Ask an AI assistant about your competitors' prices and you get a helpful answer to copy-paste somewhere. Ask this month's crop of AI agents and you get back a finished spreadsheet, formatted, sourced and ready to send. That difference — an answer versus a deliverable — is what two launches in the same July week were really about, and it changes what a one-person business can sensibly hand over to software.
One week, two launches, one direction
On 9 July, OpenAI opened its GPT-5.6 model family to the public — three tiers named Sol, Terra and Luna — alongside a new agent called ChatGPT Work. The pitch, in OpenAI's own framing, is finished work, not conversations: the agent pulls context from your connected apps and files, breaks a goal into steps, keeps at it, and hands back completed reports, spreadsheets, presentations and even small web apps. The separate Codex app was folded into one unified ChatGPT desktop application in the process.
Two days earlier, on 7 July, Anthropic pushed its own agent product, Claude Cowork, out of the desktop corner it launched in: Max subscribers can now use it on mobile and on the web, with wider rollout to follow. The design assumption is telling — you start a task at your desk, glance at progress from your phone, and collect the finished output later. Neither company is describing a chat window anymore. Both are describing a colleague you brief and come back to.
The numbers say this isn't about coding
Alongside the Cowork expansion, Anthropic published usage data covering 1.2 million anonymised sessions from more than 600,000 organisations, sampled between 11 and 31 May. The largest category, at 33.4 per cent, was business process and operations work — assembling reports, reconciling spreadsheets, the administrative connective tissue of running a company. Content creation and copywriting came second at 16.4 per cent. Software development, the use case these tools grew up on, was third at just 8.7 per cent, with DevOps at 7 per cent, research at 6.4 per cent and data analysis at 5.8 per cent behind it.
VentureBeat, reporting on the data, highlighted Anthropic's own description of the dominant use cases: the work around the work — tasks that exist in every organisation but appear in nobody's job description. For a European solopreneur, that is exactly the pile that eats evenings: the quarterly overview, the price-list update, the client report that is 80 per cent copy-paste from three systems.
What a five-person business can realistically delegate
The honest test for delegating to an agent is the same as for a junior hire: is the task well-defined, is the raw material accessible, and can you check the result faster than you could produce it?
Recurring reports. Monthly revenue summaries, project status updates, a weekly overview of open quotes. The inputs live in systems the agent can read; the output format repeats. This is the 33.4 per cent category for a reason.
Spreadsheet cleanup and reconciliation. Matching invoices against bank lines, merging a messy export with your product list, hunting duplicates. Tedious for you, mechanical for an agent — and easy to spot-check.
First drafts with a real brief. A proposal built from your past three proposals, a product description in your established tone. The agent drafts from your material; you edit rather than start from blank.
Research briefs. A structured comparison of suppliers, regulation summaries, a market overview with sources named — delivered as a document, not a chat scroll.
What still doesn't delegate well: judgement calls, anything involving a promise to a customer, and one-off tasks so unusual that briefing the agent takes longer than doing the work.
You still sign the work
An agent that hands back a polished file creates a new risk precisely because the file looks finished. A wrong number in a confident spreadsheet is more dangerous than a wrong number in a chat reply you were going to rework anyway. The rule of thumb: the agent produces, you approve. Spot-check totals against the source system, verify any figure that will reach a customer or a tax office, and keep the human signature on anything contractual. Notably, the augusto.digital review of these launches makes the same point from the enterprise side: what limits agent adoption is not model capability but whether processes, permissions and data are clean enough to act on safely.
Permissions and GDPR: the boring part that decides everything
Finished work requires access — that is the trade. ChatGPT Work is built around connected apps; Cowork reads the files and folders you give it. Before connecting your mailbox, drive or CRM, three checks matter for a European business. First, least privilege: connect the folder the task needs, not the whole archive, and prefer read-only access where writing isn't required. Second, GDPR basics: customer data flowing to an AI provider makes that provider a processor, so you need a data-processing agreement in place — both OpenAI and Anthropic offer them on business tiers — and a line in your records of processing. Third, revocability: know where the connected-app list lives and check it quarterly, because permissions granted for one task have a way of outliving it.
When a chatbot is still all you need
If your AI use is occasional questions, brainstorming and the odd rewritten email, an agent subscription buys you little — the chat tier already covers it. The agent shift pays off when a task recurs, has structured inputs, and produces an artefact someone is waiting for. It is also fine to wait a quarter: both products launched this month, and early adopters will map the sharp edges for free. The direction, though, is set — the industry's own usage data shows the centre of gravity moving from answering questions to producing work, and prices for that work are falling.
The practical question for a small business is no longer whether AI can help, but which of your recurring tasks are shaped like something an agent can finish. That inventory — what recurs, where the inputs live, what checking looks like — is precisely what Cresly's AI Readiness Scan walks through with you, so that when you do hand work over, you know exactly what you delegated and what you still sign.