AI on the PDT Project
Navigating the "AI Wild West" with structured approaches and realistic expectations for pipeline digital transformation.
The Reality of AI Experience
1
Initial Awe
Promises of perfection in 2 hours capture our imagination
2
The Rabbit Hole
4 hours later, emerging deflated and disheartened
3
Hard Truth
Jobs remain safe; AI cannot read minds or fill unclear requirements

Context is everything. Working with AI requires the same product management skills that have always been necessary.
AI: The Brilliant New Starter
No Common Sense
Understands everything but lacks practical judgement
Convincing Liar
Often doesn't know it's lying, making errors seem credible
No Self-Checking
Brilliant but doesn't verify its own work
Forgetful Genius
Forgets yesterday's work and over-complicates simple requests
These issues waste time, but structured approaches can unlock AI's true potential.
Building Effective AI Agents
Instead of relying on a single, generic AI model, leverage specialist agents—AI models fine-tuned or augmented with highly targeted knowledge bases for specific tasks or domains. This approach dramatically reduces "hallucinations" and increases accuracy by ensuring the AI operates within a defined, relevant information scope.
Product Spec Agent
Trained on Product Requirements Documents (PRDs), design specifications, and user stories. It can answer questions about feature requirements, scope, and dependencies.
Code Review Agent
Accesses your codebase, coding standards, and historical bug reports. It can identify potential issues, suggest optimizations, and ensure compliance with best practices.
Customer Insight Agent
Digests customer feedback, support tickets, and user survey data. It helps the team understand pain points, feature requests, and market sentiment.
Structured Agent Instructions
Role and Identity
Define the agent's specific role and expertise
Clear Objectives
Set measurable goals and success criteria
Contextual Background
Provide current state and pain points
Decision Framework
Establish priorities and escalation paths
Boundaries
Define limitations and compliance requirements
Why Structure Still Matters: Beyond Agent Instructions
Even with a meticulously configured agent – one with a clear role, defined objectives, and established boundaries (as discussed in the previous card) – the way we phrase our individual requests remains critical. A well-trained agent can still falter with vague instructions, delivering generic or off-target responses.
Vague Prompt Example
"Summarize this document for me."
Outcome: The agent might provide a lengthy, unformatted summary, miss key details, or focus on less relevant information, despite its overall capabilities.
Structured Prompt Example (SRS Applied)
"SITUATION: We need a summary of the attached Q3 financial report for our board meeting. REQUEST: Provide a 3-paragraph executive summary focusing on key revenue drivers and challenges. SUCCESS: The summary is concise, highlights financial performance, and is suitable for a non-technical audience."
Outcome: The agent delivers a highly targeted, actionable summary that meets specific needs, leveraging its deep understanding of its role and objectives.
This is where the SRS Method comes in. It provides a simple yet powerful framework for crafting clear, effective prompts that guide agents to deliver precise and valuable outputs, maximizing their potential even when they have detailed initial instructions.
1
SITUATION
Brief current state or change that's occurred
2
REQUEST
One specific action you want the agent to take
3
SUCCESS
How you'll know the request is complete

Break problems into small chunks. Ask AI how to create a plan rather than solving everything at once.
Breaking the Static Data Barrier with MCP
While Retrieval-Augmented Generation (RAG) significantly enhances AI by incorporating external knowledge, its reliance on static datasets limits its application in dynamic, real-world environments. To truly unlock AI's potential, seamless integration with live, evolving data sources is essential.
The Problem: Static Data Limitations
RAG excels at leveraging vast, pre-indexed knowledge bases. However, for applications requiring real-time insights or interaction with live systems, static data falls short. AI needs direct, dynamic access to operational data to remain relevant and effective.
The MCP Solution: A Standard for Dynamic Integration
The Model Context Protocol (MCP) provides a standardized, open-source framework for AI systems to dynamically integrate with any external data source or application. Think of it as the OPC (Open Platform Communications) of AI-system integration, ensuring interoperability across diverse platforms.
Compelling Example: Digital Twin Pipeline
Imagine a digital twin agent for a complex pipeline system. Linked via an MCP server, this agent gains complete, real-time knowledge of the pipeline's operational state, historical data, and engineering parameters, transforming it into a truly dynamic, actionable intelligence.
Real-time SCADA + Model Data Access
AI can query live operational data from SCADA systems combined with engineering models, providing instant insights.
Natural Language Queries
Users can ask the AI directly, "What's the temperature in temperature zone 3?" or "Summarize pressure in pressure segment 5 over the last hour," for intuitive data analysis.
Continuous Monitoring with Automated Alerts
AI continuously monitors data to detect anomalies in flow rates, vibration, or temperature, and issue threshold alerts automatically, enhancing proactive maintenance.

MCP transforms AI from a static knowledge base into dynamic, real-world intelligence, capable of interacting with and understanding live operational environments.
AI Applications for PDT Project
Scope Definition
Crawl existing documents to extract key features and create scored scope lists.
Vendor Proposal Vetting
Extract key points and score proposals against requirements automatically.
Test Data Generation
Generate and verify test data for vendor PoCs and system comparisons.
Meeting Transcription
Record and transcribe multilingual meetings for comprehensive documentation.
Industry AI Success Stories
92%
Line Pack Utilisation
Increased from 78% using AI-optimised compressor scheduling
87%
Faster Convergence
Reinforcement learning for flow optimisation
$12M
Cost Savings
Early valve failure detection 14-16 weeks in advance
30x
Performance Boost
Rust-based numerical solvers vs NumPy
Physics-Informed Neural Networks and hybrid architectures deliver measurable operational improvements.
AI as Force Multiplier
Structured approaches, specialist agents, and human oversight unlock AI's potential across all PDT project phases—from scope definition to operational optimisation.
Context is King
Clear requirements and structured prompts
Human in the Loop
AI augments, doesn't replace human judgement
Iterative Approach
Break complex problems into manageable chunks