Every CFO in the world is now asking the same question, and it's the hardest question in enterprise AI: What are we actually getting for this?
The default answers are terrible. Seats sold. Tokens consumed. Credits burned. These numbers describe cost, not value.
Here's the reframe that matters: AI adoption is a language problem before it's a technology problem. The words you use to describe AI work decide how it gets funded, staffed, and scaled.
OpenAI's new piece, How to connect AI usage to business value (September 16, 2026), is quietly a vocabulary lesson wearing an admin-console update: usage, task insights, engineering outcomes, five questions, and a reproducible ROI — a five-step discipline for turning tokens into trust.
[1]Step 1 — See where attention goes. The Usage view maps active users, credits, and tokens across ChatGPT Work and Codex. It shows where adoption grows and where spend concentrates.
Step 2 — See what work AI actually does. The Insights task classifier groups messages into use cases and tasks: account research and planning, feature development, campaign planning. For the first time, an AI platform gives you the nouns of AI work, not just the numbers.
Step 3 — See what AI produced. The Outcomes view tracks Codex's share of merged commits and lines of code, compared against review time, defects, and rework, to answer whether AI is making teams ship better.
Admins measure the mechanism; business owners measure the outcome. Neither can do it alone — that's the premise of the five questions.
[1]要点
- What would you like to improve? Pick an outcome that matters to the team.
- How does the process look today? Establish a baseline: frequency, duration, what good looks like.
- What changes with AI? Compare over a fixed period, including review and correction time.
- What does this make possible? Does saved time flow to customers and growth?
- Is the benefit worth the investment? Compare gains against AI, setup, training, and support costs.
Do the math, then attack the assumptions. The article's worked example: sellers each prepare two account briefs a week, four hours each; with AI, one hour — three hours back per brief.
- 20 sellers × 2 briefs × 3 hours × 46 weeks = 5,520 hours a year
- 50% redeployed into productive work, fully loaded at $75/hour = $207,000 of annual capacity value
- First-year costs (AI, setup, training, support): $60,000
- Result: 245% illustrative ROI (all figures hypothetical in the original)
The number that deserves skepticism isn't the ROI — it's the 50%. Will the saved time actually flow back into customer conversations? The proof is accumulating: 1Password estimates 553% ROI and $0.8M in annual engineering capacity value from Codex; ATV Big Air Tour cut listing reviews from eight hours a week to one; Playco reported 50% fewer manual fixes.
The playbook: stop buying AI and start reading it. Pick one common task that supports a business priority, review it with the business owner, agree on a baseline and the outcome to measure, and set a date to review progress. Small scope, named outcome, repeated often.
The companies that win the AI era won't be the ones spending the most — they'll be the ones who can answer five questions faster than anyone else.
[1]