Assessing ChatGPT Dots for business is harder than it looks, because the product is days old and there are no independent deployments to point at. What does exist is specific: five role-based workflows OpenAI published, five internal functions it says it has been running dots on, and a set of controls that decide whether any of it is safe to try. This piece sticks to those.
Updated October 2026. Dots launched on 29 September 2026 and are rolling out gradually, so check OpenAI own pages for the current position before relying on any detail here. Nothing here is a customer case study, because none has been published yet.
The five workflows OpenAI documents
Each of these is published as an illustration of how a dot is meant to be used, not as a measured result. Read them as the shape of the tool rather than as evidence it works at scale:
- Engineering. Watching customer feedback for recurring requests, scoping smaller fixes, building and testing them, and returning complete pull requests with videos of the change.
- Product marketing. Learning your audience, positioning and creative standards, then revising launch materials when the scope of a product changes.
- Research. Rerunning analyses as new data arrives, investigating unexpected results, updating figures and flagging what needs review.
- Sales engineering. Checking customer requirements and account history against product docs, building a proof of concept for a key integration, and updating the proposal as test results change.
- Content. Turning a new transcript into clip candidates, show notes and draft posts, then carrying your edits across all of them.
What OpenAI runs dots on internally
More telling than the illustrations is the list of functions OpenAI says it has been learning from in its own early testing: procurement, invoice processing, email marketing, customer support and commercial contracting. Four of those five are back-office processes with clear rules and an audit trail, which is a useful signal about where this technology fits first.
The announcement also describes dots inside OpenAI starting to investigate a bug as soon as it appears in Slack, and turning a new design into a working app while the team stays on customer feedback.
Where the value actually comes from
In every documented example the dot does the gathering, drafting and re-checking, and a person makes the decision. That is the part worth designing around. A process where the expensive step is assembling context is a good candidate; a process where the expensive step is judgment is not. For the general pattern, see our explainer on agentic AI.
What to pilot first
A sensible first pilot has four properties, each of which maps onto a control OpenAI actually provides:
- It is read-only to begin with, so proactive research and scheduled checks carry no risk of an unwanted action.
- It touches one connected app, not six, because every app you connect widens what the dot can reach and what it can form memories from.
- It produces a draft a named person approves, which is what the Custom Rules setting hand off to you is for.
- It is reversible. Keep anything that spends money, signs something or messages a customer behind an approval until you have watched it work.
The questions to settle before you start
Three practical matters decide whether a business pilot is even possible. Dots are available on Pro, excluding the European Economic Area, Switzerland and the UK, and on Business Premium across supported regions; Enterprise, Edu and Healthcare can try the beta only once an admin enables it, and it is off by default. Data handling differs by plan: content from Business, Enterprise and Edu workspaces is not used to improve OpenAI models by default. And the controls themselves are an admin decision, covered in our guide to dots admin controls.
If your interest is a dedicated agent that belongs to a team rather than a person, that is a different product tier, covered in our guide to specialist dots.
The honest caveat
OpenAI states plainly that dots can make mistakes and that consequential work needs reviewing. Its own system card reports a measured rate of severe misalignment of 0.84 percent in a deliberately difficult evaluation of realistic work environments, and that the flag rate for moderate scope violations roughly doubled, from 8.6 percent to 19.7 percent, when the number of intervening tasks went from five to ten. Those numbers come from adversarial tests rather than production, but the direction is the useful part: the longer the chain, the more supervision it needs.
What ChatGPT Dots for business will not fix
Three limits are structural rather than teething problems, and a pilot that ignores them will fail for reasons that have nothing to do with the technology.
The first is that a dot inherits your access, not your authority. It can assemble the case for a decision; it cannot make the decision stick in an organisation where the decision was always the hard part. The second is that connected apps are the ceiling. A dot is only as useful as the systems it can reach, and the business processes most worth automating are often the ones living in the system nobody has connected to anything. The third is that memory is per dot and per person at launch. One dot belongs to one user; institutional knowledge does not accumulate in it on behalf of a team.
That third point is precisely what specialist dots are meant to address, which is why they are a separate pilot programme rather than a setting.
How to measure a ChatGPT Dots for business pilot
Because there are no benchmarks to compare against, a pilot has to generate its own. Three measures are worth capturing from the start:
- Approval rate. What share of the work the dot brings you is accepted without change. A low rate early is normal; a flat rate after a month means it is not learning your standards.
- Intervention cost. How long review takes against how long the task took before. A dot that halves the work and doubles the checking has gained nothing.
- Escapes. Anything the dot did that you would not have approved. This should be rare, and every instance is worth turning into a custom rule.
Common questions
Is ChatGPT Dots for business available everywhere? No. Pro access excludes the European Economic Area, Switzerland and the UK at launch. Business Premium is available across supported ChatGPT regions.
Does OpenAI train on our workspace content? OpenAI says content from ChatGPT Business, Enterprise and Edu workspaces is not used to improve its models by default.
What is the lowest-risk pilot? A scheduled, read-only check that produces a draft for a named person to approve.
Can a dot approve a purchase on its own? Only if your rules allow it. Auto-review checks actions that affect accounts, and financial commitments are one of the threat models it is built to catch.
How many dots can a team have? One per eligible user at launch. OpenAI says the ability to add more dots is planned but has not published pricing for them.
Sources and further reading
Where the figures and rules above come from, so you can check them:
- The announcement, including proactive research and the control model: OpenAI, Introducing dots
- Appendix B, the red-teaming and alignment results for dots: OpenAI, GPT-6 Astra system card
- Setup, connected apps, scheduling, memory, rules and reset: OpenAI Help Center, Getting started with your dot
- Everything else announced alongside dots: OpenAI, DevDay 2026 Recap
Photo credits: OpenPlanRedBalloon1 by VeronicaTherese, CC BY-SA 3.0, via Wikimedia Commons. Volkswagen Invoice by Casey Serin, CC BY 2.0, via Wikimedia Commons. Horasis Global China Business Meeting 2013 (10865462556) by Richter Frank-Jurgen, CC BY-SA 2.0, via Wikimedia Commons.
Join the discussion