped the conversation surface and capability boundary.

    Step 2

    Build

    Trained the assistant on the practice's knowledge, wired in WhatsApp + escalation paths.

    Step 3

    Pilot

    Soft launch on a real customer base, with humans on standby.

    Step 4

    Live

    Full production traffic, including a campaign surge handled on day one.

    Client: TTT Financial Group

    Want one of these answering your inbox?

    If your team spends hours every week on repetitive replies, this is solvable. Let's scope what your version looks like.