AI has moved from asking to doing — inventory the tasks before writing the policy
Data published yesterday by OpenAI shows, within its own individual-account dataset, a shift from answers toward producing work. The signal is useful, but it does not measure company approval, productivity, or risk: that requires an internal task inventory.
- Nature
- Adoption analysis with an operational task-inventory method
- Audience
- Leadership, operations, data, security, legal, and process owners
- Published and verified
- 7 August 2026
- Boundary
- OpenAI data is attributed to the vendor and not extrapolated to Portugal or all companies
In brief
OpenAI says that, in contexts classified as work, users in the analysed set are more than twice as likely to use ChatGPT to produce or complete something than outside work. The dataset covers individual Free, Go, Plus, and Pro plans, not a census of companies. Separately, Eurostat recorded AI use in 20.0% of EU enterprises with ten or more employees in 2025. The useful decision is not to extrapolate a vendor statistic: it is to identify where AI already touches data, decisions, and deliverables, and assign an owner to every flow.
- The 6 August study is a vendor measurement of individual accounts; it does not prove return, compliance, or authorised company adoption.
- Eurostat measures enterprises with ten or more employees and found 20.0% use in 2025, up from 13.5% in 2024.
- A tool list is not enough: the governable object is the task, the data used, the output produced, and the decision it may change.
- In seven days, a company can close a first register with an owner, human gate, evidence, autonomy limit, and stop condition.
The signal is real; its unit of analysis is still wrong for governing a company
OpenAI's 6 August publication says professional use observed in its data is more oriented toward doing — writing, coding, analysing, or creating — than use outside work. The same page limits the dataset to individual accounts generally managed by the person and describes automated classification. It does not identify which tasks an employer approved, whether an output was accepted, or whether quality improved.
Eurostat provides an independent, comparable measure: in 2025, 20.0% of EU enterprises with ten or more employees used AI technologies, 6.5 percentage points more than in 2024. Written-language analysis was the most common use, at 11.8%. The two sources measure different objects; combining them as if they covered the same population would create a false conclusion.
What each piece of evidence can support
- 2025European business adoption
Eurostat measures enterprises with ten or more employees and publishes the statistical methodology.
- Use inside ChatGPT
OpenAI publishes patterns from its individual-account base and states dataset boundaries.
- Internal decision
The company replaces general estimates with a register of the tasks it actually performs.
Register tasks, not only tool names
Asking whether a team uses AI produces incomplete answers. The same tool may summarise a public document without consequence or prepare a contractual communication containing personal data. Risk comes from the combination of purpose, data, ability to act, affected user, and reversibility.
For every task, retain an approved input-output example, data class, validating person, destination system, available telemetry, and the error that requires stopping. NIST AI RMF organises management around Govern, Map, Measure, and Manage; NIST AI 800-4 notes that production monitoring remains fragmented and that functionality, operations, impact, and human intervention are not the same metric.
Minimum register for an AI-assisted task
| Field | Verifiable question | Evidence | Blocker |
|---|---|---|---|
| Purpose | What concrete work is completed? | Input, output, and acceptance criterion | Generic goal such as “be more efficient” |
| Data | What information enters and may leave? | Class, origin, basis for use, and retention | Sensitive data without a rule |
| Authority | Does the output advise or execute? | Permissions, human gate, and limit | External action without an owner |
| Quality | Which error matters? | Sample, failure rate, and review | Only a positive impression exists |
| Operations | How is degradation detected? | Logs, alert, fallback, and stop | No observability or reversal |
Seven days to turn diffuse use into a governable map
On day one, ask each owner for five real tasks, not five tools. On days two and three, select recent examples and classify data, impact, and autonomy. On day four, remove duplicates and merge equivalent tasks. On day five, define gates and stop conditions. On day six, test a sample. On day seven, leadership accepts, conditions, or suspends each flow.
The result need not be a platform. A versioned evidence table is already better than a policy that bans or authorises “AI” in the abstract. Review the register whenever the model, integration, data source, affected user, or ability to perform actions changes.
- Do not collect real prompts containing personal data merely to prove adoption; use minimised, approved examples.
- Separate perceived productivity from measured quality.
- Prioritise tasks that publish, pay, delete, contract, or communicate externally.
- Attach a review date to every flow and retire those that lose their owner.
Sources consulted
Consulted on the stated update date.
Opens in a new window.- 01From asking to doing: How the world is putting ChatGPT to workOpenAI · AI-OPENAI
- 02OpenAI Signals — data and methodologyOpenAI · AI-OPENAI
- 0320% of EU enterprises use AI technologiesEurostat · EU-EUROSTAT
- 04The use of artificial intelligence technologies in the European Union — 2026 editionEurostat · EU-EUROSTAT
- 05AI Risk Management FrameworkNIST · GL-NISTAI
- 06Challenges to the Monitoring of Deployed AI SystemsNIST · GL-NISTAI
Participate
Share the analysis. Add context.
Distribute the article through your channel or leave an observation connected to the topic.
Comments
Loading the conversation.