What office workers actually finish when they learn AI tools: data from 1,150 lesson attempts (Oct-Nov 2026)

What this report measures

Stepthrough is a free site where office workers learn AI tools by doing: each lesson is a practice copy of ChatGPT, n8n, Claude, Make or another tool, and the learner clicks and types every step while the lesson checks their work. That makes finishing measurable: a lesson is finished only when the last step has been done, not when a video has played to the end.

This report covers every interactive lesson attempt on learn.ynteractive.com from October 1 to November 3, 2026: 12,116 visitors (distinct browsers), of whom 619 opened at least one lesson, making 1,150 lesson attempts across 153 lessons. 202 people finished at least one. Test accounts, QA browsers and known crawlers are removed first (see Methodology); nothing here identifies a person. For scale, 6,280 visitors (52%) read one of the site's 1,676 text guides in the same period: roughly ten readers for every person who opened a lesson.

How many people finish an AI lesson?

Three in ten attempts open a lesson and never do the first step (345 of 1,150): people who looked at the intro screen and left, or who were reading rather than practicing. Among the 805 attempts that did step one, 68% finished. The median time from opening a lesson to its last step was 3.3 minutes (507 timed finishes), pulled down by the five-step orientation lesson; task lessons typically run 5 to 15 minutes.

Where do people quit?

If an attempt is going to stop, it stops early. 64 of the 259 unfinished attempts (25%) did exactly one step, and 54% stopped at step 3 or earlier. After the 30% mark the curve flattens: only 27% got past halfway, and 7% stopped in the final tenth. For anyone designing training, the first three steps carry most of the risk; a long typed answer or an unfamiliar screen belongs later, not first.

Which tools do people finish?

People who open a ChatGPT lesson almost always do the first step (85%), and three quarters of those finish. People who open an n8n lesson do the first step only 62% of the time. Part of that is length (n8n lessons here run 25 to 93 steps, ChatGPT lessons 14 to 26), part is that n8n is an unfamiliar canvas while ChatGPT looks like a chat box everyone has seen. Once started, n8n attempts still finish 61% of the time. The outlier is the automation basics concept lesson: 77 opens but 19 first steps, which is what happens when a page ranks for a definition and most arrivals wanted the definition.

Does lesson length matter?

Yes, but less than you might guess. Between the 11-20 step band and the 51+ band the finish rate per open falls from 48% to 35%, about 13 points, and most of that is the start rate (69% against 60%): a long lesson scares some people off before step one. Among people who did start, a 51+ step lesson still finished 58% of the time. A lesson a learner can see progress through is not a lost cause; one whose first screen looks like an hour of work is a harder sell.

Does "Do it for me" help or hurt?

Every Stepthrough step has a "Do it for me" button that performs the step and moves on. The worry is that it turns practice into a slideshow. The data says the people who use it are the people who finish: 134 of 346 (39%) pressed it at least once, and their attempts finished 77% of the time against 56% for everyone else. This is a correlation: someone who presses the button has hit a step they could not do and kept going, while "never pressed it" includes everyone who left at step two. Use is uneven, too; a handful of people fast-forwarded whole lessons (one 53-step lesson logged 167 presses from two people). Still, an escape hatch on every step does not appear to lower completion, and it keeps stuck learners in the lesson.

When do office workers learn?

Weekday mornings. 74% of attempts happened Monday to Friday (Monday and Tuesday alone: 36%), and weekday attempts finished 50% of the time against 39% on weekends. 37% of attempts began between 6 am and noon Eastern and finished 51% of the time, the best of any band; evening attempts (6 pm to midnight) were the fewest (15%). The 7-8 am shoulder and 10 am peak read like "before the day starts" and "after the first meeting". The site does not record time zones, so visitors outside the US East are shifted on this chart, and some overnight opens are daytime visitors elsewhere.

Is ChatGPT a learning referrer?

This is the finding that surprised us. Browsers that arrived from a ChatGPT answer at least once (a chatgpt.com referrer or ChatGPT's utm_source=chatgpt.com tag) were 8% of visitors, 48% of the people who opened a lesson and 66% of the people who finished one. Their attempts finished at the highest rate of any source (72% of started attempts, against 68% for direct and other traffic). The contrast with social is stark: Facebook, Instagram and Threads sent 1,031 visitors (9%), of whom 72 opened a lesson and 3 finished one. Search engines sent 259 visitors and 4 finishers.

Two cautions. ChatGPT referrals grew fast inside the period, from 15 visitors in the week of October 5 to 521 in the week of October 26, as Stepthrough published more pages that ChatGPT cites; this is one site's experience, not a market-share figure. And a browser counts for a source if any of its visits came from it, so a learner who found the site through ChatGPT and came back directly is counted as ChatGPT. Even so, the pattern is consistent: someone who asks an assistant how to build a custom GPT and clicks the lesson it suggests arrives ready to do the work.

Courses, single lessons and coming back

32% of attempts happened inside a course (a sequence of lessons with a certificate) and the rest on a single lesson page. The finish rate per open was almost identical: 46% inside courses, 48% for single lessons. Packaging lessons as a course did not, in this period, change whether any one lesson got finished.

Coming back is where the difference shows. Of 619 people who opened a lesson, 31% opened a second and 15% a third. Of the 202 who finished one, 72% opened another, 52% finished a second, 27% finished three or more and 12% finished five or more. Finishing the first lesson is the moment that predicts a learner, which is one more reason to keep step one easy.

What we take from it

• The first three steps decide most outcomes. A lesson's opening step should be short and obviously doable; long typed inputs and unfamiliar screens belong later.
• Length is a start-rate problem more than a finish-rate problem. A long workflow is better shown as a sequence of 15-25 step modules in a course than as one screen that looks like an hour of work.
• An escape hatch on every step does not appear to lower completion, so "Do it for me" stays.
• Text guides are read ten times more than lessons are practiced, and AI assistants are now citing them. Every guide should point to the matching free lesson, because the people who arrive that way finish.

Methodology

• Source. Stepthrough's own first-party event log for learn.ynteractive.com: lesson_open when a lesson is opened, step each time a step is completed, lesson_complete when the last step is done. Events carry an anonymous per-browser id, lesson id, step number, page path, referrer host and any utm_source. No names, emails or IP addresses are used; most learners never sign in.
• Period. October 1 to November 3, 2026 (UTC). Hours and weekdays are in US Eastern time.
• Units. A visitor is a distinct browser. An attempt is one visitor on one lesson, however many sessions it took. Started means at least one step event; finished means a completion event. Lesson length is the lesson's current step count.
• Filtering. Test accounts, browsers flagged as internal QA, the admin dashboard, automated browsers and user agents matching common crawler and link-preview patterns are never recorded or are removed before counting. Browsers that could not store an id (325 events) stay in totals but cannot form attempts.
• Sources. ChatGPT: any event with a chatgpt.com referrer or utm_source=chatgpt.com. Social: Facebook, Instagram or Threads referrer. Search: Google, Bing, DuckDuckGo or Yahoo. Everything else, including email and direct visits, is "direct, other".
• Not collected. Device type, country and time zone, so this report says nothing about mobile versus desktop or where learners are.
• Small samples. Rates for tools other than ChatGPT and n8n rest on 15 to 30 attempts and are directional. Every chart carries its sample size and a "Show the numbers" table.