There is a considerable difference between asking an AI assistant to draft a message and asking one to finish a piece of work. The first returns material for a person to handle. The second may gather information, move through software and complete steps with human supervision. That practical distinction sits behind a new OpenAI analysis of how European organisations are using its tools. The findings suggest that AI agents for business are moving out of the demonstration phase in at least part of the market, including smaller firms. The research does not establish that every shop, consultancy or manufacturer has embraced autonomous software. OpenAI's EU analysis, published on 14 October 2026, examines activity among its own customers.
For a smaller business, the attraction is easy to see. Owners frequently have processes that require repeated switching between email, spreadsheets, websites and business applications. An assistant that can follow a clear instruction through those stages might remove friction without requiring another employee to learn a new system. The difficult question is whether the same assistant can maintain quality, permissions and accountability once its work reaches customers or changes records.
AI agents for business are not just longer chatbot conversations
An agent is designed to carry out a task with steps, rather than merely produce a response. The distinction is not perfectly sharp: a conversation can contain tools, and an agent may require several questions from its user. What matters operationally is who does the next action. A conventional chatbot might explain how to organise invoices; a capable, properly authorised agent might classify a batch of documents, identify exceptions and prepare a report for approval.
That final approval is crucial. Software can carry out repetitive checks, but it should not silently grant itself authority over payments, personal records or contracts. A useful agent needs clearly defined boundaries. Those might include access only to selected folders, restrictions on sending messages, an audit record of changes and a rule to stop if the evidence conflicts. These are practical design requirements, not claims that every commercially available agent already meets them.
OpenAI reports that weekly active users of Codex and ChatGPT Work among EU small and medium businesses using its products rose more than eightfold since May 2026. The company’s definition in this analysis covers businesses with no more than 200 associated users. Such growth can occur from a relatively small starting point, and OpenAI has not provided in this public account the full population of European businesses needed to generalise the result to the entire economy.
The report also says its sample of EU enterprise customers generated more than half of their output tokens through these agentic tools. Tokens are units of model-generated or processed text, not completed jobs, employee hours saved or company profit. A complicated automated task may use many more tokens than a brief question. Readers should not interpret the percentage as proof that half of all work in European businesses is now performed by AI.
The same tool can be useful in very different firms
A bookkeeper might use an agent to assemble preliminary reporting materials while retaining final responsibility for reconciliation and the meaning of each figure. A marketing agency might use it to investigate changes across its website and create a proposed repair plan. A retailer might use it to compare suppliers and prepare an order for review. These are illustrative workflows, not independently verified outcomes attributed to OpenAI’s dataset.
The common feature is that much of the work takes place between applications. Businesses already understand the goal, but reaching it requires small, tedious actions that interrupt more valuable decisions. If software can perform those actions accurately, the productivity gain can be greater than simply receiving a better paragraph from a chatbot. If it gets the permissions or facts wrong, however, the cost of fixing the mess may exceed any apparent saving.
This is where earlier automation experiments in small businesses become relevant. The practical challenge has rarely been finding a task that sounds suitable for AI. It has been making sure the task can be completed reliably, that people trust the results and that responsibility remains clear when exceptions arise. Agents make the opportunity larger, but they also increase the importance of those basic controls.
European adoption is uneven, even among the customers studied
The OpenAI report shows that intensity varies by country. In its enterprise sample, the proportion of output tokens associated with agentic products ranged from 22 per cent in Slovenia to 85 per cent in Lithuania. Those are differences within a vendor’s customer sample, not a reliable ranking of each country’s entire technological sophistication. Customer composition, company size and the tasks represented could all affect the result.
The firm also divides customers into an especially intensive group and a more typical group. It reports that high-intensity European organisations generated substantially more output tokens per active user, while high-intensity users elsewhere in the world still exceeded their European peers. Those comparisons provide a view of changing usage patterns. They do not independently measure whether the businesses involved serve clients better, grow faster or make more profit.
That measurement distinction deserves attention. A company with heavy AI use might have found an effective process. Equally, it might be experimenting, correcting large amounts of output or simply processing tasks with more textual material. Adoption is an input. Outcomes such as fewer mistakes, shorter turnaround times and happier customers need to be examined separately.
Skills and process matter more than the demonstration
The introduction of an agent often reveals problems with a process that employees had quietly worked around. Files are named inconsistently, different departments interpret a status differently, and instructions are not written down because experienced colleagues know what they mean. An agent can expose that ambiguity by making the wrong assumption repeatedly. Before automation is worthwhile, someone may need to simplify the workflow and decide what counts as success.
Access rights are another constraint. An AI tool that can read a document may not need permission to edit it. An agent preparing a supplier comparison does not necessarily need the ability to buy the product. An assistant collecting overdue invoices should not be able to change a customer’s bank details. Properly separating read access, proposed changes and execution protects the business and makes results easier to check.
The analysis also belongs in a wider account of changing jobs. An earlier OpenAI report about work crossing occupational boundaries suggested that AI lets people undertake tasks outside their traditional specialisms. That work provides context, but is not proof that every person can suddenly perform another profession safely. The most promising uses expand someone’s ability to prepare work while leaving specialist judgements and legally significant approvals with qualified people.
LiveAIWire has also covered workers' experience of growing AI demands and the possibility of automation fatigue when promised benefits fail to appear. Those concerns make a useful counterweight to vendor adoption figures. People need to understand not only how to initiate tasks, but how to check work, pause processes and report a mistake without being penalised for it.
A practical test for an agentic business workflow
A small firm considering agents can start by identifying one repeatable process with clear inputs and a reversible output. It can compare the agent’s draft with the result that an experienced worker would produce, record the types of corrections required and decide which steps remain human. That gives a business evidence from its own operations rather than from a generic product demonstration.
The next question is what happens when information is missing. A dependable agent should stop or ask for assistance when it cannot safely complete the task. It should not fill gaps with confident inventions. Monitoring should focus on the quality of finished work and the cost of supervision, not simply the number of times an employee opened the AI application.
The new European data suggests that more businesses are exploring the transition from answers to actions. What it cannot tell each owner is whether a particular process will be worth automating. That judgement belongs in the detail: the permissions granted, the errors avoided, the time actually recovered and the human decisions the software leaves intact.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.