MJ explained: what an AI assistant for client conversations actually does

"AI assistant" is a phrase that has become so vague it means nothing. So let me tell you exactly what MJ — the AI assistant inside ClientPulse — does and what she explicitly does not do.

What MJ replies to on her own

MJ handles general client conversation autonomously. That includes:

• The first reply to a new lead, based on the lead's question and your knowledge base.
• Routine follow-up exchanges in an active thread.
• Birthday messages on the day, anniversary messages the day before — in the client's preferred language.
• Review requests when a review date is set on the client.
• Recurring invoices and scheduled messages.

These send automatically, in your voice, grounded in your knowledge base. You can step in any time — take over the thread, edit, redirect — but you do not have to be there for the reply to go out.

When MJ asks first

Some actions are too consequential to fire without you. For these, MJ proposes and waits:

• Sending a summary PDF to a client (e.g. a meeting-recording summary).
• Kicking off a marketing campaign.
• Sending a quote — MJ can build the quote; you click Send.
• Creating an invoice on instruction — MJ asks whether to send it to the client.
• Manual review requests — when no review date is set on the client, MJ asks you to trigger.

The line is not "everything goes through approval". The line is "the things that change the business relationship — money, summaries, marketing — pause for you".

What MJ never does

• MJ never invents a price, a commitment, or a fact about your business. She only uses what is in your Knowledge Base or what you have explicitly told her.
• MJ never claims you "want", "need", or "require" something that you have not explicitly stated. (We added this guard after an incident where another system fabricated an intent on behalf of the operator.)
• MJ never guesses past her confidence. When she is uncertain, she cuts the conversation short and creates a task for you to handle. She does not improvise.
• MJ never escalates a sensitive matter on her own. Compliance complaints, refund requests, legal threats — she flags them and hands off.

How MJ learns your voice

MJ does not "learn from your data" in the way that phrase usually implies. She does not train on your messages and resurface them in someone else's account.

Instead, your voice comes from three concrete inputs:

1. The Knowledge Base — menus, prices, FAQs, policies, service descriptions you have uploaded.
2. Prior conversation context — MJ uses the existing thread with a client as stylistic and factual ground for the next reply in that thread.
3. Per-client language preference — English or Afrikaans, set per client.

That is it. There is no shadow training, no shared model fine-tuning across operators. Each operator's MJ is grounded in that operator's own context.

Why this split exists

The reason MJ handles routine replies autonomously and pauses for big actions is positional, not technical.

Routine replies happen at WhatsApp speed. If you have to approve every "what time do you open" question, MJ is just a typing delay between you and the client — she has not actually saved you anything. So those go.

But sending a summary, kicking off a campaign, issuing an invoice — those carry weight. Those should pause for you to read, edit, and own. So those wait.

That split is the design choice. Most "AI assistants" get it wrong in one of two directions: they either run fully autonomous (and you discover what they did at the worst possible moment) or they gate everything (and the assistant is just an autocomplete you have to baby-sit). MJ runs the first half autonomously and the second half on your nod.

What this means in practice

A typical morning for an operator running ClientPulse looks like this:

• Overnight: 12 client WhatsApps land. MJ replies to all 12 in your voice, books two appointments, answers three pricing questions, and creates one task for the message she could not confidently answer.
• 7am: you open the dashboard. One task waiting (the one she handed back). Three campaign drafts and one meeting summary waiting for your nod from yesterday's meetings.
• 7:05: you handle the task, approve the summary, edit one campaign and approve all three.
• 7:06: you start your real day.

The work gets done while you sleep. The decisions wait for you. That is what an AI assistant for client conversations actually does.