Skip to main content

ADR-001: Use Agentic AI for Market Sentiment Analysis

Status: Accepted Date: 2025-09-15 Deciders: Nick Spiers, Michael Woodard

Context and Problem Statement​

Traditional commercial real estate appraisals rely heavily on appraiser subjectivity for market sentiment adjustments. We need a systematic, repeatable approach to determining market sentiment that:

  • Reduces individual appraiser bias
  • Provides consistent methodology across all valuations
  • Can learn and improve from historical outcomes
  • Is defensible to regulators and investors

Key Question: What type of AI should power CREstimate.ai's sentiment analysis?

Decision Drivers​

  • Objectivity: Need for consistent, repeatable sentiment scoring
  • Learning Capability: Ability to improve from historical valuation outcomes
  • Scalability: Handle multiple property types and geographic markets
  • Investor Confidence: Technical approach that differentiates from commodity AI
  • Regulatory Defensibility: Explainable methodology for compliance review
  • Competitive Moat: Defensible technical approach vs easily replicated solutions

Considered Options​

  1. Generative AI (LLM-based narrative generation like ChatGPT)
  2. Agentic AI (Goal-directed autonomous agent) ← CHOSEN
  3. Traditional Statistical Models (Regression, time-series analysis)
  4. Human Expert Panel (Consensus from multiple appraisers)

Decision Outcome​

Chosen Option: Agentic AI (Smitty - CRESI agent)

Rationale​

Agentic AI provides goal-directed autonomous behavior where the AI:

  • Takes actions to accomplish the goal (accurate sentiment scoring)
  • Learns from outcomes (actual market results vs predictions)
  • Self-corrects by adjusting future analysis based on feedback
  • Operates independently without constant human oversight

Unlike generative AI which creates text, agentic AI makes decisions and learns—exactly what's needed for systematic sentiment analysis.

Implementation​

Smitty is our agentic AI economist that:

  1. Analyzes 6 sentiment data layers autonomously
  2. Calculates asset-specific weighted sentiment scores
  3. Stores outcomes in Memory Bank
  4. Adjusts future weighting based on historical accuracy
  5. Operates continuously without human intervention per valuation

Consequences​

Positive​

  • ✅ Consistent Methodology: Every valuation uses the same systematic approach
  • ✅ Continuous Improvement: Each valuation makes the system smarter
  • ✅ Competitive Differentiation: Technical moat vs commodity generative AI
  • ✅ Regulatory Defensibility: Transparent, auditable decision-making process
  • ✅ Scalability: Handles unlimited valuations without human bottleneck
  • ✅ Explainability: Can trace exactly how sentiment score was calculated

Negative​

  • ❌ Higher Complexity: More sophisticated than simple LLM text generation
  • ❌ Development Time: Agentic systems require more engineering than prompt-based AI
  • ❌ Education Needed: Investors/users may not understand agentic vs generative distinction
  • ❌ Initial Training: Requires historical data to establish baseline accuracy

Pros and Cons of Options​

Option 1: Generative AI (e.g., ChatGPT-style LLM)​

Pros:

  • ✅ Faster to implement (prompt engineering vs building agent)
  • ✅ Easier to explain (familiar "ChatGPT-like" mental model)
  • ✅ Good at generating appraisal narratives and reports
  • ✅ Large pre-trained models available

Cons:

  • ❌ Doesn't take actions - only generates text
  • ❌ No goal-directed behavior - responds to prompts, doesn't autonomously analyze
  • ❌ Limited learning - no built-in feedback loop from outcomes
  • ❌ Easily replicated - any competitor can use ChatGPT API
  • ❌ Not optimized for numerical analysis - better at text than scoring

Why Rejected: CREstimate.ai needs autonomous analysis and learning, not text generation

Option 2: Agentic AI (CHOSEN)​

Pros:

  • ✅ Goal-directed autonomous behavior - analyzes sentiment to achieve accuracy goal
  • ✅ Self-correcting via feedback loops from actual outcomes
  • ✅ Can optimize for valuation accuracy over time
  • ✅ Memory bank enables continuous learning
  • ✅ Competitive differentiation - defensible technical approach
  • ✅ Handles complex multi-factor analysis better than generative models

Cons:

  • ❌ More complex to build than prompt-based generative AI
  • ❌ Requires domain expertise to design feedback loops
  • ❌ Longer development timeline
  • ❌ Need to educate stakeholders on agentic vs generative difference

Why Chosen: Perfect fit for autonomous sentiment analysis with continuous learning

Option 3: Traditional Statistical Models​

Pros:

  • ✅ Well-understood by appraisers and regulators
  • ✅ Mathematically transparent (regression equations, coefficients)
  • ✅ No "black box" AI concerns
  • ✅ Fast to run, low computational cost

Cons:

  • ❌ Doesn't adapt to new market conditions without manual re-tuning
  • ❌ Still requires subjective weight tuning by humans
  • ❌ Can't capture complex non-linear relationships in market sentiment
  • ❌ No competitive differentiation - standard approach anyone can copy
  • ❌ Brittle - breaks when market conditions change (e.g., 2008 crisis, COVID)

Why Rejected: Lacks adaptability and learning capabilities needed for dynamic markets

Option 4: Human Expert Panel​

Pros:

  • ✅ High trust from traditional appraisers
  • ✅ Incorporates deep domain expertise
  • ✅ Can handle novel situations with human judgment
  • ✅ No AI "black box" concerns

Cons:

  • ❌ Doesn't scale - limited by human availability
  • ❌ Introduces same subjectivity problem we're trying to solve
  • ❌ Expensive - requires paying multiple experts per valuation
  • ❌ Slow - human consensus takes time
  • ❌ Inconsistent - different panels may disagree

Why Rejected: Reintroduces the subjective, non-scalable problems we're solving

Notes​

Agentic AI Definition: An AI system that autonomously takes actions to accomplish goals and learns from the results of those actions, as opposed to generative AI which creates content in response to prompts.

Smitty: Our agentic AI economist, named after CRESI (Commercial Real Estate Sentiment Index). The personification emphasizes its autonomous, intelligent behavior.

Last updated: 2025-10-01 · Version 1.0.0