How ChatGPT Work Accelerated Stampli’s Product Launch
💡See how Stampli used Codex and ChatGPT Work to turn weeks of launch work into days.
⚡ 30-Second TL;DR
What Changed
Stampli combined Codex and ChatGPT Work during product launch preparation.
Why It Matters
The case suggests that AI-assisted development and work coordination can help lean teams meet aggressive product deadlines. It is particularly relevant to companies that need to compensate for limited engineering or design capacity.
What To Do Next
Pilot Codex and ChatGPT Work on one fixed-scope product-launch task, and compare completion time with your standard workflow.
Key Points
- •Stampli combined Codex and ChatGPT Work during product launch preparation.
- •The tools reduced launch production time from weeks to days.
- •The workflow addressed a resource constraint involving unavailable design capacity.
🧠 Deep Insight
Background and context from public sources — not the original article. 18 sources cited.
🔑 Enhanced Key Takeaways
- •Stampli's AI, known as "Billy the Bot," has been embedded in its Procure-to-Pay platform since its founding in 2015, processing over $80 billion in annual invoice value for 1,600+ customers.
- •ChatGPT Work, launched by OpenAI in July 2026, is an AI agent powered by the new GPT-5.6 model, designed to automate multi-step tasks across various workplace applications like Slack, Google Drive, and CRMs.
- •OpenAI's Codex, originally a coding agent, is now used by over 5 million people weekly, with more than 1 million utilizing it for non-software development tasks, and is being integrated into the new ChatGPT desktop app alongside ChatGPT Work.
- •Stampli's AI provides suggestions for ERP-structured invoice fields with an 87% coverage rate as of April 2026, with all suggestions remaining subject to human review and approval.
- •ChatGPT Work introduces "Scheduled Tasks" and enhanced "Computer Use" capabilities, allowing the AI agent to automate recurring workflows, access websites via a built-in browser, and interact with local applications and files in the background.
🛠️ Technical Deep Dive
- OpenAI Codex:
- Based on a large-scale transformer neural network architecture, descended from GPT-3 and extensively fine-tuned for code understanding and generation.
- Powered by
codex-1, a version of OpenAI'so3AI reasoning model specifically optimized for software engineering tasks. - Operates as a cloud-based software engineering agent within a sandboxed, virtual computer environment, capable of executing code and connecting with GitHub repositories.
- Features an "agent loop" at its core that orchestrates model inference and tool calls, repeating cycles of execution and re-querying until a user-facing message is produced.
- Utilizes a layered prompt structure for context management, stacking environment context, AGENTS.md file contents, sandbox permission rules, developer configuration instructions, tool definitions, and conversation history.
- Employs prompt caching as a primary mitigation for managing the quadratically growing JSON prompt in stateless requests, ensuring efficiency by caching static content.
- The Codex App Server uses a custom JSON-RPC protocol over stdio to decouple the core logic from various client surfaces (CLI, VS Code, web app, macOS app, third-party IDEs), supporting primitives like Item (atomic I/O), Turn (unit of agent work), and Thread (persistent conversation).
- ChatGPT Work (and underlying GPT-5.6):
- Powered by OpenAI's latest frontier model, GPT-5.6, which includes variants like Sol (flagship), Terra (lower-cost), and Luna (smallest, fastest, least expensive).
- GPT-5.6 models are designed for stronger performance in professional analysis, web browsing, tool use, and computer-based tasks.
- Leverages a built-in browser for accessing websites and online tools, and expanded "Computer Use" capabilities to interact with local applications and files.
- Integrates with various workplace applications via plugins, including Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, and CRM platforms.
- The underlying ChatGPT models are large language models (Generative Pre-trained Transformers) that train on vast amounts of text and undergo reinforcement learning from human feedback (RLHF) to refine output for helpfulness and contextual appropriateness.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
📎 Sources (18)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: OpenAI News ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.