Guide OpenAI · Chapter 9

    Who does what so that AI with OpenAI stays in operation.

    A business team keeps the test set, one person owns every agent, a directive says which apps are allowed, and your IT keeps roles, data and cost in view.

    As of: October 2026 · Editor: Marcel Haas

    • Frontier Firm
    • AI directive
    • Test set
    • Codex
    • Workspace agents
    • Support & Operations
    01

    The Frontier Firm designs workflows for people and agents from the start.

    In the morning the night's enquiries lie sorted in the workspace, a dot has prepared the answers, Codex has left drafts as pull requests for three tickets, and an agent on the Agents API has checked the incoming invoices against orders and put two forward for a decision. That is what the Frontier Firm looks like, and that is how we understand AI-native work: a workflow is designed for people and agents from the start, with a person leading a team of agents, and with the prospect that over time one agent carries the largest part of a workflow. The road there starts small, with a first case, a business team that owns it, and a sheet that says who decides what. The strategy before it, from the goals of the management to the order of the areas, is worked out with you by KA Partners; the implementation with ChatGPT, Codex and the OpenAI platform is ours, and both sides work from the same plan.

    02

    Owner, business team, data protection and platform carry every application.

    Every agent and every application has an owner: for a workspace agent or a dot it is the person or team lead who runs it, for an application on the API the person who owns the project. They decide on changes, answer approvals, read usage and traces and hand over to a named person on leaving. The business team keeps the test set: it collects real questions and tasks, notes the expected answer and the source, rates the answers and adds the questions the agent could not solve. The data protection office countersigns the risk sheet, demands the impact assessment where the risk is high, keeps the application in the record of processing activities and receives the evidence from the Compliance API. And the platform owner in IT keeps the admin console with SSO, SCIM, roles and Connector Registry, the Codex settings, the projects with spend limits and data regions and the platform changelog in view. These names stand on a sheet per application, and where one is missing, the agent does not start.

    03

    The directive names the allowed apps, and the workspace is the offer against shadow AI.

    At the centre of the AI directive is a table of the approved tools, their contracts and the highest sensitivity level that may be entered there. For OpenAI it lists the workspace with company sign-in, the apps the Connector Registry allows, the rules for Codex Cloud and the groups that get a dot; added to that is the labelling when an agent writes to customers, and who reviews an agent's result before it is passed on. To shadow AI the directive answers with offer and boundary: the offer is the workspace itself, in which staff work under the company sign-in, the plan's data controls and the directive, with apps that show only what the person may open themselves. The boundary is set by the data loss prevention rules on managed devices, which warn when confidential content is pasted into AI websites outside the workspace and block it for personal data, with the pointer to the directive. Whoever makes the approved route convenient rarely needs the ban.

    04

    Training happens on real tasks: the good brief, the source, the handover.

    On their own tasks a group learns in one session what it needs in the afternoon: writing a good brief, with goal, context, sources and the desired form; opening the reference that Company Knowledge delivers with the answer before adopting it; spotting an invented figure and an outdated version; and knowing when the person takes over. The development team also learns how Codex works in the sandbox, when it asks and how a draft is reviewed before it goes into the main branch; OpenAI's guide for administrators describes the steps of the rollout. In each area a champion starts ahead of the others and answers questions in their own team, the communication calendar says who sends what to whom and when, quick guides from every session land on the learning site, and in the prompt-a-thon the areas practise on tasks from their own work. The directive belongs in every session, with the cases in which it applies.

    05

    The test set is the release gate, versions and handover follow it.

    From pilot to operation lead the stages the Agent Explorer shows in chapter 3: the task, the test set, the pilot, the release, the operation and the evolution, and every stage ends with your decision. At OpenAI the test set runs in your application against the Responses API or the Agents SDK, with traces per step; because the Evals platform shuts down on 30 November 2026 and existing evals become read-only from 31 October 2026, the evaluation belongs in your repository, as part of the code Codex helps maintain. A version goes live only when it scores no worse than the previous one and the owner approves, and every change to model, instructions, sources or tools is a version; a model switch demanded by a retirement takes the same route. The handover into operation is a sheet: project with key, spend limit and data region, data, actions and approval levels, owner and deputy, test set, cost per case, the retirement dates of the models and the date of the next review. For an agent on the Agents API it also says who stops it.

    06

    Support after handover and a fixed review rhythm keep the operation going.

    If the agent answers wrongly on Tuesday, on Tuesday someone knows what to do: the champion in the area takes the report, the owner decides, and our Support & Operations team checks trace, sources and version, keeps the spend limits and failed runs in view and installs updates. The review rhythm is in the sheet: the platform owner reads the changelog and the deprecations page of the platform continuously, because a day like 26 August 2026, when the Assistants API was shut down, hits every integration nobody maintains; the retirement dates of the models from chapter 2 are in the data sheet; facts, plans and prices we check every quarter; and usage per group comes from the workspace, evaluated under a data rule that the data protection office and the staff committee sign. What the quarterly review puts on the table:

    • Test set and evaluation: which questions the agent solves, where it hands over, which sources are out of date and what belongs in the next release.
    • Cost per case: seats, tokens, runtime and hours from chapter 7 against the limits, and whether the agent gets more people.
    • Rights and rules: owner, roles, tool list, approval levels and Codex policy, checked against the risk sheet.
    • Facts and plans: retirements, plan changes and new features from changelog and deprecations page, with the date of the review.

    What you decide.

    Who owns the agent after the pilot?

    Options

    1. 1IT, because it administers workspace and projects
    2. 2An owner in the business team, the platform owner in IT, and a second line for support
    3. 3The project team that built it

    Our recommendation: An owner in the business team, because they decide what the agent may do and whether its answers hold up; for a dot it is the person who runs it, for an application on the API the person who owns the project. The platform owner in IT keeps admin console, projects, limits and changelog; the second line, with you or in our Support & Operations team, checks trace and version. Both names and a deputy stand in the handover sheet before the pilot ends.

    What we advise against: The project team or nobody. A project team dissolves, and an agent whose project nobody owns keeps running until the spend limit stops it.

    Ban private ChatGPT accounts or offer the workspace?

    Options

    1. 1Block AI websites on managed devices
    2. 2Tolerate private accounts
    3. 3Offer and boundary: a seat in the workspace for everyone who needs it, a warning for confidential content, a block for personal data

    Our recommendation: Offer and boundary. Whoever uses ChatGPT for work gets a seat in the workspace with company sign-in, the plan's data controls and the apps from the Connector Registry; the directive names what may go in there, and the training shows it on real tasks. On managed devices the data loss prevention rule warns when confidential content is pasted into AI websites outside the workspace and blocks it for personal data. The usage in the workspace shows whether the route is taken.

    What we advise against: A pure ban, because the work then moves to the private phone, and equally tolerating private accounts: without company sign-in, without SCIM and without the Compliance API.

    When does a new version go live?

    Options

    1. 1When the developers say it works
    2. 2When the test set in your repository passes with no regression against the previous version and the owner approves
    3. 3On fixed dates

    Our recommendation: When the test set passes and the owner approves. The evaluation runs in your repository against the Responses API or the Agents SDK, with traces per step, because the Evals platform is being shut down; every change to model, instructions, sources or tools counts as a version, a model switch after a retirement as well. Agents that read documents or web pages additionally go through the attack cases of the test set, and the predecessor stays ready until the new version has shown in operation what the test set promised.

    What we advise against: Rolling out a version because OpenAI released a new model. A new model answers differently, and that shows first in the test set.

    How often do we review facts, plans and rights?

    Options

    1. 1When something stands out
    2. 2Per quarter per application, plus changelog and deprecations continuously and the cost monthly
    3. 3Once a year

    Our recommendation: Per quarter per application, with a fixed date in the sheet: test set and evaluation, cost per case, rights and rules, facts and plans. The platform owner reads the changelog and the deprecations page continuously, the spend limits report monthly or earlier, and the retirement dates of the models are in the data sheet. The record of processing activities is reviewed by the data protection office once a year.

    What we advise against: Once a year only. The Assistants API was shut down on one day, and whoever reads about it in the annual review has lived through the outage.

    Status and sources.

    As of: October 2026

    • As of October 2026

      Codex Cloud starts switched off in Enterprise workspaces.

      Administrators enable use through roles, define environments, secrets and internet access, and distribute rules for local runs in a requirements.toml file. Codex Cloud respects the repository permissions of the source system.

      Source: OpenAI: Codex admin rollout guide
    • Since 29 September 2026

      OpenAI brings dots.

      Dots keep working towards goals in ChatGPT on their own cloud computer and are reachable by chat, voice, Slack or Teams. In Switzerland through Business Premium and as an Enterprise beta, not for Pro accounts at launch.

      Source: OpenAI Help Center: Getting started with your dot
    • Since 26 August 2026

      Assistants API shut down, Agent Builder next.

      OpenAI removed the Assistants API on 26 August 2026; the successors are the Responses API and the Conversations API. Agent Builder and the Evals platform shut down on 30 November 2026, and existing evals become read-only from 31 October 2026. ChatKit remains.

      Source: OpenAI: Deprecations
    • Since 15 April 2026

      The Agents SDK brings a harness and a sandbox, first in Python.

      The Agents SDK runs the loop of model call, tool call and handoff to further agents inside your application, with guardrails for inputs and outputs and traces for every step. Since April 2026 it brings a sandbox for working with files and tools, first in Python.

      Source: OpenAI: The next evolution of the Agents SDK
    • As of 12 October 2026

      Apps and Company Knowledge in Business and Enterprise, the Connector Registry in Enterprise.

      With apps, ChatGPT reads documents from SharePoint and Outlook, issues from GitHub or data from your own systems, and Company Knowledge answers questions across several connected sources; both belong to ChatGPT Business and Enterprise. Which apps are allowed in the workspace is set by the administration, and Enterprise lists them in the Connector Registry. In an app, ChatGPT sees only what the signed-in person may open there.

      Source: OpenAI: Business plans
    • As of 12 October 2026

      ChatGPT Enterprise brings SCIM, role-based access and the Compliance API.

      Business and Enterprise give teams a workspace with sign-in through SAML SSO; Codex is included in both plans. Enterprise adds user provisioning through SCIM, role-based access, the Compliance API for audit and eDiscovery, and the Connector Registry.

      Source: OpenAI: Business plans

    Frequently asked questions.

    Who writes the AI directive for ChatGPT and Codex?

    The platform owner in IT and the data protection office write it together, the management signs. For OpenAI the table lists the workspace with the highest sensitivity level that may go in, the apps from the Connector Registry, the rules for Codex Cloud and the groups that get a dot, plus the labelling for agents in customer contact and who reviews results. It rarely needs more than two pages.

    What do we do about private ChatGPT accounts in the team?

    Offer the workspace: a seat with company sign-in, the plan's data controls and the allowed apps for everyone who uses ChatGPT for work, and the training shows the route on real tasks. Added to that is the boundary on managed devices, which warns when confidential content is pasted into AI websites outside the workspace and blocks it for personal data. The usage in the workspace shows whether the route is taken.

    How do we test once the Evals platform is shut down?

    In your repository: the business team's test set runs as code against the Responses API or the Agents SDK, with traces per step and a score per case, and Codex helps maintain it. Existing evals become read-only from 31 October 2026, which is why we move them into the repository before then. So the release gate stays independent of whichever tools the vendor retires next.

    How do we hand an agent on the Agents API over into operation?

    With a sheet that names project, key, spend limit and data region, plus data, actions and approval levels, owner and deputy, test set and evaluation, cost per case, the retirement dates of the models and the date of the next review, and with the person who stops the agent when it is stuck in a loop. The second line knows trace, sources and version, and the champion in the area knows where a report goes.

    What does KA Partners do in an OpenAI initiative, and what do we do?

    KA Partners works out the AI strategy with your management: goals, the order of the areas, roles and the directive, independent of the vendor. We implement: the workspace with Entra sign-in, the first case with ChatGPT, Codex or an agent on the platform, test set and approvals, training on your tasks and the operation afterwards. Both work from the same plan, and in the first conversation we establish where your initiative stands.

    How do we notice that the workspace is being adopted?

    By the sign in daily work that every area head set in the adoption plan, for example that drafts come into the meeting with a reference rather than a guess, and by the usage per group from the workspace, evaluated under the data rule. The usage tracker shows where the rollout is holding and where more support helps, and the test set shows whether the agent is getting better.

    How often do we review models, plans and prices at OpenAI?

    Per application every quarter, with a date in the sheet: facts, plans, prices, rights and test set. In between, the platform owner reads the changelog and the deprecations page of the platform, and the spend limits report on their own. The dated articles of this guide carry the date of their last review, and the chronicle shows what has changed since.

    To the guide

    Understanding OpenAI, from ChatGPT to the Agents API.

    The guide shows what ChatGPT, Codex and the OpenAI platform take on, where your data runs and what your IT decides, with date and source.