OpenAI's story about Oracle suggests a first step: time one slow task, map its data, and check new answers against the old method.
OpenAI describes Oracle's own staff using two OpenAI tools in a customer story reported on October 8, 2026. ChatGPT Work is OpenAI's AI assistant for business staff. Codex is an AI tool that writes computer code, including database queries. Oracle reports 130,000 active users of ChatGPT. It also reports more than 95,000 active users of Codex.
Oracle's recruiting team built a tool with ChatGPT Work. The tool takes a job description, researches similar roles and compares pay. Oracle says this research used to take two to four days. Oracle's head of recruiting says recruiters now prepare for conversations with hiring managers in about 15 to 20 minutes.
A second Oracle team wrote a data map: a list of the company's records, how they connect, and the rules that apply to them. With the data map, Codex turns a business question in everyday words into a database query. A business user describes the result they want and gets back an analysis, a report or an application. A vice president of that team says responsible system design, architecture and security remain essential.
Oracle's engineers who keep live systems running also use Codex on incidents, meaning faults in those systems. Codex gathers the details and finds the right written procedure. Oracle says a typical simple incident once took an hour and now takes minutes. The times and user counts in the story are Oracle's own, published by OpenAI, which sells both tools.
Few firms using AI have changed how they handle their data. A Census Bureau working paper from April 2026 used a survey of more than 117,000 US firms, run November 2025 to January 2026. Among firms using AI, 64 percent reported making no changes to their business to use AI in the past six months. About 7 to 8 percent had changed how they manage and store data in the past six months. The Census Bureau has not reviewed this paper the way it reviews its own publications.
Executives also report bigger gains from AI than their own figures show. A working paper from the National Bureau of Economic Research, dated March 2026, surveyed corporate executives, most of them chief financial officers. Its main survey drew 603 responses from November 11 to December 16, 2025. The paper has not been peer reviewed.
On average, the executives said AI raised output per worker by 1.8 percent in 2025. Their own revenue and headcount estimates implied a smaller gain, 0.6 percent. The authors say the gap likely reflects a delay before the gains show up in sales.
Until a gain shows up in sales, it is only a claim. So a company should time one slow task the same way before and after adding an AI tool. The saving should then be counted in a cost line that finance already reports. Someone must also own the AI tool and its cost: the manager whose team does the task, paying from that team's budget.
What to do first:
- Chief operating officer: by October 30, pick one task that takes staff a day or more, and time it on the next ten cases.
- Chief information officer: by November 13, write the data map for that task, covering its records, how they connect and the rules that apply.
- Manager whose team does the task: from December 1, run the old method beside the new AI tool on the same cases, comparing answers.
- Chief financial officer: from the December month-end close, report the hours on that task next to the cost line it should lower.
The Oracle story includes one check of this kind. The vice president of the data map team says one user got an answer from the new data map tool almost at once. Its numbers matched the old manual method exactly.
