From zero to AI practice in 6 weeks: program template, new ChatGPT ads, purchasing agent with approval

Important AI news

Six-week program gets business teams building

An AWS playbook shows how you lead non-developers to working AI prototypes in six weeks. Four hours per week, fixed mentors, an MVP from day 2, daily standups, and open office hours. The example team built five specialized agents on Amazon-Nova models with orchestration via the Strands Agents SDK and storage in DynamoDB, with response times under five seconds according to the vendor. The effectiveness of the structure and the claim of triple higher skill retention come from vendor program data and are not externally validated. If you already have these talks, give your team permission to fail and set clear guardrails. Solid skills will form.¹

ChatGPT rolls out a new visual ad format and more measurement

OpenAI is adding a new visual ad format in ChatGPT and expanding measurement features, attribution partners, and brand suitability options. The announcement is from four days ago. Attribution gets broader support. That is the assignment of ad impact to channels. If you run paid campaigns, you can now test ChatGPT inventory more precisely and assign impact better. Details on availability depend on the ad account.²

Purchasing agent with human approval from LangChain

LangChain shows in an X post a purchasing agent that finds products, prepares orders, and pays via a Stripe link after a human approval. Control happens in Slack. Payment only after your confirmation. This is a design pattern, not a production-ready solution with documentation. You can prototype such a flow and bake in budget limits and two-person approvals.³

Highlights for your day-to-day

• Cut costs like LegalOn. The team cut estimated Codex daily costs by 65 percent by assigning tasks to fitting models and actively managing budgets. You can split your workloads by task and model today and set hard budget caps.⁴
• Establish longer AI workflows with review. Oracle describes workflows across multiple data sources with human review in ChatGPT Work and Codex. You define quality gates instead of only automating single steps.⁵
• Make data agents more robust. The NVIDIA team progressed in the KDD competition with a smaller, clearer agent harness. Constrain tools, inputs, and validation paths. Check answers systematically against heterogeneous sources.⁶

Tools and updates

• Copilot for Windows with Hybrid Intelligence. Microsoft announces that Copilot will use PC context with your permission, take actions, and use local models when needed to save tokens. Check privacy before you share context.⁷
• Simulations with AI agents in NVIDIA Omniverse. Examples show how agents assemble simulation apps by voice, wire up test scenarios, and improve digital twins via sensor comparisons. Useful for teams with Omniverse and access to frontier models.⁸
• Synthetic data for enterprise agents. AutoSynthData describes a way to generate training data for enterprise agents. The announcement is from seven days ago. This helps when real data is not usable for compliance reasons.⁹
• GitHub Copilot with on-device AI on new Windows PCs. Local code suggestions without the cloud can cut latency and keep data in house. Requires new Windows hardware.¹⁰

Try it in five minutes

1) Open your AI chat tool and ask for a one-page plan for a six-week AI Builders run. Four hours per week, MVP by day 2, daily short check-ins, weekly office hours, clear guardrails, and a human approval before every production step. Give an example problem from your day. For example email routing or contract summarization.¹ 2) Have it generate an invite for participants plus a manager briefing note, including time needed, target state, risks, and data sources.¹ 3) Check whether the plan includes the guardrails. Data use clarified, a kill switch defined, a budget limit per experiment, and a demo date at the end of week 2.¹

Expected result: a crisp page and an invite you can send internally. Check criterion: the plan meets four hours per week, shows an end-to-end MVP by day 2, and names human-in-the-loop approvals explicitly.¹

Sources

1. Building AI builders: Playbook for closing the AI knowledge-capability gap (aws.amazon.com)
2. Building advertising for the way people use AI (openai.com)
3. @langchain: You can now build an agent that: 🔎 Finds real products 🛒 Prepares purchases 💳 Pays …with a person approving every order 👨‍💻

Ask in @slackhq, review the order, approve in @Stripe's (x.com) 4. LegalOn halves Codex costs while maintaining development speed (openai.com) 5. How Oracle turns days of work into minutes with ChatGPT and Codex (openai.com) 6. Building Reliable Data Analytics Agents: Lessons from the KDD Cup (developer.nvidia.com) 7. We’re supercharging Copilot on Windows with Hybrid Intelligence. With your permission, Copilot can tap into the context on your PC, take action for you, and use local models when it makes sense, giving you more capability while helping your tokens go further. [Read more] (linkedin.com) 8. Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents (blogs.nvidia.com) 9. AutoSynthData: Generating Training Data for Enterprise Agents (huggingface.co) 10. GitHub Copilot brings on-device AI coding to new Windows PCs (commandline.microsoft.com)