Building a Custom AI Sales Platform: Agents That Sell — With Human-in-the-Loop and Full Visibility

TL;DR

A client wanted an AI agent that could sell on its own — across the channels their customers already use — without turning into a black box they couldn't see into or step into. This is what custom AI agent development services look like in practice: we built the whole platform from scratch, a multi-agent system that reads a live catalog, builds carts, generates hosted payment links, and hands off to a human the instant it matters.

This is the story of how it came together — from a single idea on a discovery call to a working product the client now runs their sales on.

The one-line version: A multi-agent system that sells across three channels, lets a human take over any conversation at any moment, and surfaces every message in one inbox so the business always knows what's happening.

See the live demo →

The problem: nobody wants an AI they can't see

Most businesses don't hesitate on chat automation because they doubt the AI can hold a conversation. They hesitate because they're afraid of what they can't see. Hand your customer conversations to a bot and you risk losing the one thing that matters — knowing what was said, what got sold, and where deals quietly fell apart.

Our client felt exactly that. The brief wasn't just "a bot that replies." It was:

An agent that sells across WhatsApp, Messenger, and Instagram.
That a human can take over at any moment.
That gives the business a complete, message-by-message view of everything it does.

We designed the platform around that requirement first.

What we built

A single AI agent that handles a real buying conversation end to end, across every channel. A customer messages — say, asking about hoodies — and the agent:

Reads the live catalog and replies with the product, sizes, and prices — not a canned menu.
Answers follow-up questions naturally ("which colors does it come in?").
Builds the cart on confirmation — "Added one Everyday Hoodie in medium to your cart."
Generates a real hosted payment link when the order is confirmed.

A live WhatsApp buying conversation — the agent reads the catalog, answers questions, and builds the cart in real time.

Then the moment the client cared about most: when a customer asks to "talk to a real person about a bulk order," the agent doesn't guess or stall. It hands off — "I've shared your request with our team; a teammate will join you shortly" — and steps aside so a human can join the same thread. No dropped sale. No bot loop.

Here's the whole system in one picture:

flowchart LR
    C[Customer]
    WA[WhatsApp]
    MS[Messenger]
    IG[Instagram]
    R[AI Agent<br/>Router]
    CAT[Live Catalog]
    PAY[Hosted Payment Link]
    H[Human Takeover]
    INBOX[Unified Inbox<br/>+ Tracking]

    C --> WA --> R
    C --> MS --> R
    C --> IG --> R
    R --> CAT
    R --> PAY
    R -.escalate.-> H
    R --> INBOX
    H --> INBOX

Human-in-the-loop, built into the platform

The handoff isn't a bolt-on. Takeover is a core part of the platform: the client's team can step into any conversation at any moment, and the agent gets out of the way the instant they do. AI handles the volume; a human handles the moments that matter — the bulk order, the upset customer, the deal worth a personal touch.

Human-in-the-loop in action — a teammate takes over the live thread, and the agent steps aside automatically.

A deterministic safety net sits behind the agent's own judgment, so escalations never slip through even when the model is unsure. The customer experience stays in one continuous thread the whole time, whether they're talking to the agent or a person.

flowchart TB
    M[Incoming message]
    AI{Agent decides:<br/>can I handle this?}
    NET{Deterministic<br/>backstop triggered?}
    REPLY[Agent replies<br/>catalog · cart · payment]
    ESC[Escalate]
    HUMAN[Human joins<br/>same thread]
    PAUSE[Agent steps aside]

    M --> AI
    AI -->|Yes| REPLY
    AI -->|Asks for a human| ESC
    REPLY --> NET
    NET -->|Yes| ESC
    NET -->|No| REPLY
    ESC --> HUMAN --> PAUSE

Full visibility: the part that closed the deal

Behind every message is a workflow we built and control. Each conversation is captured and surfaced in one inbox, tagged by channel — WhatsApp, Messenger, Instagram — and by what it's doing. That gives the business owner answers to the questions that actually matter:

What are customers asking for?
What is the agent selling, and at what price?
Where does it hand off to a human — and how often?
How does each conversation end: sale, handoff, or drop-off?

The same agent on a Facebook Messenger thread — one platform, every channel, all surfaced in one place.

Every conversation resolves into one of three outcomes, and every one is logged:

flowchart LR
    START[Conversation] --> AGENT[AI Agent handles]
    AGENT --> SALE([Sale ✓<br/>payment link sent])
    AGENT --> HANDOFF([Handoff<br/>human takes over])
    AGENT --> DROP([Drop-off<br/>flagged for follow-up])
    SALE --> TRACK[Tracked in inbox]
    HANDOFF --> TRACK
    DROP --> TRACK

Selling is the easy part. Knowing when not to — and being able to prove the agent is behaving — is the hard part. That's what we built the platform around.

How it works under the hood

The agent isn't one giant prompt. A router reads each incoming message and dispatches it to a specialized sub-agent — catalog lookups, payment and checkout, or general questions — so each part of the conversation is handled by something purpose-built.

flowchart TB
    MSG[Incoming message]
    ROUTER[Router agent<br/>classifies intent]
    CATALOG[Catalog sub-agent<br/>live products · sizes · prices]
    CHECKOUT[Checkout sub-agent<br/>cart · hosted payment link]
    GENERAL[General sub-agent<br/>questions · smalltalk]
    HANDOFF[Human-handoff layer<br/>+ deterministic backstop]
    WF[(Workflow + tracking layer)]

    MSG --> ROUTER
    ROUTER --> CATALOG
    ROUTER --> CHECKOUT
    ROUTER --> GENERAL
    ROUTER --> HANDOFF
    CATALOG --> WF
    CHECKOUT --> WF
    GENERAL --> WF
    HANDOFF --> WF

The catalog is read live from the store, so prices and stock are never stale.
A confirmed cart becomes a hosted payment link through the client's own payment provider.
The human-handoff layer lets the agent escalate cleanly, with a deterministic backstop.

Every step runs through the platform's workflow layer — which is also where the conversation tracking lives. The same plumbing that makes the agent work is what makes it transparent.


Channels	WhatsApp · Messenger · Instagram
Capabilities	Live catalog · Cart building · Hosted payment links · Human-in-the-loop takeover · Full conversation tracking
Approach	Multi-agent routing · Workflow-driven · Custom-built from scratch around the client's idea

From one idea to a full working platform

This started as a single idea on a call. We took it through discovery, design, and build, and shipped it as a complete, working platform the client runs their sales on.

That's what we do. OffLift is a digital product studio — we build custom web apps, mobile apps (iOS & Android), and AI-powered products from first idea to production. No templates, no off-the-shelf. We build it around your idea and how your business actually works.

Want to build something like this?

Got an idea for an AI agent or platform that fits how your business actually runs? That's exactly what we build.

Book a free 30-minute discovery call → — no pitch, just a real conversation. We'll come prepared with ideas.

You can also reach us by email at [email protected].

Keep up with what we're building

We ship new automation case studies, deep-dives, and demos almost every week:

LinkedIn — linkedin.com/company/offlift
Blog — offlift.com/blog
Email — [email protected]

About OffLift — We're a digital product studio that designs and ships AI-powered automation, mobile apps, and web platforms from first idea to production. Related reading: How AI Chat Automation Works (the step-by-step process), Sentira (AI mental wellness), God Encounters (full-stack platform with AI moderation).