OpenAI DevDay 2026 for procurement: six practical takeaways

OpenAI wrapped DevDay 2026 on 29 September with more than 20 announcements. Among them: dots for ongoing work, a Decisions API for fixed-choice questions, easier plugin creation, Private Intelligence, and ChatGPT in Slack and Microsoft Teams. Its new Marketplace may also let eligible enterprise customers spend part of an existing OpenAI commitment on partner software. OpenAI's DevDay recap.

Four core changes and two supporting tools deserve procurement's attention. They could help track contract value and savings, route purchase requests and reuse sourcing methods. None is a procurement product; each proposed workflow needs configuration, testing and a named owner. For a wider view of workflows already in use, see our procurement AI use cases guide.

The short version: Start with work a person can check. Let AI gather evidence, flag exceptions, prepare drafts and route requests. People retain authority over supplier selection, commercial commitments and external actions. Build that boundary into the workflow and the contract.

Dots are ongoing agents with a cloud computer and access to permitted apps. They can work in the background while users inspect progress and set rules for actions. Specialist dots with their own identities and system access remain in focused enterprise pilots. Check which version your organisation can use before designing a shared workflow. OpenAI's dots announcement.

What this means for procurement

Start with contract value. Give an authorised dot the contract, POs and invoices. Ask it to compare volumes and prices with agreed discount tiers, rebates and credits. If volume crosses a threshold, it could flag invoices still using the old rate and link to the clause, PO and invoice lines. The category owner checks units, dates, amendments and credit notes before raising a claim.

It could also watch savings targets. Give it the approved initiative register, baseline method, target dates and actual spend. It could flag initiatives with no visible spend change, delayed supplier changes or reported savings that no longer match the baseline. A finance or procurement owner decides what counts as realised savings. Each alert needs its calculation and source records.

Both watches depend on matching contract terms to the right supplier, item and transaction. Test amendments, mixed units, credit notes and partial periods. A dot's proactive research can read permitted sources, but cannot directly send messages or change connected content; later actions have separate checks. OpenAI's dots FAQ. Measure valid exceptions and reviewer time, alongside hours saved.

The Decisions API answers developer-defined questions from a fixed list of choices. It accepts text or image context and is designed for classification and routing. OpenAI announced a limited preview at DevDay, with broader release planned for the following days. DevDay recap.

What this means for procurement

Consider the first minute after someone submits a purchase request. The requester has chosen a category and written a free-text description, often with missing scope or evidence. Procurement has to send it to the right owner or return it for detail. Threshold and category rules handle clear cases. They miss requests whose description conflicts with the selected category.

A team building an intake tool could ask the API three bounded questions: Is the request complete? Does it need security review? Which procurement queue should receive it? The tool could request missing information or route a complete submission while the requester is still in the process. OpenAI cites request routing as an intended use; this procurement workflow is our proposed application. DevDay recap.

Test whether it cuts clarification loops and gets requests to the right reviewer faster. Use past requests involving company data, several categories or incorrect requester labels. Log each input and route so staff can correct mistakes. Supplier selection and commercial approval stay with authorised people.

DevDay brought Plugin Creator, a revised submission flow and better plugin recommendations. A plugin can package instructions, resources and connections to external tools. OpenAI lists creation and discovery across plans; users choose plugins and approve access. DevDay recap; plugin documentation.

What this means for procurement

An RFP plugin could carry the approved checklist, templates and review steps. It could take the scope, evaluation criteria and timetable, then prepare a draft pack and questions for the category manager. Give the checklist an owner and version before adding live data connections.

Supplier due diligence needs more care. With an approved connection to supplier records, a plugin could identify the right entity, flag missing evidence and prepare a review file. Specify the records it can read, the fields it can change, whose account it uses and how to revoke access. Directory discovery does not replace company approval of the plugin and its permissions.

We are preparing an OpenAI version of Molecule One's Procurement OS plugin. Request the OpenAI Procurement OS plugin with your name and email, and we will send it when it is ready.

Private Intelligence covers new data controls for business use. OpenAI announced Zero Data Retention with Private Safety Processing for API workflows and previewed Private Inference for later this autumn. The first uses customer-controlled storage for automated safety review without giving OpenAI personnel access to the underlying content through that review. DevDay recap.

What this means for procurement buyers

Organisations with strict privacy requirements may now have a better route to deploy AI. The buying work is to trace where data goes: which product and endpoint receives it, which connected services can access it, who sees the output and what remains afterwards. OpenAI's API data controls say standard abuse-monitoring logs may hold customer content for up to 30 days. Zero Data Retention needs approval and has endpoint-specific limits. The Responses API keeps application state for at least 30 days when stored. Third-party MCP servers apply their own retention policies. An API setting does not automatically cover a dot, meeting page or connected supplier system.

Zero Data Retention with Private Safety Processing still keeps selected safety records in customer-controlled cloud storage for 30 days. It must be enabled and verified for each API project. OpenAI's technical guide sets out the storage and operating requirements. Get the terms for the exact product, model and endpoint in writing, then test with approved data.

Employees can mention @ChatGPT in a Slack or Microsoft Teams thread. It can use the thread and connected tools that an admin or user has authorised; colleagues can refine the result together. OpenAI lists the feature for Business and Enterprise. Replies are visible to people in the conversation, while source files retain their own permissions. DevDay recap; product guide.

What this means for procurement

It could ask a requester for missing scope or budget information in the original thread. In a deal thread, it could keep a draft list of open legal, security, finance and commercial questions. The deal owner decides what to share and what to commit to a supplier.

The Meetings plugin takes notes from calls and in-person conversations, then saves a summary and suggested actions in ChatGPT Space. Users can edit those actions before starting follow-up work. It is in beta on macOS for Pro and Business, with limited Enterprise alpha access. The user must obtain participants' consent. OpenAI says it deletes audio once the notes are ready; buyers should check retention for the page and transcript separately. OpenAI's Meetings guide.

What this means for procurement

Across departments, the plugin can draft a record of decisions, owners and open questions after a meeting. The meeting owner checks commitments and owners before sharing or acting on the notes. When buying, assess it as a collaboration tool with its own consent, access and retention terms.

If we were running a procurement team today: five moves this quarter

Choose one small workflow. Record errors, review time and cost before expanding access.

1. Pilot a contract value or savings watch. Name an owner and include difficult records such as amendments, credit notes and mismatched units. Record the source, proposed action, reviewer and correction.
2. Map connected accounts and permissions. Record each dot's or plugin's identity, permitted actions and offboarding process. Disconnecting a plugin stops new access but does not erase information already in a dot's context. Dots FAQ.
3. Set approval points. Require a person to approve supplier messages, notices, material tracker changes, selections and commitments. Specify any advance approvals and the audit record. Dots permit some advance approvals; certain actions require the user to take over. Dots FAQ.
4. Get a retention schedule. Cover every product, endpoint, connected service, transcript, page and output file. Confirm what deletion removes: deleting a dot does not delete its separate files and conversations. Dots FAQ; API data controls.
5. Price the whole workflow. Include the plan, deeper-work allowance, API usage, connected apps, setup, review and maintenance. Dot conversations sit outside ChatGPT usage limits; tasks started in Codex or ChatGPT Work count towards them. Dots announcement. For Marketplace, confirm eligible partner products, the usable commitment and the effect of direct partner contracting on renewal leverage. Marketplace FAQ.

Expand only when reviewers can verify the source and approval behind each action, and the total cost holds up. If you need to choose a first workflow and design a controlled pilot, our procurement AI consulting team can help.