Digitizing Real-World Workflows: Enhancing LLM Accuracy in Decision-Making

Your AI application has to fit your current workflows—don't try to make your workflows adapt to the LLM.

The Fundamental Problem

Asking an LLM a question like "Should I use a condensing boiler or steam boiler for my new building?" is not going to yield a trustworthy answer. The response will be generic, potentially inaccurate, and lacks the context needed for a real engineering decision.

But why is this the case? And more importantly, how can we fix it?

The 75% Correctness Goal

So, how can we get to a 75% level of correctness in AI-generated answers? The solution lies in digitizing the workflows that we all perform in the real world.

Once digitized, we can bound these workflows to our LLM for significantly better, more reliable answers that align with actual engineering practice.

Understanding Real-World Workflows

Before we can digitize workflows, we need to understand what they actually look like in practice.

Example: Boiler Selection Workflow

When an experienced engineer selects a boiler, they follow a systematic process:

Step 1: Building Analysis

• Building type and size (square footage)
• Occupancy levels
• Location and climate zone

Step 2: Load Calculation

• Heating load (BTU/hr)
• Domestic hot water requirements
• Simultaneity and safety factors

Step 3: System Requirements

• Operating pressure (psig)
• Temperature requirements (°F)
• Turndown ratio needed
• Redundancy considerations

Step 4: Economic Analysis

• Capital budget available
• Energy costs ($/therm)
• Expected lifespan (years)
• Maintenance budget

Step 5: Constraints

• Space availability (footprint and height)
• Utilities (gas availability, pressure, electrical service)
• Code requirements (emissions limits, efficiency minimums, seismic)

This is what engineers do in the real world—they follow a structured decision-making process based on multiple factors and constraints.

The Digitization Process

Step 1: Map the Decision Tree

Create a comprehensive map of how decisions are actually made:

1. Initial Assessment - What information is gathered first?
2. Calculations - What formulas and methods are applied?
3. Constraints - What limitations must be considered?
4. Options Evaluation - How are alternatives compared?
5. Final Selection - What criteria determine the final choice?

Step 2: Capture Domain Expertise

Work with experienced professionals to document:

• Rules of Thumb - Quick estimates and sanity checks
• Common Mistakes - What pitfalls to avoid
• Best Practices - Proven approaches for different scenarios
• Edge Cases - Unusual situations and how to handle them

Step 3: Create Decision Parameters

Define the inputs and outputs clearly:

Required Inputs:

• Building data (characteristics, size, usage)
• Load data (heating load analysis)
• Site data (conditions and constraints)
• Budget data (financial constraints)

Decision Factors and Weights:

• Efficiency (25% weight, minimum 85%)
• First cost (20% weight, maximum $150,000)
• Operating cost (30% weight, 20-year lifecycle)
• Reliability (15% weight, mean time between failures)
• Maintenance (10% weight, complexity level)

Output Recommendation:

• Primary recommendation (condensing, steam, or hybrid)
• Confidence level (0-100%)
• Reasoning for the recommendation
• Alternative options to consider
• Important warnings or considerations

Step 4: Build the Digital Workflow

Transform the real-world process into a digital framework:

Digital Boiler Selection Process:

1. Validate Inputs - Ensure all required data is provided and accurate
2. Calculate Heating Requirements - Determine actual heating loads based on building data
3. Filter Viable Options - Eliminate options that don't meet site constraints
4. Perform Lifecycle Cost Analysis - Calculate total cost of ownership over expected lifespan
5. Apply Decision Factors - Score each option using weighted criteria
6. Generate Recommendation - Provide final recommendation with detailed reasoning

Bounding LLMs with Digitized Workflows

Once we have digitized workflows, we can effectively bound our LLM queries:

Before Digitization:

User: "Should I use a condensing boiler or steam boiler?"

LLM: "Condensing boilers are more efficient (90-98% efficiency) and better for low-temperature applications, while steam boilers are better for high-temperature processes..."

Generic, unhelpful for actual decision-making.

After Digitization with Workflow Bounding:

User: "Should I use a condensing boiler or steam boiler?"

System: Runs through digitized workflow, gathering required data

LLM (Bounded): "Based on your building profile:

• Office building, 45,000 sq ft, Climate Zone 5A
• Calculated heating load: 1.8M BTU/hr
• Low-temperature heating distribution (130°F supply)
• Natural gas available at 10 psig
• 20-year lifecycle cost analysis

Recommendation: Condensing Boiler

Reasoning:

1. Your low supply temperature (130°F) allows the boiler to condense, achieving 95%+ efficiency
2. Lifecycle cost savings vs. standard boiler: $87,000 over 20 years
3. Natural gas pressure sufficient for condensing unit
4. Return on investment: 5.8 years

Specifications:

• Size: 2.0M BTU/hr input (1.9M output with safety factor)
• Configuration: Two 1.0M BTU/hr units for redundancy
• Estimated installed cost: $125,000

Important Note: Your code requires NOx emissions below 20 ppm—specify ultra-low NOx burners.

Confidence: 87%"

Specific, actionable, based on actual engineering workflow.

Key Digitization Principles

1. Start with Common Scenarios

Don't try to digitize every edge case immediately. Focus on the 80% of decisions that follow standard patterns.

2. Maintain Flexibility

Build in the ability for engineers to override or adjust the workflow as needed. AI should assist, not replace, professional judgment.

3. Continuous Refinement

As the system is used, capture feedback and refine the workflows. AI should learn from real-world outcomes.

4. Transparent Reasoning

Always show how the AI arrived at its conclusion. Engineers need to understand the "why" behind recommendations.

Workflow Examples for Different Domains

HVAC System Selection

• Load calculations → Equipment sizing → First cost analysis → Operating cost analysis → Selection

Lighting Design

• Space analysis → Illumination requirements → Energy code compliance → Cost comparison → Specification

Energy Audit Process

• Baseline establishment → Data collection → Analysis → Measure identification → ROI calculation → Prioritization

Measuring Success

How do we know if our digitized workflows are working?

Key Metrics:

1. Accuracy Rate - Percentage of AI recommendations that match expert decisions
2. Time Savings - Reduction in time to make decisions
3. Consistency - Variance in decisions for similar scenarios
4. User Confidence - How often users accept AI recommendations
5. Outcome Validation - Did the recommendation prove correct in practice?

Target: 75%+ accuracy with 50%+ time savings

Implementation Roadmap

Phase 1: Workflow Documentation (Weeks 1-4)

• Interview subject matter experts
• Map decision trees and processes
• Identify key parameters and constraints

Phase 2: Digitization (Weeks 5-12)

• Build workflow logic
• Create decision engines
• Develop parameter databases

Phase 3: LLM Integration (Weeks 13-16)

• Connect workflows to LLM
• Test bounded queries
• Refine prompts and responses

Phase 4: Validation & Refinement (Weeks 17-20)

• Test with real scenarios
• Gather expert feedback
• Iterate and improve

Conclusion

The power of AI in professional applications doesn't come from the LLM alone—it comes from properly digitizing the real-world workflows that experts use every day.

By taking the time to:

1. Map how decisions are actually made
2. Digitize the process with all its nuances
3. Bound the LLM within this structured framework

We can achieve AI systems that deliver trustworthy, accurate, actionable recommendations that fit naturally into existing professional practice.

Remember: Don't force your workflows to adapt to the AI. Make the AI adapt to your workflows.

"Come engineer with me." - Raj Setty

Ready to digitize your engineering workflows? Contact us to discuss how we can help transform your decision-making processes.

Read the original LinkedIn post here.