AI can make a busy week feel lighter while the team's output stays exactly where it was. People report saved time, send more prompts, share impressive examples. None of that tells you whether the workflow improved.
A small marketing team can get the answer without a data analyst. Choose one repeated workflow, record how it performs before AI enters it, then compare the same work over 90 days. Keep the measurement narrow enough that people will actually maintain it.
One baseline, reviews at 30, 60 and 90 days, three measures. It will not give you laboratory proof. It will give you enough to decide whether to continue, change the setup, or stop spending time on it.
Start with one clean baseline
Pick a specific unit of work. "Create the weekly email" can be measured. "Use AI for marketing" mixes too many jobs, people and standards. Write down where the work starts, what a finished item looks like, who approves it, and what does not count.
Record two normal weeks before you change anything. Use completed work, not numbers people recall later. Hold the brief, the approval standard and the counting rules steady. If a campaign peak or a public holiday makes those weeks odd, note it and extend the baseline.
OpenAI's AI Workflow Test Brief recommends a named group, fixed dates, observable success signals and a recorded quality or correction rate. That is enough structure for a small internal test. You do not need a reporting system before the first measurement.
Use the 30/60/90 reviews for different decisions
At day 30, check the test is being run the way you designed it. Look for missing records, changed briefs, work that quietly moved outside the chosen workflow. Read the rejected drafts. Early good news is provisional, because the sample is small and people are still learning the new process.
At day 60, compare the full period with the baseline, item by item as well as in total. One unusually large project can drag an average around. Check whether faster drafts are creating extra revision work later, and whether the same quality standard is still being applied.
At day 90, make an operating decision: continue, revise for another fixed test period, or stop. OpenAI's 90-Day AI Transformation Starter Kit uses an evidence gate and ends with a decision to advance, reshape, defer or stop. It also warns teams off presenting estimated time saved, or first use, as proven impact.
The dates are management checkpoints, not scientific thresholds. Day 30 does not prove adoption. Day 60 does not guarantee a stable effect. They exist to catch bad measurement early and to stop a weak pilot drifting past a decision date.
Track only these three measures
- Median time to first usable draft. Start the clock when the agreed brief is ready. Stop when a reviewer receives a draft that can be assessed properly. Use the median, because one delayed item can pull a small average far from typical work.
- Accepted pieces shipped per week. Count only work that passes the existing approval standard and reaches its intended channel. Draft volume is easy to inflate and it creates review work, so it stays out of this count.
- Average revision rounds per piece. Count each return from reviewer to creator as one round. This catches the common failure where the first draft arrives sooner but needs more correction before it can ship.
Keep the definitions beside the tracker. "Usable draft" and "revision round" have to mean the same thing in the baseline and at day 90. If someone changes a definition halfway, mark the date and stop combining the two periods.
Give the next decision an owner. In a small team, a pilot can run on for months simply because nobody has the authority to end it. The owner also notes any process change that makes the current period hard to compare with the baseline.
Do not call a winner on tiny samples
Four pieces before AI and five after it can swing hard because one brief happened to be easy. A single viral post can lift weekly output while saying nothing about the production process. Early movement is a reason to look closer, not to announce a result.
A June 2026 Microsoft Research study compared sustained users with similar coworkers who adopted later. That comparison is more credible than counting activity after a rollout. Your team probably cannot copy the design, but it can stop treating every before-and-after change as proof.
Our view: a modest, repeatable gain at day 90 beats a spectacular week at day 30. Inspect the underlying pieces. Note the staff changes, campaign peaks and approval delays that moved the numbers. If removing one item flips the direction, collect more work before you decide.
Choose the workflow this week and start the two-week baseline before anyone changes the process. Put day 30, 60 and 90 in the calendar now, with one named owner for the final call.