In a freemium product there is no clock. The limit is the only thing that ever asks anyone to pay, which makes it a product decision disguised as a pricing detail.
A free trial has a forcing function built in: the clock runs out whether or not the user is ready. Freemium removes the clock. Nothing will ever interrupt a happy free user unless you design the interruption. So the cap has to do three jobs at once, and they pull against each other:
That third sentence is the whole strategy. Get it right and you are not selling features at the moment of conversion, you are asking someone to give up a routine, which almost nobody does.
Before you pick a number, pick what you are counting. A good unit meets four conditions:
| Condition | Why |
|---|---|
| It tracks value received | Otherwise the user is billed for something they do not experience as benefit, which reads as arbitrary |
| It tracks your cost | Otherwise your most-loved free users are your biggest losses, and generosity becomes financially unsurvivable |
| The user can feel it coming | A limit that arrives without warning feels like a trap; one you can see approaching feels like outgrowing |
| It grows with habit, not luck | If usage spikes for reasons outside the user's control, the cap punishes circumstance rather than rewarding adoption |
Words dictated. Meetings recorded. Minutes transcribed. Images generated. Each is a defensible unit for a different product. "Clients" or "properties" usually is not, because it punishes the user for having a good year, which is the opposite of what you want from someone whose success you want to be part of.
| Shape | Converts when | Fails when |
|---|---|---|
| Usage cap metered units per period | The user outgrows the allowance | The allowance throttles the habit before it forms |
| Feature cap functionality withheld | The user hits a wall they need past | The free tier feels crippled and the magic never lands |
| Seat cap number of people | The team adopts around the first user | Your buyer works alone, and no amount of success adds a seat |
Everyone repeats that Wispr Flow gives you 2,000 free words a week. The actual documented structure is more interesting, and it is the design worth stealing.2
| Soft cap | Hard cap | |
|---|---|---|
| Desktop free tier | 2,000 words per week | 5,000 words per week |
| iOS free tier | 1,000 words per week | 1,500 words per week |
Two lessons sit in that little table.
Two thresholds, not one. The soft cap creates the pressure; the hard cap is where it actually stops. In between, the user is over the line and still working - reminded, but not cut off. That gap is what stops a bad week from breaking the habit you spent two months building. A single hard wall converts some users and quietly destroys others.
The caps differ by surface. Mobile gets half the desktop allowance. Whatever the reason - cost, behaviour, strategic emphasis - the point is that the cap was tuned per context rather than set once as a policy. Caps are instruments, not principles.
No founder interview explaining how these numbers were chosen could be found. Assume, as with everyone else, that they were arrived at by iteration.
The document under review states a good design principle: the cap sits on capture, never on memory. Free users can record less, but the product never forgets what it already learned. The reasoning is sound - a product that forgets to punish you cannot be trusted with the thing that makes it valuable.
Now look at the opposite bet. Granola's free tier does the reverse: capture stays open, and it is the history that gets limited.3 That is not a mistake, it is a different theory: make the accumulated memory the thing you pay for, because it is the thing that grows in value.
Both are legitimate. What is not legitimate is choosing without noticing there was a choice. State it as a bet:
Published benchmarks: for self-serve freemium, 3 to 5 percent free-to-paid is good and 6 to 8 percent is great; Kyle Poyar's more recent framing raises "great" to 8-12 percent and puts AI-native products at 15-20 percent.4 Treat the exact figures as soft and the ordering as reliable.
Now do the arithmetic that strategy documents skip. Suppose the target is 1,000 paying users at 49 EUR per month, the historical Leedflow entry tier shown in the supplied screenshot. That is about 588,000 EUR a year. Treat 49 EUR as a modelling assumption, not a current competitor fact: Leedflow's current price card and FAQ say entry pricing starts at 89 EUR excluding tax.5
| Conversion rate | Free signups required | As a share of an estimated 100,000 French real-estate professionals |
|---|---|---|
| 3 percent (good) | 33,300 | a third of everyone |
| 5 percent | 20,000 | one in five of everyone |
| 15 percent (AI-native claim) | 6,700 | one in fifteen |
The 100,000 professionals figure comes from an industry blog and has no official source behind it. It is used here to show the shape of the problem, not to be right. See the fact sheet.
The lesson is not that the plan is impossible. It is that the conversion rate is not a detail - moving from 3 percent to 15 percent is the difference between needing a third of the country and needing a niche. Which means the cap design, which sets that rate, deserves more of the team's attention than the choice of which product to build. Most documents get this backwards.
Read Wispr Flow's own plans page - it takes two minutes and it is the primary evidence for the soft-and-hard-cap pattern - then Poyar on free-to-paid conversion for the benchmarks.
Want to design a specific cap for the assistant, with the arithmetic done properly against real inference costs? That is a good next session. Ask your teaching agent.