The Newsroom

Vol. III · No. 1 · September edition

Execution

OpenAI's August research shows agents moving into legal, sales and other business functions. The next design problem is responsibility, not access.

Every automated step meets a named point of review, refusal or accountable decision. Common Intelligence

OpenAI describes enterprise AI moving from assistance to execution. Agents are no longer limited to answering questions; they can use tools, create files and complete longer pieces of work for review.

The use is spreading quickly outside engineering. OpenAI reports strong growth in enterprise Codex activity across legal, sales, recruiting and marketing, where decisions depend on policy, context and relationships as much as on speed.

The boundary becomes more important

That does not remove the human role. It makes the boundary more important. Someone must decide what context the system receives, which actions require approval, which evidence is sufficient and when uncertainty should become a refusal.

Those choices should be made with the people who already carry the responsibility. They know the difficult cases, understand the consequences of a false positive and can tell the difference between a shortcut and a broken control.

Make the handoff visible

A good deployment gives the system repetitive work it can perform consistently and gives people better information at the point of judgment. It also makes the handoff visible enough that a reviewer can understand what happened and challenge it.

People plus AI is not a slogan for preserving the old process. It is a practical design rule: use machines for scale and repetition, keep accountable judgment with named people, and build the connection between them so neither side is guessing.

Sources

Talk it through

Leo Largillet

Leo LargilletChief executive

Available this week

Bring a process your team knows should work better. In forty-five minutes, we will listen to how it really runs, identify where AI could help, and be clear about what we would leave with people.

Frequently asked

  • Our work starts from the people already responsible for the operation. The aim is to remove repetitive search, comparison and drafting, while making their judgment easier to apply and keeping accountable decisions with named human owners.

  • They help map the real process, choose the difficult cases, test early versions and define where review or refusal is required. An internal group learns to operate and improve the system before handover.

  • One senior sponsor, access to the people and systems involved, and agreement on the business measure before building begins. Leadership also protects time for operators to participate, because their knowledge is part of the system.

  • It should be frequent, expensive, evidence-heavy and close to a measure the company already trusts. The two-week diagnostic compares candidate workflows and tests the strongest one on real company data.

  • That is why it is tested on real exceptions before production. Corrections become part of the evaluation set, and the system is designed to ask for review when evidence is weak rather than appearing certain.

  • We build in the client environment wherever possible, use the minimum permissions each action needs, and make activity auditable. Data boundaries, retention, human review and prohibited actions are agreed before a production release.

  • The engagement is designed to reduce that dependency. The client receives the source, prompts, evaluation cases, runbooks and controls, and proves the handover by running the next cycle with its own team.

  • We establish the current cost and operating baseline together, then review the same measure after each release. If the workflow does not produce a credible financial or operating improvement, we do not expand it.