How Power Users Actually Use AI (What Our Data Shows)

TL;DR: Here's how power users use AI, based on our own data on 2,170 customers (July 15, 2026). They don't just send more prompts. They build a system. The top 10% of active users ran 55.8% of all prompts, and our most engaged tier averaged 2.91 of seven features against 1.16 overall. 60.6% of that tier had connected AI through MCP, against 0.7% of the least engaged.

Most writing on how power users use AI is a list of habits somebody believes in. This one comes from a database. On July 15, 2026 we pulled everything our production database knew about our customers into a single table, one row per person: 2,170 rows, about 26,200 prompt enhancements, and a little under 200,000 MCP events. Then we asked a plain question. What do the people who get the most out of AI actually do that everyone else doesn't?

The answer surprised us, and it isn't the one we would have written in a brainstorm. Power users don't write longer prompts or know secret phrasing. They build a small amount of structure around their prompts and wire it into the place where they already work.

How concentrated is AI usage among our customers?

Very. A small group does most of the work. Of the 1,512 customers who had run at least one prompt, the heaviest 10% generated 55.8% of all prompts.

The full distribution makes the shape clearer:

The median active user had run 6 prompts. The mean was 17.3. When the average is nearly three times the median, a handful of people are pulling it up, and "the average user" describes almost nobody.

This matters if you are trying to learn from other people's AI habits. The loudest advice comes from the long tail of heavy users, but most people are the median user who tried something six times. The useful question isn't "how do I prompt like an expert?" It's "what did the heavy users set up that let them keep going after prompt number six?"

Who counts as a power user in our data?

We didn't define power users by volume alone, because volume rewards anyone running an automated script. The analysis scored every customer with a weighted formula that favours breadth first, then habit, then raw activity:

engagement_score = (distinct features used × 3)
                 + (active days × 2)
                 + (total events × 0.1)

Customers were then placed into five tiers, evaluated top-down:

For the rest of this post, "power users" means the 109 Champions unless we say otherwise. They're the group whose behaviour is most different from everyone else's, and the group the analysis profiled in depth.

What do power users do that other AI users don't?

They use more of the tool, and specifically the parts that save work for later. Champions averaged 2.91 of the seven features we track. The whole base averaged 1.16.

Here is where Champions and Dabblers (activated users who drifted away) differ most:

Look at the last row first. Nearly everyone in both groups used the core enhancer. That's the button that turns a rough prompt into a structured one. The button isn't what separates the groups. What separates them is everything around the button: somewhere to keep good prompts, somewhere to store background the AI needs every time, and a connection that brings all of it into the tool they're already typing in.

The Dabbler row tells the other half of the story. These people activated. They got a result. Then nothing made the second session easier than the first, so there wasn't much of a second session. (Why so many people stop at one feature is its own question, covered in why most people use one feature of seven.)

Why does MCP predict heavy AI use so strongly?

Because it moves the prompt tool into the AI client itself. In our data, it was the single strongest marker of a Champion: 60.6% of Champions had connected MCP, against 0.7% of Dabblers.

MCP, the Model Context Protocol, lets clients like Claude Desktop, Claude Code and Cursor call external tools directly (see tool use). Once it's connected, improving a prompt stops being a detour to another tab and becomes something the assistant can do inline.

The numbers around it are lopsided in a telling way. Only 104 customers used MCP at all, 4.8% of the base. Those 104 accounted for roughly 199,000 MCP events, which is extreme concentration. Claude Code alone produced about 177,000 of those events, and most of them are session handshakes rather than individual prompts, so treat the event count as a sign of an always-on connection, not of 177,000 separate requests.

A second number points the same way. Only 4.5% of all customers used both the enhancer and MCP. Among Champions, 57.8% did. Power users don't pick one surface. They use whichever one is closest to the work in front of them.

We want to be careful about the direction of cause. The data shows that MCP users and heavy users are largely the same people. It doesn't prove that connecting MCP turns anyone into a heavy user. Developers, who already live in Cursor and Claude Code, are both more likely to connect it and more likely to use AI all day. What we can say is that nearly everyone who connected it ended up using the product heavily.

If you work in one of those clients, our MCP setup guide covers the connection, and managing prompts through MCP in Cursor and Claude Desktop walks through the day-to-day workflow.

Do AI power users save and reuse their prompts?

Yes, and this is the most copyable habit in the data. 67.9% of Champions had saved prompts to a library, against 17.7% of Dabblers. 35.8% had stored reusable context, against 3.6%.

A saved prompt is a decision you only make once. A saved context (your role, your audience, your constraints, your house style) is background you only type once. Together they explain why a Champion's hundredth session takes less effort than their first, while a Dabbler's second session feels exactly like their first.

The surprising part is where power users came from. At signup we ask where people kept prompts before. Among Champions, the most common answer was "none", at 45.9%, with another 22% answering "other". Across all customers, 1,085 people, half the base, said they had no system at all.

So the typical power user in our data wasn't a Notion organiser who moved their collection over. They had nothing, found a place to put things, and started putting things there. You don't need an existing system to become a heavy AI user. You need a first place to save the second prompt.

If you're starting from nothing, a simple context block is the highest-leverage thing to save. This is the kind of reusable context Champions stored, with the specifics swapped for placeholders:

ROLE: You are helping [your role] at [company or project].
AUDIENCE: [who reads the output, and what they already know]
GOAL: [what a good result lets them do]
VOICE: [3 adjectives], no hype, short sentences.
ALWAYS: state assumptions; flag anything you are unsure of.
NEVER: invent statistics, quotes, or sources.
FORMAT: [default structure, length limit, headings or not]

Paste it once, save it, and attach it every time instead of retyping it. For teams, the same idea scales into shared variables; reusable prompt variables for dev teams shows how, and building a brand-voice context covers the marketing version.

Who are the AI power users in our customer base?

Mostly founders and builders. Among the 109 Champions, founders were the largest group at 33.0%, followed by content creators at 13.8%, developers at 11.0%, marketers at 8.3% and freelancers at 7.3%.

Group size and conversion rate tell different stories, though. Across all customers, the share of each job title that reached Champion looked like this (job titles are self-reported at signup):

Developers and product managers reached the top tier at the highest rates, which fits the MCP finding: they're the people already working inside AI-enabled editors. The product manager group is small, 44 people, so read that figure as directional.

One more pattern stood out. Customers who left their job title blank, or typed something that didn't fit a category, did worst on every measure: activation as low as 47.5% and 30-day retention of 26.2% for the custom-title group. We don't read that as the title causing anything. It's more likely that people who can say what they do can also say what they want from AI, and that clarity carries into how they use it.

Two facts about how Champions arrived are worth knowing too. 90.8% bought through our AppSumo lifetime deal, and 66.1% came to us directly rather than through an ad or search. Heavy users mostly arrived already knowing what they wanted.

Which AI tools do power users bring their prompts to?

ChatGPT first, by a wide margin, with Gemini second. The per-platform counts from our browser extension in the same analysis ranked ChatGPT at 1,910, Gemini at 1,175, Claude at 563 and Perplexity at 265, with a long tail below that.

Two things follow. First, even customers who pay for a prompt tool mostly use it with the assistant they already had, so good prompting habits have to travel across tools rather than being tuned for one. Second, the MCP traffic sits almost entirely on the Claude side (Claude Code dominated it), which means the most engaged workflow in our data isn't a chat window at all. It's an agent or editor calling tools while the person works.

Different models respond differently to the same instruction, which is why heavy users keep reference material close. Our negative prompt support matrix is one example of the kind of lookup table that saves a power user from relearning the same lesson on each model.

How can you copy the AI power user workflow?

Copy the structure, not the volume. Our data suggests four steps, in this order:

1. Save the second prompt. The first time you run something is an experiment. The second time is a pattern. Save it with a name you'll search for later.
2. Write your context down once. Use the block above. Attach it instead of retyping your background into every chat.
3. Keep a short library, not a big one. Champions saved prompts they reused, not every prompt they ever wrote. Favourite the handful you run weekly.
4. Bring the tools to where you work. If you live in Claude Desktop, Claude Code or Cursor, connect your prompt tools there. If you live in ChatGPT or Gemini, use a sidebar so you never switch tabs.

For a worked example of those four steps applied to one real job, an email workflow that actually saves time shows a single repeated task turned into a saved, reusable system. Founders can go further with the founder's prompt system, and developers can treat prompts like code with prompt versioning.

A quick self-check, if you want to know where you stand today:

1. How many prompts have I run more than twice this month?
2. How many of those are saved somewhere I can find in 10 seconds?
3. Do I retype my role, audience or constraints in most chats?
4. Can I improve a prompt without leaving the tool I'm writing in?

Mostly "few / none / yes / no"  -> you are where our Dabblers were.
Mostly "many / most / no / yes" -> you are working like our Champions.

What can't this data tell you about AI power users?

Quite a lot, and it's worth being plain about it.

It describes one product's customers. People who pay for a prompt tool already care about prompting more than the average AI user does. Adoption figures measure whether someone used a feature at least once, not how deeply. Most customers signed up in May or June 2026, so absolute retention figures are inflated by recency; overall 30-day retention was 42.6%, but the analysis itself warns that relative comparisons are the durable signal. Job titles are self-reported, and several groups are small. MCP event totals are dominated by connection handshakes. And the snapshot is from July 15, 2026; we haven't re-run it since our database moved.

What it does show, clearly, is a direction. The people who got the most from AI in our data weren't better at writing a single prompt. They had a place to keep prompts, a place to keep context, and a connection that put both inside the tool they already used. That's a system anyone can build in an afternoon.