How to tell whether your AI rollout actually saved time

Three months after rolling out AI tools, an owner asks the team whether it was worth it. Everyone says yes. The owner asks how much time it saved. Someone says "loads." Another says "maybe an hour a day." Nobody has a number, because nobody wrote one down before the rollout, and nobody is writing one down now.

That leaves the renewal decision to mood. If the team likes the tools, they stay. If the subscription line catches the accountant's eye in a bad month, they go. Neither outcome tells you anything.

You can fix this in two weeks with a sheet of paper.

The four columns

Choose three tasks the team now does with AI help. Pick tasks that repeat at least weekly and that end in something concrete: a quote, a client update, a reconciliation, a job report. For each instance over the next two weeks, the person doing it fills in four numbers, in minutes.

Making it is the time from starting to having a first draft. This is the column AI shrinks, and the one everyone remembers.

Checking and fixing it is reading the draft, correcting figures, rewriting the paragraphs that sound wrong. People undercount this one because it feels like part of the job rather than a cost of the tool.

Explaining context to the tool is the time spent telling the AI things it should already know. Who the client is. Which price list applies. That the client prefers formal address. That the cheap option was discontinued. Count every minute of retyping background.

Redone after it went out is any time spent later because the output was wrong: the corrected invoice, the apology email, the second site visit. Usually zero. When it is not zero, it is often large.

If you can, run the same tasks without AI for a few instances too. If the team refuses, a rough estimate of the old time from the person who did the task before is better than nothing.

Reading the sheet

Add the four columns and compare with the old time. Three patterns tend to show up.

The first is a genuine saving: making time falls a lot, checking time is modest, the other two columns stay near zero. Keep the tool and expand it carefully.

The second is a wash: making time falls, but checking and explaining eat most of the gain. That does not mean the tool is useless. It means the setup around it is costing you.

The third is a loss, with the redo column carrying it. One wrong quote that had to be withdrawn can cancel a month of fast drafting. That calls for an approval gate before anything customer-facing leaves, not more speed.

The column a company memory shrinks

The explaining column is the one most firms can cut without touching the model. Everything in it is a fact or a rule that lives in someone's head and gets retyped into a chat window every day. Write those down once, in a shared memory the AI reads before it starts, and that column falls for everyone at the same time.

We do this ourselves, one correction at a time. On Friday a correction about punctuation in outgoing messages went into our company memory as a written rule with its reason, so no session has to be told it again.

The memory will not shrink the checking column much. Someone still has to read the draft. And it will not touch the redo column on its own. That depends on whether a person actually approves the output before it goes out, and on whether the facts in the memory are current. A stale memory moves time from the explaining column straight into the redo column.

One more thing to watch for. If the explaining column is high for one person and low for another doing the same task, the second person has usually built a private prompt or a private document that the first does not have. That private document is company memory sitting in the wrong place. Ask for it and share it.

Start Monday

Print the table, or make it a shared sheet. Pick the three tasks today. Ask the people doing them to fill it in for two weeks, and promise them it is about the tools, not about them. On the last Friday, add it up together.

Book 20 minutes at pgco.world/book.