FounderWiseDecisions, not feeds
← All articles FounderWise · Long-form

One customer uses it and you cannot remove it

The feature costs you a day a month and shows up in four percent of accounts. Removing it is judged as a loss, it is holding up workflows you cannot see, and putting it back later is not an option.

04 Nov 2026 12 min read By Joshua Pi’Rwot
Share X LinkedIn

The export scheduler is used by four percent of accounts, costs roughly a day of engineering a month, and blocks a change you want to make. The obvious answer is to remove it.

The four percent is not the number to look at. Whatever those accounts built on top of it is invisible to you, and it is the thing you are actually removing.

Why these three models

The decision is whether to deprecate a feature with low usage and a real maintenance cost. The features that fire are a removal that will be evaluated against having it rather than against a clean product, a set of user-built dependencies nobody designed or documented, and a change that is much harder to undo than to make.

Three lenses. Reference dependence produces an equilibrium answer about how the removal lands. Self-organised use produces a complex answer about what is actually attached to it. Options produce a random answer about the reversibility. The second is the one that produces the surprises, and the only way to see it is to look before you announce anything.

1. Removal is a loss, and simplification is not a defence

You are evaluating the feature against an ideal product: fewer surfaces, less code, a cleaner roadmap. Your customers will evaluate the change against what they have today, and against that reference the change is a removal.1

That asymmetry is why the announcement always lands worse than the internal discussion predicted. Internally the feature is a cost. Externally it is a thing that exists and works, and taking it away is a loss that weighs more than a comparable improvement elsewhere would gain.

Two practical consequences follow and both are about framing rather than about substance.

The first is that a removal paired with a replacement lands very differently from a removal alone, even when the replacement is worse in some respects. The reference shifts from what is being taken to what is arriving. A migration path, however crude, is worth more than the engineering it costs.

The replacement carries its own failure mode. Announcing a path is making a commitment with a date attached, and if the replacement slips you own two problems instead of one: the removal, and a broken promise about what was going to stand in its place. Offer only a path you can ship before the removal date, never one you intend to build in parallel with it. A crude path that exists beats a good one that is planned. If you have neither, move the date rather than announce a path you hope to finish.

The second is that the reason you give changes the reference. Because almost nobody used it invites the affected customer to notice that they are the almost nobody, which is a statement about their importance rather than about the feature. Because it blocked a change that a majority of you asked for names a trade, and a trade can be accepted.

Do not tell a customer their workflow was not popular enough to keep.

2. Find out what was built on top of it

The second lens is the one that turns a small decision into an incident, and it operates entirely out of your view.

Users solve their own problems with local rules, and what they build on top of a feature is not designed by anyone and not visible in your analytics. Somebody’s month-end close depends on that export landing in a particular folder at a particular time. Somebody’s client report is assembled from its output. A downstream script parses its file format. None of that shows up as usage; it shows up as a low percentage and a quiet account.2

Which means the usage number measures the wrong thing. Four percent of accounts touch the feature and the question is what stops for them when it goes, and whether what stops is a convenience or a monthly obligation to somebody else.

The investigation is cheap and almost never done. Take the accounts that use it, look at what proportion of their overall activity runs through it, whether the usage is on a schedule, and whether they are among your larger customers. A feature used lightly by many is a different object from one used constantly by four accounts, and the two produce the same headline percentage.

A worked instance of what that looks like. A distributor on your product exports a stock movement file on the last working day of every month, opens it in a spreadsheet, reconciles it against their mobile money statement, and sends the result to a supplier who releases the next consignment on the strength of it. In your analytics that account touched the feature twelve times in a year, which reads as trivial. In their operation it sits between them and their inventory. So the thing to look for is regularity rather than frequency. Usage clustered on the same days of every month is somebody’s process. Usage scattered at random is somebody’s curiosity.

If you cannot get three conversations, the logs carry most of it. Plot each account’s use against the day of the month and the day of the week. Check whether the output leaves your system through a download, a webhook, an email or an integration token, because output that leaves has a recipient you have never met. Check whether any account calls it with an API key rather than through the interface, which nearly always means a script, and a script means the file format itself is a promise you made without noticing. That is an hour of queries, and it narrows the conversations you still need to one or two.

Then ask three of them what happens if it stops. Not whether they would mind, which produces politeness, but what specifically breaks and who else notices. That conversation regularly surfaces a dependency the customer themselves had stopped thinking about, and it is the only way to find one.

There is a second reading of the usage data that is worth taking. Low adoption sometimes means the feature is unnecessary and sometimes means it was never discoverable, and those are opposite findings that produce the same number. If the four percent who found it use it constantly, you may be looking at an undermarketed capability rather than an unwanted one, and removing it is the more expensive of the two mistakes because it is the one you cannot reverse.

The occasional finding is that the feature is load-bearing for your largest account through a chain nobody documented, and that discovery in advance costs one afternoon while the same discovery afterwards costs the account.

3. It is much harder to put back

The third lens is about the asymmetry between the two directions, which is what should set the sequence.

Removing a feature is a week. Restoring it after the code has been deleted, the tests removed, the data migrated and the team moved on is not a week, and in practice it does not happen: the company lives with the consequence instead. So this is a decision that is cheap in one direction and expensive in the other, which is exactly the condition under which preserving flexibility is worth paying for.3

The instruments are ordinary and mostly free. Hide it before you remove it, so that new accounts cannot adopt it while existing use continues, which stops the problem growing while you learn. Announce the date long before the change, because the cost of the surprise is larger than the cost of the wait. Keep the code behind a flag for one release rather than deleting it, so that a discovered dependency can be handled without an emergency.

The reason to act on the reversible half immediately rather than after the debate is that the structure keeps working while you deliberate. A feature that remains available keeps accumulating attachments, and it does so regardless of anyone’s intention to remove it later.4 Every week of discussion adds to the thing you are discussing removing.

And distinguish two things that get conflated in these discussions. Stopping new adoption is nearly free and reversible. Deleting the capability is expensive and permanent. Doing the first immediately and the second slowly gets you most of the benefit at a fraction of the risk, and the maintenance saving you were chasing mostly arrives with the first step anyway, because the cost was in supporting new questions rather than in the code sitting there.

Budget for the second-order cost of the hidden state, because it is not free. You are now supporting two populations, one that can see the feature and one that cannot, and every support reply, every documentation page and every demo has to know which of the two it is addressing. That is tolerable for a quarter and corrosive over a year. Put an end date on the hidden state at the moment you create it, even if all that date does is force a review, because the failure mode of a reversible step is that it becomes permanent through neglect and you finish holding every maintenance cost you set out to remove.

The same discipline applies to the flag. One release means one release, with the removal named in a ticket and dated. A flag kept indefinitely means maintaining the old path and the new one at once, which costs more than either decision would have.

What the three say together

  • Look at usage intensity per account and whether it runs on a schedule, not at the percentage of accounts.
  • Ask three users what specifically breaks and who else notices. Not whether they would mind.
  • Hide it for new accounts immediately. That is reversible and it stops the problem growing.
  • Announce the date early, offer a path however crude, and give a reason that names a trade rather than their unimportance.

Where they disagree

The dependency argument and the option argument point to different speeds.

Investigating dependencies says go slowly: the thing you need to know is invisible in your data, it takes conversations to surface, and the discovery you miss is the one that costs an account. Every week of investigation is cheap against that.

The option argument says the cost of holding is real and compounding. Every month the feature stays, more accounts adopt it, more workflows attach to it, and the removal you are contemplating gets strictly harder. Investigation has a shelf life and the situation degrades while you conduct it.

The reconciliation is the split between the two actions, which is why it is the recommendation rather than a middle path. Hide it now, which is reversible, costs nothing and immediately stops the option argument’s clock. Then investigate at whatever pace the dependencies require, because the situation is no longer worsening. Teams routinely do the opposite, deliberating for two quarters while adoption continues, and then removing under pressure without ever having done the investigation.

What none of them contain

None of the three prices what the removal signals about your product’s direction. Customers read deprecations as statements about which use cases you are committed to, and a removal in an area someone depends on makes them wonder what else is on the list. That effect reaches accounts that never used the feature.

None of them handles the feature that is cheap to keep and expensive to think about. Some maintenance costs are not engineering hours but the attention tax of a surface that has to be considered in every design discussion, and that cost is real, hard to measure and frequently the actual reason for the proposal.

None of them handles the feature that is quietly load-bearing for your own team. Support scripts, internal reporting and onboarding demos accumulate around a capability just as customer workflows do, and the internal dependencies are usually discovered a week after removal rather than before it. The three-conversation exercise works internally too and takes twenty minutes.

And one property the ensemble will not produce: the customers who use it are not a random sample of your base. They are the ones who explored your product furthest, which correlates with engagement and often with tenure. Losing them is a different loss from losing four percent of accounts, and the analytics report both the same way.

The one action that survives the ignorance: today, hide the feature from new accounts. That is reversible, it costs almost nothing, and it stops the dependency base growing while you find out what is actually attached to it.

Who has to move

The founder or whoever owns the roadmap, and the sequencing matters more than the decision. The instinct is to debate removal for a quarter and then remove it quickly once the argument is won, which is exactly backwards: the reversible part is being delayed and the irreversible part is being rushed. The cheapest first test is three customer conversations asking what breaks, which takes an afternoon and, in the case worth catching, tells you that a low-usage feature is holding up your largest account’s monthly reporting.

Sources and notes

  1. Sanjit S. Dhami, The Foundations of Behavioral Economic Analysis, Oxford University Press. Reference dependence, under which outcomes are evaluated as movements from a comparison point rather than in absolute terms, and the asymmetry between losses and gains relative to that point, are developed in the part on behavioural decision theory. Used in section 1 for the direction only. No loss-aversion coefficient is quoted, because the commonly cited figure is contested and varies with domain.
  2. Stefan Thurner, Rudolf Hanel and Peter Klimek, Introduction to the Theory of Complex Systems, Oxford University Press. Agent-based treatments, in which individuals following local rules produce structure that no participant designed, are developed in the chapters on complex adaptive systems. Used in section 2 for the dependencies users build on top of a feature, which are not designed, not documented and not visible in usage data. Used qualitatively; no simulation is claimed.
  3. Sanjit S. Dhami, The Foundations of Behavioral Economic Analysis, Oxford University Press, and the real-options literature it references. The value of preserving flexibility rises with the irreversibility of the commitment, so where one direction is cheap and the other is effectively permanent the sequence should be chosen to keep the expensive direction open. Used in section 3. No option value is computed.
  4. Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green Publishing. The observation that a system’s structure keeps producing its behaviour after the intentions behind it are forgotten supports the point in section 3 that a feature accumulates attachments for as long as it remains available, so stopping new adoption is a structural change while deliberation is not.

A note on a number this article does not give. There is no usage threshold below which a feature should be removed. The percentage of accounts is the wrong measure entirely, and the right one, what breaks and for whom, is not a number you can benchmark. Get it from three conversations.

Joshua Agonya Pi’Rwot, Founder.

Lock in your calls.

You’ve marked 0 of 5. Now choose how often you want the signals.

Step 1 · Pick your cadence

The DispatchWeekly · your Monday 5 callsFreealways

Step 2 · Where to send it

Personalize your BriefThe Brief

Tune every edition to the markets and industries you actually act on.

🔒 Unlock personalization — The Brief, $19.99/mo →
Free Dispatch forever · upgrade anytime · we never share your details.
Know where you stand?
The Dispatch tells you what changed. Knowing what to do about it is a different question, and it is the one FounderWise answers. Start with the free Traction Audit: 12 questions, about 3 minutes, scored out of 100.
Find out where you stand →

For teams, syndicates & programs

Recommended
Team
$15/seat · mo
Daily Brief for the whole team (min 3 seats).
  • Everyone on the same signal
  • Admin + shared watch-list
  • One invoice · ~25% off solo
Get Team →
Cohort License
$1,500/cohort
One cohort of up to 30 founders, for accelerators, funds & programs.
  • MVP Studio and GTM Studio for every founder
  • Your logo, your cohort
  • A funder report on what each founder did
Talk to us →
Pass the Dispatch on
Know a founder making these calls blind? Send them this week’s five — free, every Monday.

Decisions, not feeds. · Curated by Joshua Pi’Rwot · FounderWise · Free Audit · Store · parent of Business Growth Accelerator

Call committed. We’ll hold you to it.
Know where you stand →