What does OpenAI’s usage policy update mean for your business?

By Andrej Ruckij · October 28, 2026 · updated November 4, 2026

TL;DR: OpenAI’s usage policies published on October 21, 2026 apply from November 20, 2026. Less is new than the coverage suggests. A ban on automating high-stakes decisions without human review has been in the policy since October 2025, and OpenAI’s business contract already banned using outputs to build competing models. Three things did move. The human-review rule now covers any automated step that acts on a person (actions, transactions, submissions), not only “decisions”. The training ban now spells out labels, rubrics, rankings and data filtering, whether you or a contractor does the work. And any service that gives third parties direct access to OpenAI models needs OpenAI’s prior approval, with no published application route as of November 1. OpenAI’s bar is also lower than Anthropic’s, which additionally requires telling the person that AI was involved, and the EU AI Act treats a CV-ranking tool as high-risk from December 2, 2027, even with a person approving every result.

What changed on October 21, word for word

The page’s changelog entry for October 21, 2026 says OpenAI updated the policies “to clarify restrictions on using OpenAI models and outputs to train other models”, “clarified protections for children and teens and requirements for high-stakes uses”, and added “approval requirements for national security and law enforcement uses, as well as for model-access intermediaries.” It ends: “These changes are effective November 20, 2026.” On November 1 that was still the latest entry.

Set side by side with the archived October 2025 version, this is what moved for a business:

Topic	Old version (effective October 29, 2025)	New version (effective November 20, 2026)
High-stakes automation	”automation of high-stakes decisions in sensitive areas without human review"	"automation of high-stakes actions in sensitive areas without human review. This includes actions, decisions, transactions, submissions, and other steps that materially affect a person’s rights, safety, access to essential services, or other consequential interests”
Areas listed	critical infrastructure, education, housing, employment, financial activities and credit, insurance, legal, medical, essential government services, product safety components, national security, migration, law enforcement	the same areas, with law enforcement merged into “national security, intelligence, and law enforcement” and intelligence newly added
Licensed advice	”provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional"	"building or offering AI tools that provide others with tailored advice that requires a license, without appropriate involvement by a licensed professional. You also may not state or imply that the AI itself holds professional credentials or authorization.”
Training other models	no section in the usage policies	a new “Training and developing models” section (details below)
Direct model access	nothing in the usage policies	”If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval.”

Two of these matter more than they look. The licensed-advice line now targets whoever builds the tool, so an agency that ships a legal or medical advice bot for a client is the one “building or offering” it. And the change from “decisions” to “actions, decisions, transactions, submissions, and other steps” closes the reading that an agent filing a form or sending an email wasn’t making a “decision”.

My read: the human-review rule is the one that catches small companies, because “screen the CVs and reject the weak ones” is one of the most common automations people ask for. The intermediary rule catches agencies and anyone running a model gateway. The training section catches teams quietly distilling GPT into a cheaper open model.

What was already in your contract

If you use the API or ChatGPT Business or Enterprise, you also signed the OpenAI Services Agreement (effective January 1, 2026). Three of the “new” rules were already there in contract form. Section 3.3 says the customer will not “(e) except for a Permitted Exception, use Output to develop artificial intelligence models that compete with OpenAI’s products and services” and will not “(g) buy, sell, or transfer API keys from, to, or with a third party”. Section 3.1 says the customer “may not resell or lease access to its Account or any End User Account.”

The agreement defines the Permitted Exception as using Output to “(a) develop artificial intelligence models primarily intended to categorize, classify, or organize data (e.g., embeddings or classifiers), if these models are not distributed or made commercially available to third parties; and (b) fine tune or customize models provided as part of OpenAI’s fine-tuning or other Services set forth on the Pricing Page.” The new usage policy keeps that door open: “These restrictions do not prohibit uses expressly permitted by your applicable Agreement with OpenAI, including any Permitted Exception under the Services Agreement.”

So the update mostly widens and spells out rules that already bound API customers. If an automation breaks the new training or resale wording, it probably broke the contract before October 21 too.

What counts as a “high-stakes automated step”

OpenAI’s test has three parts, and all three have to be true:

An automated step. An action, decision, transaction, submission “and other steps”, taken without a person reviewing it.
It materially affects a person. Their “rights, safety, access to essential services, or other consequential interests”.
It sits in a listed area. Employment, housing, credit and financial activities, insurance, education, legal, medical, essential government services, and the rest of the list above.

OpenAI doesn’t publish examples. Here’s how I’d sort the automations agencies and SaaS teams build most often. The OpenAI column is my reading of the policy text; the Anthropic and EU columns follow their own lists, quoted further down.

Automation	OpenAI from Nov 20	Anthropic usage policy	EU AI Act (from Dec 2, 2027)
CV screener that sends rejections on its own	needs human review	high-risk: qualified reviewer plus disclosure	high-risk (recruitment and selection)
CV ranker whose shortlist a recruiter checks and sends	has human review, so allowed	still high-risk (“screening, ranking”)	still high-risk
Loan or BNPL pre-check that declines applicants	needs human review	high-risk (“approving, denying”)	high-risk (creditworthiness)
Insurance claims bot that closes or denies claims	needs human review	high-risk (“determining the outcome of a claim”)	high-risk only for life and health insurance pricing and risk assessment
Tenant screening that rejects applications	needs human review	high-risk	not in the Annex III point I read
Agent that files a legal or government form for a person	needs human review (“submissions”)	high-risk if it prepares legal documents to “sign or file”	depends on the use
Support bot answering refund questions	outside the list, in my reading	outside the list	outside the list
B2B lead scoring for your sales team	outside the list, in my reading	outside the list	outside the list

The second row is the one to notice. Adding a person who clicks approve satisfies OpenAI’s wording. It doesn’t satisfy Anthropic, which counts ranking itself as high-risk and requires disclosure, and it doesn’t take a tool out of the EU’s high-risk category, which covers systems “to analyse and filter job applications, and to evaluate candidates” whether or not a human signs off.

What human review should look like in practice

OpenAI doesn’t define “human review”. Three other texts do, and they agree on the essentials, so I’d build to them.

Anthropic’s usage policy: “A qualified person must meaningfully review a High-risk AI Recommendation, with the authority to change or modify it, before it is delivered as advice to an individual who might rely on it or before it is used to implement a decision.” The person needs “the training or experience to evaluate the output in that field” and “remains responsible for the accuracy and appropriateness of what is delivered or decided.”

The EU AI Act, Article 14(4): the people assigned to oversight must be able “to remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias)” and “to decide, in any particular situation, not to use the high-risk AI system or to otherwise disregard, override or reverse the output”. Article 26(2) adds that deployers must assign oversight to people “who have the necessary competence, training and authority, as well as the necessary support.”

Colorado’s new law (SB 26-189), as Davis Wright Tremaine summarises it: when a consumer asks for human review, the reviewer must have “authority to approve, modify, or override the decision”, must consider “relevant and available primary evidence”, and must not “simply defer to the automated output”.

Put together, a review step that would hold up looks like this:

A named person, not a role inbox. Someone whose job already includes that decision: the recruiter, the credit officer, the claims handler.
Before the action. Nothing reaches the candidate, applicant or customer until that person has approved it.
The evidence, not only the verdict. The reviewer sees the CV, the application or the claim file next to the model’s suggestion and its reasons.
Real authority. They can edit, override or reject the model’s output, and the system doesn’t make overriding harder than approving.
A log. Who approved what, when, and whether they changed it. The EU AI Act asks deployers to keep system logs for “at least six months” (Article 26(6)), so I’d start keeping them now.

The failure mode is the rubber stamp: a reviewer who approves 200 AI rejections in ten minutes. My read is that a regulator would call that automated, and so would OpenAI if it ever looked. One cheap check: track how often reviewers change the model’s output. If the override rate is zero over a few hundred cases, either the model is perfect or nobody is really reviewing.

How this lines up with the law

OpenAI’s policy says its rules are “no substitute for legal requirements”, and for most businesses the law is the stricter layer.

Rule	Who it binds	What it requires	From when
OpenAI usage policies	anyone using OpenAI	human review of high-stakes automated steps	November 20, 2026
GDPR Article 22	anyone deciding about people in the EU using personal data	”the right not to be subject to a decision based solely on automated processing” with legal or similarly significant effects; where an exception applies (contract, law, explicit consent), at least “the right to obtain human intervention” and “to contest the decision”	in force now
EU AI Act, Annex III high-risk	deployers of AI for recruitment and selection, creditworthiness, life and health insurance pricing, education and more	competent human oversight (Art. 26(2)), logs kept at least six months (Art. 26(6)), telling workers’ representatives and the affected workers before workplace use (Art. 26(7)), telling people they’re subject to the system (Art. 26(11))	December 2, 2027
EU AI Act transparency duties	providers and deployers of AI that talks to people or generates content	disclosure and labelling duties	August 2, 2026; generative systems already on the market have until December 2, 2026 for machine-readable marking
Colorado SB 26-189	deployers of automated decision-making technology that materially influences consequential decisions about Colorado consumers	notice before use, an adverse-outcome notice within 30 days, data correction, human review on request “to the extent commercially reasonable”	January 1, 2027

Two of these dates moved this year. The EU’s Annex III deadline was August 2, 2026 until the AI Omnibus, Regulation (EU) 2026/1744, entered into force on July 27, 2026 and pushed it to December 2, 2027. And the Colorado AI Act you may have read about no longer exists in that form: in May 2026 it was repealed and replaced by SB 26-189, a narrower law built around notices and consumer rights. Enforcement sits with the Attorney General only (no private lawsuits), with a 60-day cure period unless a violation is knowing or repeated, and the effective date could still move because of xAI’s legal challenge in federal court.

I didn’t cover New York City’s hiring-audit law, California’s automated decision-making rules or Illinois; check those if you hire or lend there.

How Anthropic and Google compare

If you build on more than one model, the three policies don’t line up.

Point	OpenAI (effective Nov 20, 2026)	Anthropic (usage policy marked effective Nov 12, 2026)	Google (Gemini API terms updated Apr 28, 2026; Prohibited Use Policy Dec 17, 2024)
Human review in high-stakes areas	required for automated steps that materially affect a person	a “qualified” reviewer with authority to change the output, before delivery	bans “automated decisions that have a material detrimental impact on individual rights without human supervision” in high-risk domains
Telling the affected person	not required by the policy	required: the person “must be clearly told that AI was used to produce it”, without naming the model	not in the policy
Training on outputs	banned for models meant as “alternatives to or substitutes for” OpenAI’s; Permitted Exception for internal classifiers and OpenAI fine-tuning	banned for any AI model: “Use outputs to train an AI model (e.g., “model scraping” or “model distillation”) without prior authorization from Anthropic”	banned for models “that compete with the Services (e.g., Gemini API or Google AI Studio)“
Reselling or proxying access	prior approval for model-access intermediaries	banned through “unauthorized means, including services that route requests through consumer subscriptions or misrepresent the product or client being used”	I found no general resale rule in the Gemini API terms; Grounding with Google Search results can’t be resold

Anthropic is the strictest on both points that matter to builders: every high-risk recommendation needs disclosure, and its training ban isn’t limited to competing models. If your stack switches between providers, I’d design the review step and the disclosure to Anthropic’s standard and you’ll clear the other two.

The training ban: what’s in and what’s out

The new section bans using OpenAI models or outputs “to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services”. It lists three groups:

“Creating or preparing training data, training tasks, rubrics, or reinforcement-learning environments.”
“Providing feedback used in training, including labels, scores, rankings, preferences, rewards, or corrections.”
“Selecting, filtering, or validating training data, tasks, or trajectories.”

The scope lines are broad: “These restrictions apply whether you perform the work yourself, for another party, or through someone acting on your behalf, even if you do not train the resulting model. They also apply when outputs from product use or evaluation are later used or supplied for those training purposes.” The allowed side: “research, software development, model evaluation and benchmarking, and product development, within these boundaries.”

OpenAI’s stated reason is safety: “Training another model on OpenAI outputs can transfer capabilities without carrying those safeguards over.”

How I’d read common cases (my reading, not OpenAI’s guidance):

Fine-tuning a GPT model through OpenAI’s own fine-tuning: allowed; it’s part (b) of the Permitted Exception.
An internal classifier or embedding model trained on GPT labels, never sold or shared: allowed under part (a) of the Permitted Exception.
Training an open model on GPT outputs so you can drop your GPT calls: out. That model is a substitute by design.
Using GPT as a judge to score your own model’s answers, then training on those scores: out; “scores, rankings, preferences, rewards” are named.
Using GPT to benchmark vendors before you buy: allowed as evaluation, as long as the scores don’t later feed training.
Selling a GPT-generated dataset to a client who trains a model: out, because “for another party” and “even if you do not train the resulting model” both apply.

Not everyone thinks these bans are fair. Microsoft’s Satya Nadella called it ironic that labs train on public data while banning distillation, and wrote: “In consuming intelligence, you are creating intelligence.” GovInfoSecurity notes that Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg have defended distillation. None of that changes what your account can be suspended for.

Reselling: what the approval involves

What the policy says is short: an intermediary needs “our prior review and approval”, and the rule “does not apply solely because an application integrates OpenAI models to power its features or functionality.”

What it doesn’t say is how to apply. As of November 1, the policy page has no application form or criteria (the form it links to is for update notifications), GovInfoSecurity’s report describes no process, and I found no published timeline. OpenAI may also grant case-by-case exceptions, weighing “the partner, proposed use, applicable safeguards, and the partner’s compliance commitments”, and says such approvals “may require written terms”.

One reason OpenAI cares about routers: in its February 2026 memo to the House Select Committee on China, as GIGAZINE reported it, OpenAI said accounts linked to DeepSeek employees reached its models through third-party routers and hid where the traffic came from.

How I’d sort agency and SaaS setups:

Clearly in: a pass-through endpoint on your API key that clients call directly; a model router or AI gateway you sell; a white-label “AI API” plan; handing clients API keys you pay for (which section 3.3(g) of the contract already bans).
Clearly out: a support bot, summariser or document tool where OpenAI powers your feature.
Grey: a client portal with a free-text prompt box and a model picker, where the client is in effect using GPT through your account. My read is that it looks closer to direct access, so I’d ask.

If you’re in the first or third group: write one page on what you offer, who the end users are, and how you stop abuse (rate limits, per-client keys, logging, content filters), then send it to your OpenAI account contact or through Contact Sales on openai.com before you offer direct access; the rule applies from November 20, 2026. Keep the written answer.

Enforcement so far

OpenAI’s transparency page describes how enforcement works: “classifiers, reasoning models, hash-matching, blocklists, and other automated systems”, then possible “human review”, with actions ranging from warnings to “Terminating or limiting access to our products”, and an appeals form.

The public cases I found are all about distillation:

A small builder, July 2025. A developer posted on OpenAI’s community forum that their account received this notice: “We have identified ongoing activity in your OpenAI account that is not permitted under our policies for: Distillation”. They said they were building an orchestrator with repeated JSON calls for client automations. A forum moderator replied that fine-tuning a smaller model on GPT outputs is documented and allowed, and pointed them to support. The thread ends unresolved. This is the failure mode that matters for automation builders: high-volume, repetitive, structured calls can look like extraction.
DeepSeek, February 2026. OpenAI told the House Select Committee on China that DeepSeek used “unfair and sophisticated techniques to extract results from leading American AI models”, per GIGAZINE’s report.
Anthropic, February 2026. Anthropic reported that DeepSeek, Moonshot AI and MiniMax “generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts”. It responded with detection classifiers and tighter verification for educational, security-research and startup accounts, which it called the pathways most often exploited.
US government. GovInfoSecurity reports that CISA warned about DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai over distillation of OpenAI and Anthropic models.

I found no public case of OpenAI enforcing the human-review rule or the intermediary rule. The second applies only from November 20, 2026.

What to do with this

For: businesses with automations that call OpenAI through the API or through Zapier, Make, n8n and similar tools; agencies and freelancers who build AI tools for clients; SaaS teams with a model picker or an API; teams that fine-tune or train their own models. Not for: people who type into ChatGPT and read the answer. The rules apply to them too, but nothing here changes how they work.

The checklist, in the order I’d run it:

#	Step	Owner	Time	How to check it’s done
1	List every automation that sends data to an OpenAI model, including no-code tools. One line each: input, output, what happens next.	ops or tech lead	30-60 min	every workflow that calls OpenAI has a line
2	Tag each line: does the next step act on a person in a listed area (employment, credit, insurance, housing, education, legal, medical, government services)? Use the table above.	product owner, with whoever knows the client’s domain	30 min	each line marked “acts”, “recommends” or “out of scope”
3	For every “acts” line, add an approval step: the model ranks or drafts, a named person approves before anything is sent, submitted or decided.	developer	1-4 hours per workflow	run one real case; nothing reaches the person until someone clicks approve
4	Give reviewers the evidence and the power to override, and log decisions (reviewer, time, changed or not).	the business owner of the decision (HR lead, credit manager) plus developer	1-2 hours	the log exists; track the override rate monthly
5	Write a one-line disclosure for affected people (“this assessment was prepared with AI and reviewed by…”). Required on Claude from November 12, 2026; it also prepares you for the EU and Colorado notice duties.	marketing or legal	30 min	the line appears in the email or screen the person sees
6	Ask your data or ML lead: do any GPT outputs become training data, labels, scores or filters for another model? If yes, is that model meant to replace GPT, or shared outside the company?	data or ML lead	15-30 min	written yes/no per pipeline; a “yes, substitute” pipeline is stopped or covered by the Permitted Exception
7	Check whether any client reaches OpenAI models directly through you. If yes or unsure, send OpenAI the one-pager described above.	founder or account owner	15 min, plus waiting for a reply	written answer from OpenAI on file
8	Agencies: add a clause to client contracts naming who the human reviewer is and who answers for policy compliance.	agency owner	1 hour	signed by each client running a high-stakes workflow
9	Sign up for policy-update notifications via the form on the usage policies page.	whoever owns vendor compliance	2 min	confirmation received

If you’re building agents that act on their own, the wider question is what limits you put on them. I collected the patterns in the wiki page on glossary/guardrails.

Sources

More of this every day on the channel: https://t.me/aidaily_business