The Digital Twin Approach to SOP Automation
How we built a system that auto-discovers how a business actually operates, compares it to documented SOPs, and uses the alignment gap to infer requirements and automate software delivery.
Every Company Has Two Versions of Itself
There’s the company that exists on paper — the SOPs, the process docs, the org chart. And then there’s the company that actually exists — the workarounds, the tribal knowledge, the “ask Sarah, she knows how this works.” These two versions drift apart every single day.
SOPs Are Written Once and Forgotten
Most companies invest real effort writing their first set of SOPs. Within weeks, reality drifts. Within months, the SOPs are fiction. Nobody has time to continuously audit and update them because the gap is invisible until something breaks.
Institutional Knowledge Is Trapped
How your business actually operates lives in Slack threads, email chains, meeting notes, and people’s heads. When someone leaves, that knowledge walks out the door. There’s no systematic way to capture the “real” version of your processes.
You Can’t Automate What You Can’t See
Every time you want to build software, deploy an automation, or onboard a new tool — you start from scratch. You interview stakeholders, map processes, write requirements. The knowledge was always there, buried in your communication data. You just couldn’t extract it.
The Expensive Consequence
Companies spend thousands of hours per year on process documentation that’s outdated before it’s published. And then they spend thousands more on consultants to “discover” what the business actually does before they can build anything. What if both of those problems were the same problem — and both had the same solution?
The Two-Layer SOP Model
The foundational insight behind the digital twin: every process has two representations, and the gap between them is where all the value lives.
Layer 1: Documented SOPs
- Official, approved SOPs from Google Drive or SharePoint
- Written by process owners, reviewed by leadership
- Versioned, with change history and approval chain
- Represents the company’s ideal operating state
- Vectorized and searchable in the Company Brain
Layer 2: Discovered Processes
- Auto-generated from communication data analysis
- Extracted from Slack, email, tasks, meetings, and more
- Reflects real-time operational patterns and behaviors
- Shows actual roles, handoffs, and decision points
- Continuously updated as new data flows in
The Gap Is the Gold
The comparison between these two layers produces three things that are enormously valuable:
- Discrepancies — where reality has drifted from documentation (and whether the docs or the team need to change)
- Coverage gaps — processes that exist in practice but have never been documented
- Inferred requirements — a structured, machine-readable understanding of what the business actually needs, derived from what it actually does
How the Digital Twin Works
The system ingests all of a company’s communication data, uses AI to discover the actual business processes hidden in that data, and then compares those discoveries against documented SOPs.
Data Ingestion
Slack, Email, Tasks, Meetings, Docs
Company Brain
RAG + Wiki + Knowledge Graph
Process Mining
Discover how work happens
SOP Comparison
Find where reality drifts
Suggestions
AI-generated improvements
Step by Step
1. Crawl Everything
Automated crawlers pull data from every communication channel the business uses — Slack messages, email threads, task updates, meeting transcripts, shared documents. The system supports 20 data source categories covering 57 platforms. Each client connects the channels they actually use.
2. Build the Brain
All data feeds into a three-layer knowledge base. RAG layer: vector embeddings for semantic search. Wiki layer: compiled theme articles and client dossiers. Knowledge graph: entities and relationships with traversal queries. This is the company’s institutional memory, made searchable and queryable by AI agents.
3. Mine the Processes
A process mining engine analyzes the brain data to discover actual business processes. It classifies communication into business activities, extracts sequences, identifies roles and handoffs, detects variants, and produces structured process models. Rolling 90-day window, configurable per client.
4. Compare Two Layers
The SOP comparison engine matches each discovered process to its corresponding documented SOP and identifies six types of discrepancies: missing steps, skipped steps, sequence differences, role mismatches, undocumented handoffs, and tool drift. Each process gets a compliance score.
5. Generate Suggestions
For each discrepancy, the system determines: should the SOP be updated (reality is better than docs), should the team be coached (docs exist for a reason), or is it just informational? Each suggestion includes the actual proposed text change, not vague advice.
6. Close the Loop
SOP owners are notified and can approve, reject, or modify suggestions. When approved, the documented SOP is automatically updated with proper versioning and provenance tracking. For processes with no SOP at all, the system generates a draft for review. The twin stays in sync.
What the Process Mining Engine Discovers
The engine doesn’t need you to tell it what your processes are. It discovers them from how your team communicates.
Activities & Steps
Named process steps extracted from communication patterns, with frequency and recency data.
Sequences & Transitions
What follows what. The engine maps the actual order of operations including branches and loops.
Roles & Handoffs
Who performs each step, who passes work to whom, where bottlenecks form, where steps get skipped.
Decision Points
Where the process branches based on context — like enterprise vs. SMB onboarding following different paths.
Tools & Systems
What platforms and tools are actually involved at each step (vs. what the SOP says should be used).
Process Variants
The same process happening differently depending on context, client type, team, or time pressure.
Confidence-Based Surfacing
Discoveries at ≥85% confidence are presented as findings. Discoveries between 70–85% are shown with “needs more data” treatment. Below 70%, the system continues gathering data silently. This prevents noisy false positives from undermining trust in the system.
Six Types of Discrepancies
Missing Steps
Steps that happen in practice but aren’t in the documented SOP. The team invented something useful that never got captured.
Skipped Steps
Steps documented in the SOP but rarely or never observed. Either the step is unnecessary or the team is cutting corners.
Sequence Differences
Steps happening in a different order than documented. Sometimes the team found a better sequence.
Role Mismatches
Steps performed by different people than the SOP specifies. Org changes that never got reflected in documentation.
Undocumented Handoffs
Transitions between people or teams that aren’t captured in the SOP. The invisible coordination that holds operations together.
Tool Drift
Different tools being used than what the SOP specifies. The team migrated to Notion but the SOP still says “update the Google Sheet.”
From SOPs to Automated Software Delivery
This is where the digital twin stops being an SOP management tool and becomes the foundation for automating software development itself. Each layer builds on the one before it.
Discover Processes
The twin discovers how the business actually operates by mining communication data.
Align SOPs
The twin compares discoveries to documented SOPs and keeps them synchronized.
Infer Requirements
Accurate, current SOPs become machine-readable process definitions. Requirements emerge naturally from what the business actually does.
Generate Automations
With structured requirements, AI agents can identify which process steps are automatable and generate the actual automation logic.
Deploy Software
Automations and custom software are built, tested, and deployed by AI agents — with human approval at every critical gate.
Why This Order Matters
Most companies try to jump straight to automation. They hire consultants, interview stakeholders, write requirements documents, and then build. But they’re building on a foundation of guesses about how the business operates. The digital twin inverts this: you start with what’s actually happening, keep it accurate, and let the requirements and automation flow naturally from reality. No more “the requirements were wrong” because the requirements come from observed behavior.
The Digital Twin Sits Inside a Larger System
The SOP automation twin is one layer of a multi-component AI platform. Each layer feeds the others, creating compounding intelligence over time.
Company Brain
Three-layer knowledge base: vector search, wiki, and knowledge graph. The memory of everything the business knows.
FoundationDigital Twin
Process discovery, SOP comparison, and improvement suggestions. Models how the business operates.
This FeatureCRM
Brain-synced relationships, deals, and team data. Connects operational knowledge to business relationships.
Relationship LayerReadiness Assessments
Three-tier AI maturity scoring. Organization, department, and personal level. Prioritizes where to start.
Discovery LayerClient Portal
Client-facing dashboards, requirements pipeline, assessments, and onboarding. The interface for clients to work with their AI team.
Interface LayerVoice Agent
Conversational AI assistant with 40+ tools. Calendar, email, Slack, brain search, delegation — all by voice.
Interaction LayerAgent Platform
Multi-agent orchestration. 14 specialized AI agents with 100+ skills, Slack routing, delegation, and autonomous execution.
Execution LayerOnboarding Engine
Brain-first client onboarding. Crawl data first, then use what we learn to streamline the entire setup process.
Growth LayerCompounding Intelligence
Each layer makes the others smarter. The brain feeds the twin. The twin improves the SOPs. The SOPs inform the requirements. The requirements drive the agents. The agents generate automations. The automations create more data. The data feeds the brain. It’s a flywheel — and every turn makes the entire system more valuable.
Roadmap: Four Phases to Full Autonomy
Each phase delivers standalone value and builds on the previous one. We’re live on Phase 1.
1
Process Discovery + SOP Alignment
Auto-discover processes, compare to SOPs, generate improvement suggestions
LIVE
↓
Process Discovery + SOP Alignment
Auto-discover processes, compare to SOPs, generate improvement suggestions
This is where we are now. The system can ingest communication data from 20 categories of data sources, build a three-layer knowledge base, discover business processes through AI-powered analysis, compare them to documented SOPs, identify discrepancies, and generate concrete improvement suggestions with proposed text changes.
Data crawlers for Slack, Gmail, Asana, GDrive, Zoom, and more
Three-layer brain: vector embeddings, wiki compiler, knowledge graph
Process mining engine with classification and sequence extraction
SOP comparison with six discrepancy types and compliance scoring
AI suggestion engine with approval workflow
Visual web interface with interactive process diagrams
2
Real-Time State Synchronization
Near-real-time data flow, live process maps, agent context exposure
NEXT
↓
Real-Time State Synchronization
Near-real-time data flow, live process maps, agent context exposure
Process models update within minutes of new activity (not waiting for nightly crawler runs). The visual interface reflects current state. AI agents can query the twin for live operational context when making decisions.
3
Simulation Sandbox
“What if” scenario testing for process, team, and tool changes
PLANNED
↓
Simulation Sandbox
“What if” scenario testing for process, team, and tool changes
“What happens if we change this process?” “What if this person leaves?” “What if we double our client load?” The twin simulates impacts before you make real changes. Save and compare scenarios side by side.
4
Closed-Loop Optimization
AI agents propose and execute improvements with graduated autonomy
PLANNED
↓
Closed-Loop Optimization
AI agents propose and execute improvements with graduated autonomy
The twin becomes self-improving. Agents propose workflow modifications based on observed patterns. Graduated autonomy: full auto for low-risk actions, human-on-the-loop for medium, human-in-the-loop for high-stakes. Cross-client pattern learning (anonymized) means every client makes every other client’s twin smarter.
How You Could Build Something Similar
We believe in sharing ideas. If you’re technical and already running agents, here’s the conceptual blueprint for building your own digital twin SOP system.
Start with a Knowledge Base
You need a place to store and search company knowledge. At minimum: a PostgreSQL database with pgvector for embeddings, a chunking strategy (we chunk by H2 headings), and an embedding model (we use text-embedding-3-small, 1536 dimensions). Add a wiki compiler and knowledge graph for bonus points.
FoundationBuild Data Crawlers
Write crawlers for whatever platforms your company uses. Key patterns: hash-based incremental sync (only process new/changed data), orphan deletion (remove docs when source data is deleted), content deduplication, and sensitivity classification. Start with Slack and email — they contain 80% of process signal.
Data PipelineBuild a Process Mining Engine
This is the hard part. You need an LLM pipeline that can: (1) classify documents into business activity categories, (2) extract process sequences from classified activities, (3) identify roles, handoffs, and decision points, (4) detect process variants. Think of it as turning unstructured communication into structured workflow models.
Core IntelligenceImport Your Existing SOPs
Pull documented SOPs from wherever they live (Google Drive, SharePoint, Notion) into the same knowledge base. Chunk them the same way, embed them the same way. Now you have both layers in one searchable system. Track ownership from document permissions.
Layer 1Build the Comparison Engine
Match discovered processes to documented SOPs (start with name/topic similarity). For each pair, use an LLM to identify the six discrepancy types. Score compliance as a percentage. Rank by frequency and impact. Surface only high-confidence findings.
Core LoopAdd Suggestions + Approval
For each discrepancy, generate a concrete suggestion: “Update SOP section 3.2 from X to Y because the team has been doing Y for 6 weeks with better outcomes.” Route to the process owner (derive from doc permissions) via Slack or email. Track the full approval/rejection history.
Closed LoopKey Decisions We Made
- Per-client isolation: Every client runs on their own server. No cross-client data sharing. The brain, twin, and all services are isolated.
- Brain-first onboarding: We crawl data before we ask questions. Then we use what the brain already knows to turn interview questions into confirmations.
- Confidence thresholds: ≥85% to present as a finding, 70–85% as “needs more data,” below 70% stays silent. Trust is everything.
- Dynamic skill deployment: Instead of a fixed rollout order, skills deploy when prerequisites are met. Credential + goal matching, not static waves.
- SOP ownership from doc permissions: Don’t create a separate ownership system. Read it from the document system that already manages it.
20 Data Source Categories
The system is designed to ingest data from whatever tools a company uses. Different clients have different tool stacks. Here’s the full taxonomy.
Core Sources (1–6)
Messaging (Slack, Teams), Email (Gmail, Outlook), Tasks (Asana, Monday, ClickUp), Documents (GDrive, SharePoint, Notion), Calendar (Google, Outlook), CRM (HubSpot, Salesforce)
Active CrawlersEnrichment Sources (7–19)
Meeting transcripts, Accounting (QuickBooks, Xero), Support (Zendesk, Intercom), Knowledge Bases (Confluence, Guru), HR, VoIP, E-commerce, ERP, Social Media, Contracts, Forms, AI Wearables, Code
RegisteredOnboarding-Driven
During client setup, the system asks which tools they use, then configures crawlers automatically. New platforms can be added without restructuring the system. The architecture is connector-agnostic — if a connector doesn’t exist yet, it gets flagged for the delivery team.
DynamicWant to See This in Action?
We’re actively onboarding companies to the platform. If you’re running a business with Slack and email, your digital twin is waiting to be discovered.