Questions became analysis and action

OpenAI launched the Data agent in ChatGPT Work on 10 September. It was designed to connect with company data, investigate changes and build interactive dashboards in conversation. Users could refine analysis without writing queries. The launch described capability rather than demonstrated outcomes in Norwegian organisations.

For a project team, shared analysis could surface variance and risk faster. Source-system permissions still applied, and recipients needed the data, filters and time behind the answer. A dashboard became decision evidence only when definitions were approved by the data owner.

Scaling needed more than a successful demo

Gartner reported on 1 September that only 22 per cent of surveyed organisations had scaled AI across several business units. Sample and method limited generalisation, but local benefit did not automatically become a common way of working. Shared data, cost controls, skills and decision rights all needed resolution.

On 9 September, Gartner described four shifts in future work involving closer cooperation between people and AI. In project work, an agent could investigate and propose while a named person approved priority and resources. Automated actions required limits and a route back.

An analysis card connected the answer to the outcome

The team could record question, data sources, filter, agent version, dashboard link, decision, approver and actual result. Monthly checks of corrected answers, stale data, cost and reversed decisions showed whether the workflow scaled. Sharing then became traceable work rather than a loose report link.

Controls in practice

OpenAI described data-platform connections and access controls as part of the Data agent. An organisation still needed to test whether row and object permissions survived questions, dashboards and shared results. A screenshot or export could lose dynamic controls and therefore required separate classification and recipient limits.

Scaling across business units also required a common language for measures. Sales, service and finance could use the same term with different calculations. Before comparing units, data owners needed to reconcile definition, currency, period and missing values. Disagreement should appear as a data-quality task rather than be smoothed over in a summary.

Sources

OpenAI: Now everyone can put data to work, 10 September 2026.

Gartner: Only 22% have scaled AI across business units, 1 September 2026.

Gartner: Four shifts shaping the future of work, 9 September 2026.

Microsoft Dynamics 365 Blog: One always-on AI at Work roadmap, 25 August 2026.

For discussion

Which data-driven decision can be shared, checked and acted on without obscuring ownership?