The engine

Lead sources in. Channels orbiting. Qualified conversations out.

Five lead-source slots flow into the engine on the left. Four channels orbit the core (Chat, SMS, Email, DM). Three qualified outputs flow out on the right — handed off to your team with the full conversation context attached.

Web forms
Paid ads · LSA
Marketplaces
CRM imports
Inbound calls
The engine
MagicBlocks
conversion engine
EngageQualifyFollow upReengage
Across CHAT SMS EMAIL DM
Qualified conversations
Booked appointments
Briefed handoffs

The flow, in three moves

The whole thing, in three moves.

The architecture beneath everything is more involved than three steps. The day-to-day a buyer actually sees is exactly three.

01 · Connect your leads

Plug in where your leads come from.

Web forms, paid ads, marketplace leads, CRM imports, inbound calls. The AI starts working the second a lead arrives. Live in 7–14 days.

02 · MagicBlocks takes over

Every lead gets a real conversation.

Text, email, chat, or call — on whatever channel they reply on. We qualify, nurture, follow up, and only hand off when the lead is ready.

03 · Your team closes deals

Warm handoffs, not cold ones.

Your reps get pre-qualified handoffs with the full conversation context. No cold pitches to people who filled out a form three days ago. No wasted hours on leads that weren't going to close.

The architectural difference

One-prompt AI works for chatbots. It breaks the second a real customer pushes back.

Most AI tools run one set of instructions on a loop. Push it off-script — objection, topic change, three-day gap — and it loses the plot. The task gets almost done, but not finished.

Sales is messy by design. The lead changes direction. Raises objections you didn’t expect. Goes quiet for a week and comes back with a different question. One set of instructions can’t hold all that at once.

Most AI tools (one prompt)
MagicBlocks (multi-prompt)
Instruction stack
One instruction handles everything
A specialised playbook for every stage
Conversation memory
Loses track of what it’s asked
Tracks where the conversation is at every step
Hallucination resistance
Hallucinates more under pressure
Optional second AI can check replies before send
Required-field discipline
Forgets fields under pressure
Doesn’t declare done until the checklist clears
Multi-turn recovery
Falls apart on multi-turn chats
Pulls back on track in 4.55 turns avg
Task completion (n=400)
59% task completion
98% task completion

Tested across 400 simulated "customer from hell" conversations. Z-score = 9.33 (p < 0.00001). The difference is the build — not luck. See how we tested it

Memory across channels

Remembers your leads. The way a great rep would.

Most AI tools log the conversation and move on. MagicBlocks builds a living profile for each lead — the messages, objections, and preferences that matter — and pulls from it on reply.

When a lead comes back two weeks later, we don’t start over. We pick up where they left off. We remember they were worried about pricing. That they prefer text to email. That they mentioned "after the holidays." It’s not a CRM record. It’s relationship memory.

MF

Maya Fernandez

Maya Fernandez — refi inquiry, $380K, mid-700s creditFirst touch · 14 mo ago
!Asked specifically about 30-yr fixed pricing — high original intentDay 0 · web chat
Channel preference: SMS over email. Replied evenings.Day 0–7
After the holidays” — explicit timing signalDay 21
Rate-cycle reactivation triggered when refi rates moved 25bpsToday · 0:00:08
Re-engaged. Sarah Chen briefed: priority, channel pref, holiday lineToday · 0:01:30
Active relationship5 / 5 signals
  • Messages across every channel
  • Objections raised — and how they were resolved
  • Channel, timing, and communication preferences
  • Buying signals and behaviour patterns
  • Qualification state — carried across conversations
  • Everything gathered in conversation

HAPPA · how the engine actually sells

Built on how people actually buy. Not how CRMs think they do.

Every MagicBlocks conversation runs on HAPPA — our sales method, built around how people actually make decisions. It’s not a script. It’s a framework that adjusts as the conversation goes — based on what the lead says, what they need, and where they are.

HHookOpen with something relevant.AAlignMatch their priorities, not yours.PPersonaliseUse everything we already know.PPitchOffer at exactly the right moment.AActionMake the next step easy.

Most sales AI follows a rigid flow: ask the questions, tick the boxes, move on. HAPPA starts with empathy. It Hooks by understanding the lead’s situation. It Aligns with their priorities, not yours. It Personalises using everything it knows. It Pitches only when the moment’s right. And it drives Action by making the next step feel easy.

Dynamic journeys

Not a drip sequence. A follow-up that adjusts.

Traditional automation runs on fixed sequences: email 1 on day 1, email 2 on day 3 — no matter what the lead does in between. MagicBlocks doesn’t run sequences. It runs follow-ups that adapt.

Every lead’s path is different — shaped by what they do, what they reply, when they go quiet. If a lead opens an email but doesn’t reply, the next touch is different than if they hadn't opened it. It doesn’t follow a map. It navigates.

01Lead arrives
02Engage + qualify in conversation
03Read signals + adapt cadence
YesBooked + handed off
Not readyLong-cycle nurture
DormantReactivation queue
  • Branch on what the lead does (opens, clicks, replies, silence)
  • Switch channels based on what they respond on (chat → text → email → DM)
  • Adjust timing based on how they engage
  • Choose the right message for the objection raised
  • Re-enter when a cold lead comes back
  • Trigger off your data (rate changes, inventory, seasonal events)

Guardian · the compliance layer

An optional AI that checks the reply. Before it sends.

Speed and persistence only matter if your messages stay controlled. Guardian is an optional separate AI you can turn on to review replies before they go out. Your rules, your brand voice, factual accuracy, conversation quality — when Guardian is on, each message is checked. If something fails, it’s rewritten automatically or queued for review.

Guardian sits on top of a Guardrails layer you configure — so your team controls exactly how the AI behaves. No human has to remember the rules. The system enforces them.

  • Rules Engine

    Set the exact rules your business runs by. When Guardian is on, replies are checked against them before they send.

  • Rules Monitor

    When Guardian is on, replies that break a rule get caught and rewritten before they send.

  • Jailbreak Prevention

    Blocks attempts to trick the AI into ignoring your rules or going off-script.

  • Moderation

    Catches anything inappropriate or off-policy before it reaches a lead.

  • PII Collection Control

    You decide exactly what data the AI is (and isn’t) allowed to collect.

  • Redaction

    Turn on redaction and the data types you choose are stripped from conversation logs. You decide what counts as sensitive.

  • Automatic FAQ Grounding

    Product and service answers come from your verified knowledge base. Not from the AI guessing.

In our stress test of 400 simulated conversations, a one-prompt agent asked an off-topic question in 100% of runs. The MagicBlocks architecture brought that down to 55% — and limits the worst errors, because it only allows certain questions at each stage of the conversation. See the full Guardrails system

Infrastructure at a glance

Built to run at any volume. Without you thinking about it.

MagicBlocks replies in milliseconds. The system handles spikes without slowing down — whether you’re running 500 leads a month or 50,000. Your data stays where your regulations require: dedicated storage in the US, Europe, and Australia. SOC 2 Type II and ISO 27001 — independently audited, not self-assessed. And your conversations are not used to train our models.

  • Replies in <1s
  • SOC 2 Type II
  • ISO 27001
  • US · EU · AU residency
  • 99.9% uptime SLA
  • Encryption in transit + at rest
  • No training on customer conversations
Sean Clark, Co-Founder & Engineering at MagicBlocks

From the team who built it

If you want AI that actually finishes a sales conversation — not just answers a question and stops — there’s no shortcut around running specialised playbooks for every stage. We built it that way because every other approach we tested broke under real customers. Once you’ve watched a one-prompt system collect every piece of qualifying data and then fail to actually book the call, you don’t go back.

Sean co-built Bob, the mortgage industry’s first AI agent, at Beeline.

FAQ · Technical buyers

Questions technical buyers ask.

Which AI models do you run on?

We’re not locked to one vendor. Each part of the conversation routes to the best model for the job — bigger models for the harder reasoning, faster ones for speed-critical replies — with automatic backup if any model goes down. You aren’t exposed to any one model’s failure.

How is this different from ChatGPT plus a workflow tool?

ChatGPT runs one set of instructions. A workflow tool stitches together rigid paths. MagicBlocks switches between specialised playbooks for every stage of the conversation — each with its own rules, its own context, its own guardrails. That’s why it completes 98% of tasks in stress tests vs 59% for one-prompt systems.

How long does it take to set up?

Most teams are live in 7–14 days from contract signed. The bottleneck is usually content review on your side — rates, policies, FAQs, qualification criteria — not the technical work. Full rollout normally follows a 2–4 week pilot.

Do you use our conversations to train your models?

No. Your data is yours. Conversations are used to improve your AI on your data — nothing goes into shared model training, nothing crosses customer boundaries. The underlying model providers we route to are configured to suppress training on your conversations as well.

How do you stop the AI from making things up?

Three layers. (1) The way it’s built — each part of the conversation only handles what’s valid for that stage. (2) Answers come from your verified data, not from the AI guessing. (3) When you turn it on, an optional Guardian AI can check each reply before it sends and rewrite or queue anything off. In our stress test of 400 simulated conversations, a one-prompt agent asked at least one off-topic question in every run; the MagicBlocks architecture brought that down to 55% of runs — and the slips that remained were mild edge cases, because the structure limits which questions can come up at each stage.

What if your model provider has an outage?

Automatic failover to a backup model. We run hot standby across multiple providers for the critical conversation path. A single provider going dark is a routing event, not a customer-visible outage.

Can we tune the AI to sound like our brand?

Yes. During onboarding we tune the tone — warm, direct, technical, playful — to match how your team speaks. Every reply reads consistently. If you rebrand, we retune; the new voice picks up everywhere.

Can our compliance team audit conversations?

Yes. Every message is logged with timestamp, channel, content, outcome, and the rule-set it ran under. Role-based access lets compliance review anything later. Exports are audit-ready.

How does it handle our qualification criteria?

Set them up during onboarding. BANT, MEDDIC, CHAMP, SPICED, or fully custom. Rules live in your settings — no rewriting needed. As your criteria change, you update the settings and the AI picks it up the same day.

Now you know how it works. See it work.

Watch a real conversation engage, qualify, follow up, and reactivate. On your terms.