The Enginy
Intelligence Engine

From Tools to Autonomous AI Agent

Enginy graphic

ACT I

The
Diagnosis

We Have Two Problems That Are Killing Our Growth

Problem 1: Zero Stickiness

Users don't feel locked into Enginy. They can leave without feeling like they're losing anything.

  • Nothing accumulates inside our platform
  • No intelligence unique to their account
  • Competitor launches cheaper AI → they switch overnight

We're disposable.

Problem 2: Churn from "AI Expectations" & "Ease of Use"

Users sign up expecting AI to do the work. Instead, they get a toolkit.

  • They have to write prompts
  • They have to configure variables
  • They have to engineer their own solutions

Selling "access to AI" when we should sell "AI that works for you."

Problem 1: Zero Stickiness

When customers churn, they take everything with them.

  • Their prompts
  • Their workflows
  • Their learnings

Nothing accumulates that becomes more valuable over time.

Users aren't coming back daily. They're not building dependency. They're not feeling actual loss when they consider churning.

Problem 2: Churn from "AI Expectations" and "Ease of Use"

AI Expectations

Users expect AI to do the work. Instead, they get powerful tools that require them to:

  • Write prompts
  • Configure variables
  • Engineer their own solutions

The AI doesn't feel easy. It feels like a complicated differentiator.

Ease of Use

Our AI Variables feature tries to be everything to everyone:

  • Mixes prospecting with outreach
  • One confusing interface
  • Non-technical users struggle

Steep learning curve → most users bounce before they figure it out.

The Growth Imperative

To hit our growth targets, we need to be the easiest and most AI-powered platform in the market.

If we don't fix stickiness

Growth becomes a leaky bucket. We spend to acquire users who leave before they see compounding value.

If we don't fix ease of use

We keep churning users who signed up expecting magic and got homework instead.

The Path Forward

Stop making users work to get value from AI

Make the AI work for them

Build intelligence that compounds with usage so they can't imagine leaving

ACT II

The
Guiding Policy

Three Strategic Pillars

1. Best Possible Quality

Expert-level output should be the default, not something users engineer themselves.

Trading user control for guaranteed results.

2. Compound Intelligence

Every interaction makes the system smarter for that specific user.

Month 1: Good. Month 6: Knows your voice and what converts.

3. AI-First Experience

Users see value from the first session. Agents do the work, not tools.

Vibe selling, not prompt engineering.

Pillar 1: Best Possible Quality

We're prioritizing output quality over user flexibility.

What this means

  • Taking ownership of quality
  • Opinionated defaults over infinite customization
  • If AI produces mediocre output, that's our failure

Key Actions

  • AI evaluations for all default prompts
  • Split AI Variables into AI Datafields & AI Texts
  • Close default prompts to users
  • Reasoning for each sentence

Pillar 2: Compound Intelligence

Every interaction, every campaign, every piece of feedback makes the system smarter for that specific user.

Month 1

The AI is good

→

Month 6

It's learned your voice and knows what converts

That's not something you can replicate by switching platforms.

Pillar 3: AI-First Experience

First session matters

Users see value before struggling with setup

Vibe selling, not prompt engineering

Describe what you want in natural language

Agents, not tools

Give them agents that do the work

Preview and iterate

More like directing a team member, less like configuring software

Make AI feel intelligent and autonomous, not like a complicated text field.

ACT III

The AI
Outreach Agent

The Brain

The Brain: From Pipeline to Autonomous Agent

Current: Pipeline

Context in, message out

→

Future: Autonomous Agent

Makes three core decisions

WHO

Lead discovery and dynamic prioritization

WHEN

Timing based on real-time signals

HOW

Campaign structure, channels, messages

The Evolution

Before

Manual list, manual campaign, manual messages, manual replies

Now

Upload list, AI enriches data, AI generates messages, AI drafts replies

AI only touches the copy

Future

User describes ICP once

Agent handles everything else

  • Finding leads
  • Ranking by urgency
  • Picking channels
  • Writing messages
  • Sequencing steps
  • Continuous re-prioritization

Dynamic Lead Prioritization

The Core Behavioral Shift

From: Upload a list, send to all

To: Living, prioritized queue that reorders in real time

How it works

  1. ICP-driven sourcing: Agent continuously discovers matching leads
  2. Signal-based scoring: Real-time priority based on intent signals
  3. Dynamic reordering: Queue reorders based on incoming signals
  4. Channel & message per lead: Agent picks the right channel and crafts the message

Dynamic Prioritization Example

Yesterday: Lead #131 was buried at position #131

Today: They switched jobs and posted on LinkedIn about scaling outbound

Result: Agent detects signal and moves them to #1

Agent Decision:

  • Lead: Sarah Chen, new VP Sales at Acme
  • Score: 94 (was 32)
  • Signal: Job change (3 days) + LinkedIn post about outbound challenges
  • Channel: LinkedIn DM
  • Message angle: Congratulate on new role, reference scaling challenges
  • Action: Process now instead of waiting

ICP & Buyer Persona: Structured Targeting

Current Problem

One combined ICP mixes company-level and contact-level attributes

  • No multi-persona targeting
  • No reusable personas
  • No persona-specific messaging

Solution: Split It

ICP (Company Level)

Industry, Size, Revenue, Location, Tech Stack

Many-to-Many

Buyer Persona (Contact Level)

Job Title, Context, Pain Points, KPIs, Intent Signals, Messaging Framework

Buyer Persona: Much Richer Understanding

Migrated from ICP

  • Job Title
  • Key Pain Points (now extended)
  • Additional Information
  • Disqualifying Factors

New Properties

  • Typical Context of Position: What drives their decisions
  • Objective of the Position: Measurable goals they're hired to achieve
  • Extended Key Pain Points: Categorized (frequent vs. secondary)
  • KPIs That Matter: Specific metrics they track
  • Intent Signals: Events that create buying windows
  • Value Justification Approaches: Specific angles per persona
  • Messaging Framework: How to frame messages (ATL vs. BTL)

Messaging Framework: ATL vs. BTL

ATL (Above-the-Line)

VP/C-level executives

  • Think: Past and future (annual goals, risks, budget)
  • Need to hear: Validation and outcomes (pipeline, conversion, risk)
  • Signals: "Does this impact the number?", "What is the risk?"
  • Framing: Validation: outcome / risk / speed

BTL (Below-the-Line)

Managers and operators

  • Think: Present (what is broken today, process friction)
  • Need to hear: Solution to concrete day-to-day problems
  • Signals: "This wastes our time", "the data is bad"
  • Framing: Problem: workflow / friction / time

Same account, same insight, but two ways of framing it.

Conversational Feedback Loop

How the agent goes from "educated guessing" to fine-tuned precision

Learning Accuracy Over Time

Educated Guess

(First campaign)

40% accuracy

  • Broad ICP
  • Generic messages
  • Default channels
  • Static timing
→

Refined Model

(After feedback)

60% accuracy

  • Tighter ICP
  • Adapted tone
  • Preferred channels
  • Adjusted timing
→

Fine-Tuned Agent

(After weeks of use)

90% accuracy

  • Precise targeting
  • Personalized per lead
  • Optimal channel per signal
  • Dynamic timing per lead

Reasoning and Control: The Core UX Challenge

The biggest challenge: trust

Users need to understand why and have meaningful control.

The Solution: Propose-Then-Execute Loop

Inspired by agentic coding tools like Claude Code

  1. Agent proposes action
  2. Explains reasoning
  3. User can approve, reject, or redirect
  4. Agent executes

Levels of Autonomy

Level 1: Full Review

Agent proposes every action; user approves each one manually

Level 2: Review High-Impact

Agent auto-sends standard actions; pauses for high-priority leads or unusual messages

Level 3: Review Exceptions

Agent runs autonomously; pauses only when confidence is low

Level 4: Full Autonomy

Agent runs without interruption; user reviews via activity log

Most users start at Level 1-2 and graduate to Level 3-4 as they build trust.

The Action Log: Full Transparency

10:32 AM

ACTION: Sent LinkedIn DM to Sarah Chen (VP Sales, Acme Corp)

CHANNEL: LinkedIn | PRIORITY: #1 (Score: 94)

WHY THIS LEAD NOW:

  • Changed jobs 3 days ago (new VP Sales role) — high intent window
  • Company matches ICP: Series A, B2B SaaS, 45 employees, UK
  • Active on LinkedIn (2 posts this week about scaling outbound)
  • Similar leads converted at 67% via LinkedIn DMs

WHY THIS CHANNEL:

  • LinkedIn preferred: lead is highly active on the platform
  • Historical data: LinkedIn DMs outperform email 2.3x for job-change signals

Why the Reasoning Layer Is Critical

Without it

  • Users won't trust the agent with their brand
  • Users can't learn from the agent's decisions
  • Users can't catch mistakes before they damage relationships
  • Feedback loop breaks

With it

  • Users understand the agent's logic and can shape it
  • The agent becomes a teaching tool
  • Trust builds incrementally
  • Mistakes are caught and corrected

The reasoning layer is not a nice-to-have; it's the core unlock that makes the whole agent model work.

ACT IV

Additional
Capabilities

Reply Agent: Goal-Driven Reply Handling

Current

AI drafts a generic reply. User edits and sends.

Future

Agent classifies intent, picks strategy, handles reply autonomously.

Reply Strategies

  • Interested → Meeting: Push for meeting or demo
  • Objection: Address specific objection, send resources
  • Bad Timing: Switch to nurture campaign, schedule re-engagement
  • Out of Office: Pause sequence, resume after return date
  • Wrong Person: Ask for referral to the right contact

AI Conversation Tagging: Turning Replies into Intelligence

Current

Tags are a UI convenience - colored labels in the Inbox

Future

Tags become structured signal data that feeds into:

  • Reply Agent decision making
  • Feedback loops (learning what works)
  • ICP Discovery Layer (understanding audience fit)
  • Success pattern recognition

Tag Taxonomy Evolution

Intent Signals

Interested, Requesting Info, Buying Signal

Objections

Timing, Price, Competitor, Authority, Need

Outcomes

Demo Scheduled, Meeting Booked, Trial Started, Deal Closed

Disqualifiers

Wrong Person, Wrong Company Size, Already Using Competitor

Timing

Bad Timing: Short-term (<3 months), Long-term (>3 months)

Sentiment

Enthusiastic, Open, Lukewarm, Dismissive, Hostile

Key insight: Tags are the ground truth labels the entire learning system trains on.

Intent Signal: Another Face for the Intelligence Engine

Intent Signal isn't a separate product - it's a natural extension of the intelligence engine.

1. Bigger granularity for describing ICP

Instead of abstract ICP criteria, say: "I want companies actively hiring for sales roles" or "companies evaluating CRM tools right now"

2. Additional reasons to reprioritize leads

Every intent signal is a new data point that feeds into dynamic lead prioritization

Intent Signal Examples

"Hiring for sales roles"

→ Agent bumps leads at companies with recent sales job postings

→ Crafts messages around scaling challenges

"Recently funded"

→ Agent watches for funding announcements

→ Leads with fresh funding get higher priority and growth-focused messaging

"Evaluating competitor tools"

→ Agent prioritizes leads showing competitor research activity

→ Switches messaging to competitive differentiation

"Published content about [pain point]"

→ Agent monitors LinkedIn/blog activity for pain-related content

→ References the specific topic in outreach

Dialer Integration: Adding Phone as a Channel

Phone calls become another channel the agent can select - like email or LinkedIn.

The difference: For now, calls are executed by a human (you or your SDR team).

How it works

  1. Agent decides a phone call is the best next action
  2. Agent prepares a call package: script, context, talking points, objection prep
  3. Human caller makes the call using the agent's preparation
  4. Caller logs outcome, recording, notes
  5. Agent processes results and learns

Future: When AI voice agents mature, the transition is seamless - only the execution endpoint changes.

ACT V

Technical
Architecture

AI Messages Framework

The engine behind the agent's message generation capability

Stage 1: Context Assembly
↓
Stage 2: Personalization Waterfall
↓
Stage 3: AI Generation
↓
Stage 4: Quality Control
↓
Stage 5: A/B Testing
↓
Stage 6: Delivery & Learning

Stage 1: Context Assembly

1.1 Enginy Client Context

Brand voice, product info, case studies

1.2 Enginy Identity Context

Sender profile, role, territory

1.3 Combined Client Context

Unified sender profile

1.4 ICP (Company Level) Context

Firmographics, tech stack, news, intent signals

1.5 Buyer Persona (Contact Level) Context

Job title, context, pain points, KPIs, messaging framework

1.6 Historical Data RAG Layer

Learning from past interactions

Stage 2: Personalization Waterfall

Level 1: Industry baseline
Level 2: Company-specific insights (size, growth, tech stack)
Level 3: Role-specific pain points and value props
Level 4: Individual activity (job changes, LinkedIn posts)
Level 5: Timing (optimal send time, fiscal cycles, buying stage)
Level 6: Intent signals (website visits, competitor research, funding)

Context accumulates at each level. By Level 6, the message is hyper-personalized.

Stage 4: Quality Control

0-60: Poor

Regenerate with major revisions

60-80: Needs Work

Apply targeted suggestions and regenerate

80-100: Ready

Proceed to A/B Testing

Quality Dimensions

  • Personalization depth
  • Value proposition clarity
  • Call-to-action strength
  • Tone appropriateness
  • Length optimization

Continuous Learning: Six Feedback Loops

Outcome-Weighted Learning

Reply Received

1.5x weight

Light engagement signal

Meeting Booked

2.0x weight

Strong interest indicator

Opportunity Created

2.5x weight

Pipeline addition

Deal Closed

3.0x weight

Revenue generated - maximum learning signal

Closed deals teach the system 3x more than simple replies. The learning corpus prioritizes what actually converts.

ACT VI

How It All
Connects

The Enginy Intelligence Engine: Full System

STRATEGY LAYER

Best Quality
Compound Intelligence
AI-First Experience
↓

AGENT LAYER (THE BRAIN)

WHO (Dynamic prioritization) WHEN (Signal-based timing) HOW (Channel + message per lead) REPLY (Goal-driven handling) FEEDBACK (Chat + Action log + Reasoning)
↓

CONVERSATION TAGGING LAYER

Structured signal data from every reply

↓

EXECUTION LAYER (AI MESSAGES FRAMEWORK)

Context Assembly → Personalization → Generate → Quality → A/B Test → Send & Learn

Mapping Strategy Pillars to Capabilities

Best Quality
→

Agent writes each message individually with full context

Quality gates prevent bad output

Reasoning for every decision, transparent to user

Compound Intelligence
→

System learns from every interaction (40% → 90% accuracy)

Chat feedback refines ICP, messaging, channels, timing

Double-lock partner strategy

AI-First Experience
→

User describes ICP, agent does the rest

First session to live campaign in under 30 minutes

Adjustable autonomy levels for trust building

Implementation Roadmap

Phase 1: Foundation (Weeks 1-4)

  • Data source integration
  • Context assembly pipeline
  • Vector database for RAG
  • Quality scoring calibration

Phase 2: Message Engine (Weeks 5-8)

  • RAG layer activation
  • Limited deployment (100-500 emails/week)
  • A/B testing framework
  • Chat interface for campaign refinement

Phase 3: Agent Capabilities (Weeks 9-16)

  • Dynamic lead prioritization engine
  • Signal monitoring and real-time scoring
  • Channel selection logic
  • Action log and reasoning UI

Phase 4: Full Agent (Weeks 17-24)

  • Continuous lead discovery
  • Full conversational feedback loop
  • Living priority queue
  • Cross-campaign learning

Success Metrics

Email Quality Score (avg)

>85

Daily

Reply Rate by Persona

+20% vs baseline

Weekly

Agent Decision Accuracy

>80% user approval

Weekly

Lead Prioritization Accuracy

>70% of top-10 convert

Monthly

Time-to-First-Campaign

<30 minutes

Per new user

User Autonomy Level (avg)

Level 2+ after 30 days

Monthly

Pipeline Contribution

>25% of new pipeline

Monthly

Learning Corpus Growth

+500 patterns/month

Monthly

The Vision

Stop making users work to get value from AI.

Make the AI work for them.

Build intelligence that compounds with usage.

So they can't imagine leaving.

Questions?

The Enginy Intelligence Engine

Enginy graphic