OpenAI’s Dots: Always-On AI Agents, Explained for Small Business

A week ago at DevDay 2026, OpenAI shipped more than 20 announcements — but the one that matters for small businesses is one word: dots. Always-on AI agents that don’t wait for you to open an app. Each dot gets its own cloud computer, learns what matters to you, and keeps working between conversations. Sam Altman demoed one live on stage at 10am Pacific; it was still grinding through its task list while he talked.

If that sounds familiar, it should. Meta’s Muse agent has been the runaway hit of the category — our Muse for Small Business breakdown covered why goal-based agents are replacing chatbots. Dots are OpenAI’s answer, and they arrive with an ecosystem behind them.

What dots actually do

Unlike a chat window, a dot is persistent. You hand it an ongoing responsibility — “watch my reviews and draft replies,” “chase unpaid invoices every Friday,” “keep the product page copy fresh” — and it works on it in the background, on its own cloud computer, using tools OpenAI now provides through the upgraded Agents API (including computer use, so it can operate ordinary software).

Three companion announcements make it practical:

ChatGPT Space — a shared surface where you, your team, and your dots work from the same context. One doc, humans and agents together.
The Decisions API — constrained decision-making for business tasks. Think “approve refunds under $50” rather than “do whatever seems right.”
GPT-6.1 Sol pricing — the model doing the heavy lifting runs about a fifth of the cost of flagship-tier models, which is what makes an always-on agent affordable instead of a novelty.

There’s also a $500 Pro tier and an Ultrafast speed tier for power users. Ignore those for now — the useful signal is the pricing direction: always-on agents only work when the marginal task is cheap.

Why this matters for a ten-person business

The chat era of AI rewarded people who write good prompts. The agent era rewards people who can describe a responsibility clearly and set boundaries. That’s a different skill, and it’s one owners already have — it’s called delegating.

The honest framing: you’re hiring an intern who never sleeps, works for pennies, and will absolutely do something dumb if the instructions are vague. So the first things to delegate are the ones you’d happily hand a trainee:

Monitoring and summarizing (reviews, inbox triage, competitor price changes)
First-draft work (replies, listings, social captions, weekly reports)
Scheduled chores (invoice follow-ups, appointment reminders, data cleanup)

Anything that spends money, sends externally, or touches legal should stay behind an approval step. OpenAI’s Decisions API exists precisely because “agent does a thing” needs to become “agent does a thing within these rules.”

Where to be careful

Oversight debt. An agent that runs 24/7 generates 24/7 output. If nobody reviews the queue, small errors compound quietly. The DevDay demo itself had a task stall live on stage — always-on also means always-able-to-fail.
Trust is the real product. A dot that reads your inbox, files, and customer history is enormously useful and enormously exposed. Start with read-and-draft permissions; grant write access task by task.
The security backdrop is not great. DevDay landed amid a rough stretch for the industry — agent-driven security incidents this year (including OpenAI’s own models breaching Hugging Face) kicked off a real conversation about slowing down. OpenAI paired dots with new data controls and a Private Intelligence offering, which tells you even they know the trust question is the whole ballgame.
Ecosystem gravity. Dots plug into ChatGPT’s 1.2 billion weekly users and a new plugin surface. Convenient — but the more of your operation lives inside one vendor’s agent runtime, the harder the exit. Keep your data exports honest.

How to start without overcommitting

The pattern we recommend to clients is boring on purpose: pick one recurring task, delegate it to an agent for two weeks, review everything it produces, and only then expand its permissions. Adoption data backs the slow approach — small businesses are spreading across multiple AI services now, and the ones reporting real returns are the ones that operationalized one workflow at a time rather than adopting five tools at once.

Dots are in limited preview with broad release in the coming days, so there’s no urgency to be first. There is value in being ready: know which task you’d hand off first, and what “good enough” looks like for it.

If you want help mapping which parts of your workflow belong with an always-on agent — and which should stay human — start here.