Strategic LLM Implementation in Chemical Manufacturing with Focus on Documentation and Virtual Agents

chevron_philips_chemical · Energy · year n/a · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Generative AI governance and use-case selection in chemical manufacturing

Chevron Phillips Chemical needed a realistic, governed approach to generative AI across virtual agents, document processing, and internal manuals while accounting for legal, security, bias, testing, traceability, and risk concerns.

Constraints:

• The adoption effort involved legal/IP, IT security, digital workplace, data science, data engineering, and analytics stakeholders.
• Users needed education on what LLMs can and cannot do.
• The use cases included unstructured PDFs, documentation, and open-ended virtual assistants.
• Governance had to define appropriate use, risk assessment, testing, productivity enhancement, and traceability.

Architecture

CP Chem formed a cross-functional generative-AI team to set policy, educate leadership, assess realistic capabilities, and explore lower-risk use cases. The architecture combines provided models such as Dolly and GPT, Databricks/Unity Catalog governance, virtual agents for specific topics, RPA-style document extraction from PDFs, operations-manual assistants, user training, and testing/hardening for open-ended interfaces.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  cross_functional_genai_team["Cross-functional GenAI team"] --> leadership_education_program["Leadership education program"]
  cross_functional_genai_team --> use_case_portfolio["Use-case portfolio"]
  use_case_portfolio --> provided_llms["Provided LLMs"]
  use_case_portfolio --> testing_and_hardening_loop["Testing and hardening loop"]
  databricks_and_unity_catalog_platform["Databricks and Unity Catalog platform"] --> use_case_portfolio
  cross_functional_genai_team --> generative_ai_policy["Generative AI policy"]
  generative_ai_policy --> testing_and_hardening_loop
  generative_ai_policy --> databricks_and_unity_catalog_platform

Components

Data flow

The cross-functional team establishes policy and governance, educates leadership, and identifies use cases. Provided LLMs are applied to specific virtual-agent topics, PDF/document extraction, and manual assistants. Databricks and Unity Catalog support governed data access, while testing and user training constrain open-ended behavior and traceability.

Why this architecture

The approach is governance-first because CP Chem saw generative AI as useful but not a magic pill. Cross-functional leadership, provided models, governed data platforms, policy, user training, and testing let the company explore practical lower-risk use cases while controlling bias, truthfulness, and traceability risks.

Models used: Dolly · GPT · OpenAI GPT · Databricks models

Scale

• The generative-AI team included legal/IP, IT security, digital workplace, data science, data engineering, and analytics stakeholders.
  “that team included legal and intellectual property and included it security and included um digital workplace data science data engineering analytics”transcript, stated
• CP Chem reports massive amounts of operations manuals and documentation as a target corpus.
  “we have massive amounts of tations and operations manuals the main things that we're going after is figuring out how to go after the manuals”transcript, stated

Alternatives considered (roads not taken)

Treat LLMs as a magic-pill solution

Rejected because: CP Chem explicitly wanted leadership to cut through hype and see realistic capabilities.

Write deterministic code for highly variable document extraction

Rejected because: The source material varies too much for hand-coded extraction rules, so LLMs are considered for imposing structure.

Use LLM outputs as source of truth

Rejected because: The company trains users that the model's output should not be treated as authoritative.

Patterns

safety-policy-enforcement (governance)

CP Chem built a generative-AI policy that defines appropriate use, risks, testing, accountability, productivity goals, and traceability.

human-in-the-loop-gate (organizational governance)

A cross-functional team with legal, security, workplace, data science, engineering, and analytics stakeholders governs AI adoption and use-case assessment.

guardrails-io-filtering (runtime safety)

CP Chem identifies the need to harden open-ended assistants so they stay on topic and avoid unwanted answers or actions.

tool-use-orchestration (enterprise automation)

The use-case portfolio includes virtual agents and RPA-style document extraction from unstructured PDFs into structured information.

Tradeoffs

Start with lower-risk virtual assistants and documents.

Chose Specific-topic agents, document extraction, and manual assistants over Broad unconstrained enterprise agents.

Rationale: The company is exploring useful cases while explicitly testing and hardening open-ended interfaces and avoiding source-of-truth assumptions.

Use provided LLMs.

Chose Dolly and GPT from external providers over Building proprietary foundation models first.

Rationale: The interview describes exploiting provided models while focusing governance and application effort around them.

Make policy and traceability central to adoption.

Chose Generative AI policy plus traceability over Ad hoc self-service experimentation.

Rationale: CP Chem is a self-service data shop, so scalable platforms and explicit policy are needed to let users move safely.

Gotchas (what bit them)

Leadership hype can distort enterprise AI adoption.

Root cause: Generative AI promises difficult new capabilities but is not a magic-pill replacement for realistic risk assessment.

Resolution: Educate leadership and build a cross-functional team before scaling use cases.

Open-ended LLM interfaces are hard to test.

Root cause: Users can ask open-ended questions, making it difficult to ensure the app stays on topic and does not do unwanted things.

Resolution: Build explicit testing and hardening around topic boundaries and undesired behavior.

Bias has both technical and colloquial meanings in LLM adoption.

Root cause: Large language models combine data-science notions of bias with human fairness concerns.

Resolution: Train users on model behavior, prompting, examples, and the warning that the model should not be treated as source of truth.

Evaluation

CP Chem emphasizes testing, risk accounting, user training, and governance policy rather than reporting measured production evaluation results.

Outcomes

• CP Chem formed a cross-functional generative-AI team.
• The company completed a generative-AI policy.
• It identified virtual agents, document extraction, and operations-manual assistants as lower-risk starting use cases.

The interview gives organizational scope and target corpora such as large manuals, but does not report deployed traffic, latency, accuracy, or cost outcomes.