Measure AI adoption by completed work, not usage count

AI adoption reports often show "usage went up," but that doesn't tell you if your team's work actually changed. Usage counts only show someone opened the tool. If a rep generated ten draft replies but a person threw all of them away and rewrote from scratch, that's not completed work.

OpenAI's July 2026 guide to managing AI investments makes a similar point: don't judge value by token price alone. Look at completed work, better decisions, and workflows that can scale. This post turns that idea into a first measurement a small team can run. It's not a promise of savings. It's a record you can actually check.

Separate usage from completed work

Usage tells you who opened the tool. Completion tells you whether a task that started with some input actually finished to a standard. Completion needs evidence. First, a human reviews the AI's output. Then it's sent to a customer, saved to an internal doc, or moved to the next step.

You don't need a company-wide metric in week one. One task that repeats a few times a week, with an owner who can check the result, is enough. Pick something you can describe start to finish. Examples: "sort 20 new inquiries into department-ready drafts" or "pull decisions and open questions from meeting notes into a table." Not sure which task to automate first? Spend time mapping which tasks repeat and where the exceptions happen.

Four things to log in your one-week tracker

When you add cost, skip a complex allocation formula. Model and tool cost, human review time, and rework count in the same row is enough. After a week, you can move past "did we use AI a lot?" Instead, you can ask: "which type passed with no review, and where did a human have to start over?"

A one-week log template you can use today

Log one line per task, and read it together for 15 minutes on Friday. Small numbers aren't failure. If you find a task with lots of exceptions, that's a reason to avoid a bigger automation, not a bad result.

Task name:
Starting condition and input:
Format of the AI's output:
Human reviewer:
Definition of done:
Reason for this fix or stop:
Review time (rough):
One thing to change next time:

"One thing to change next time" doesn't have to mean rewriting the prompt. Standardizing input file names, adding a "must-cite-source" column to the output, or flagging exceptions as "needs review" can matter more. Change one thing at a time, so you can actually read the difference it makes.

Read time savings alongside outcomes

Don't call it a saving just because AI produced a draft fast. Maybe a person spent a long time tracking down the source. Or the next team had to reverse a bad classification. Then the cost just moved downstream. On the other hand, if a person approves the draft quickly and the same format keeps working, that's a reason to consider a wider pilot.

Match your bar to your actual risk. For anything touching customer-facing text, money, contracts, or personal data, look at the review and approval history before the completion count. If you need help mapping your workflow and where responsibility sits, a consulting review can help lay out the current flow.

Three decisions to make on Friday

• Pick one task type that passed, and repeat it next week with the same bar.
• For task types with heavy edits, make one of input, expected output, or exceptions clearer.
• For hard-to-reverse actions, scale back from live execution to read, sort, or draft only.

A good AI metric isn't a number that makes tool activity look good. It's a record that helps your team make a better call about what to hand off, and how far. Show the start and finish of one task first, and your next investment will rest on evidence, not a guess.

Where to start

Choose one recurring task this week. Log the eight fields above as one line per task. Read the log together for 15 minutes on Friday.