Classifying Business Development Tasks with COMET: A Human-AI Delegation Framework

April 7, 2026 | 8-10 minute read

Classifying Business Development Tasks with COMET: A Human-AI Delegation Framework

Within ARKONA, our multi-agent AI ecosystem, BizOps is a relatively new domain. While CoreOps (cyber-physical reverse engineering) and REOps have been operational for some time, successfully integrating AI into business development required a different approach. Simply applying LLMs to CRM data wasn’t enough; we needed a structured framework for intelligent task classification and, crucially, a *delegation* strategy balancing AI autonomy with human oversight. That’s where COMET came in.

The Problem: Beyond Simple Automation

Early attempts at automating BD tasks resulted in predictable failures. LLMs, while adept at summarizing information and drafting emails, struggled with nuance, particularly when dealing with client interactions. A Level 5 AI – one that fully automates a task – isn't appropriate for everything. We quickly realized the core challenge wasn't *automation*, but *classification* – determining the appropriate level of human involvement for each task. Tasks ranged from basic lead generation to complex contract negotiations, demanding varying degrees of AI assistance and human control.

COMET: A 7-Step Delegation Framework

COMET, or Collaborative Orchestration for Multi-faceted Engagement and Tasking, is our 7-step framework designed to address this. It’s grounded in NIST 800-30 guidelines for risk evaluation and leverages IEEE standards for human-machine collaboration. The core principle is to categorize BD tasks along a 1-5 “Human-AI Interaction Level” scale:

Level 1: Human-Led, AI-Assisted. Tasks requiring high judgment, strategic thinking, and relationship building (e.g., client negotiations, complex pricing discussions). AI provides information, context, and predictive analysis.
Level 2: Human-In-The-Loop. AI proposes solutions or drafts communications, but a human *always* reviews and approves before execution. (e.g., crafting proposals with high financial implications).
Level 3: AI-Augmented. AI handles the majority of the task, with human oversight for exceptions or unusual cases. (e.g., qualifying leads based on predefined criteria, responding to simple inquiries).
Level 4: AI-Driven, Human-Monitored. AI operates autonomously, with humans monitoring performance and intervening only if predefined thresholds are breached. (e.g., automated data enrichment from public sources).
Level 5: AI-Autonomous. Fully automated tasks requiring no human intervention. (e.g., basic lead scraping, social media monitoring for mentions).

The 7 steps of COMET are:

Task Ingestion: Tasks arrive via multiple channels – email (monitored by the ‘Inbox Agent’), CRM updates (Salesforce integration), and direct agent requests.
NLP Analysis & Intent Recognition: The ‘Lexicon Agent’ uses MuXD to parse natural language, identify the task's intent, and extract key entities.
Risk & Sensitivity Assessment: A NIST 800-30-grounded engine evaluates the task's potential risk (financial, reputational, legal). This leverages data from our ‘Fortress Agent’ (security monitoring) and external threat intelligence feeds.
Level Assignment: Based on risk assessment and task complexity, COMET assigns a Human-AI Interaction Level. This is the critical decision point.
Agent Orchestration: The appropriate agent(s) are assigned to the task. For example, a Level 5 task might be handled solely by the ‘Prospector Agent’ (lead generation), while a Level 1 task is routed to a dedicated BD manager.
Execution & Monitoring: Agents execute the task, with appropriate human oversight based on the assigned level. Progress is logged and monitored by the ‘Chronos Agent’ (system monitoring).
Feedback Loop & Refinement: Human feedback is captured and used to refine the Level assignment logic and agent behavior. This continuous learning process is vital for improving COMET’s accuracy.

Technical Implementation: Task Routing and Level Assignment

The heart of COMET is a combination of the inter-agent communication broker (a pub/sub system built on RabbitMQ) and a set of YAML configuration files that define the rules for Level assignment. Here’s an example:


task_types:
  - name: "Lead Generation"
    keywords: ["lead", "prospect", "contact"]
    level: 5
    agent: "ProspectorAgent"

  - name: "Proposal Drafting"
    keywords: ["proposal", "quote", "pricing"]
    risk_threshold: 50000  # USD
    level: 2
    agent: "ArticulatorAgent" #AI Proposal Writer

  - name: "Contract Negotiation"
    keywords: ["contract", "agreement", "terms"]
    level: 1
    agent: "HumanBDManager"
    requires_approval: true

The ‘Lexicon Agent’ matches incoming task descriptions against these keywords. The risk threshold determines whether the task automatically falls into a higher (more human-involved) level. The agent specified is the primary actor responsible for the task. This configuration is dynamically loaded and updated by the ‘Configurator Agent’, allowing us to adapt COMET to changing business needs.

MuXD and the Role of LLMs

MuXD, our hybrid LLM router, is crucial. It intelligently routes requests between local Ollama models (for speed and cost-effectiveness) and Claude (for complex reasoning and creative tasks). For example, a Level 3 task – qualifying a lead – might be handled entirely by a local Llama 3 model running within the ‘Lexicon Agent’. However, crafting a personalized proposal for a Level 2 task would leverage Claude’s superior text generation capabilities. We've implemented token savings optimizations, prioritizing local models whenever possible to minimize Claude API costs.

The Newsroom Pipeline as a COMET Case Study

Our 5-agent newsroom editorial pipeline, responsible for publishing technical articles (like this one!), provides a compelling example of COMET in action. The pipeline leverages agents dedicated to research, writing, editing, fact-checking, and publishing. A new topic is ingested, the ‘Research Agent’ (Level 4, AI-Driven, Human-Monitored) gathers information. An initial draft is created by the ‘Writer Agent’ (Level 3, AI-Augmented). The ‘Editor Agent’ (Level 2, Human-In-The-Loop) refines the content, and the ‘FactChecker Agent’ (Level 3, AI-Augmented) verifies accuracy. Finally, the ‘Publisher Agent’ (Level 5, AI-Autonomous) schedules and publishes the article. The entire process is orchestrated by COMET, ensuring quality and efficiency.

Current Status & Performance

As of today, 2026-04-07, we have 21 out of 22 services operational, with 184 commits in the last 7 days. COMET has demonstrably improved BD efficiency by 25% (measured by deals closed per agent hour) and reduced errors by 15% (measured by post-sale issue reports). We’re continuously refining the Level assignment rules and agent behaviors based on real-world performance data.

Key Takeaway: The Value of Structured Delegation

The most significant lesson we learned isn’t about the power of AI, but the importance of *structured delegation*. Simply throwing LLMs at business problems isn't enough. A clear framework for classifying tasks and assigning the appropriate level of human involvement is essential for maximizing AI's potential and mitigating its risks. COMET isn't just about automation; it's about creating a harmonious collaboration between humans and AI, ensuring that the right tasks are handled by the right resources, at the right time. This approach, rooted in systems engineering principles and informed by standards like NIST 800-30, is what truly unlocks the value of an autonomous agent ecosystem.

Blog