Thursday’s launch was less a new chatbot than a new colleague. On 9 July 2026, OpenAI released ChatGPT Work, an agent built on the GPT-5.6 family that does not answer a question and stop — it takes a brief, works across your connected apps and files for minutes or hours, and hands back a finished document, spreadsheet, deck or site. For a small business, that is a different proposition entirely from the assistant most owners have been using for the last three years.
What ChatGPT Work actually does
The agent combines the conversational side of ChatGPT with the engine behind Codex, OpenAI’s coding agent. It can read data from tools you already pay for — Slack, Google Drive, Microsoft 365 — run a multi-step workflow on its own, and keep going in the background while you close the laptop. You can watch its progress, answer its questions mid-task, redirect it, and approve the steps that matter before it commits to them.
Two details matter more than the demo. The first is Scheduled Tasks: the agent can run once, repeat on a cadence, or trigger when something changes — a supplier price list updates, a form fills, a channel goes quiet. The second is programmatic tool calling, where the model writes a small throwaway program to coordinate its own tool calls rather than pinging each one by hand. It sounds like plumbing, but it is why long, tool-heavy jobs now finish instead of timing out halfway.
Availability is staged: desktop first across plans, with Pro, Enterprise and Edu ahead of Plus and Business on web and mobile. It runs on the new GPT-5.6 line, whose three tiers we broke down in our look at Sol, Terra and Luna pricing — Terra, the everyday default, lands at roughly half the cost of the model it replaces.
Why this lands differently for a five-person business
Large companies have had agent pilots for a year. Small teams mostly have not, because the tooling assumed an engineer to wire it up. An agent that reads your existing Drive and Slack and produces a real deliverable removes that assumption. The work most likely to move first is the work that already eats an owner’s Sunday: monthly reporting pulled from three systems, proposal drafts assembled from past ones, competitor snapshots, invoice chasing, onboarding packs.
The economics are not speculative. Surveys this year put the median small business at around five AI tools in regular use, with a majority of active users reporting more than twenty hours saved a month — roughly half a full-time person redirected to work that actually needs a human. Agents compress that further, because the saving shifts from “faster drafting” to “the task ran without me.”
The project-management problem nobody demos
Here is the part the launch video skips. An agent that works for two hours unsupervised is a delegation problem, not a software problem — and delegation is where small businesses reliably struggle, with people and with machines alike.
Three habits separate the teams that get value from the ones that get an expensive mess:
- Write the brief you would give a contractor. Scope, inputs, definition of done, what it must not touch. Vague prompts produce vague deliverables, and an agent will confidently spend an hour producing them.
- Put the approval gate where the risk is. Sending an email, publishing a page, moving money, changing a shared file — those need a human yes. Reading, drafting and summarising do not. Approving everything is as bad as approving nothing; it trains you to click through.
- Log what ran. Once Scheduled Tasks are humming, a company can accumulate a dozen recurring automations nobody owns. That is how back-office automation quietly turns into back-office debt. Keep one page listing every scheduled agent, its owner and its kill switch.
Connected apps also mean connected permissions. An agent inherits whatever access you grant it, so grant narrowly: a dedicated Drive folder rather than the whole account, a single Slack channel rather than the workspace. This is not paranoia — it is the same rule you would apply to a new hire on day one.
What to do this month
Do not migrate anything. Pick one recurring task you can describe in five sentences, run it as a scheduled agent for four weeks, and measure two numbers: hours you got back, and how many outputs you had to redo. If the redo rate is above a third, the brief is wrong, not the model.
The broader shift is already visible — agents are moving out of pilots and into production work across the market, and the competitive gap this year will sit between businesses that learned to delegate to them and businesses that kept typing prompts. ChatGPT Work does not close that gap for you. It just removes the last excuse for not starting.
