Insights
Deleting a donor is not one action
Every system holding a copy of that person defines the word differently, and nobody coordinated. The disagreement between those definitions is where the money goes.
The cleanup got approved. Someone is about to sit down and start removing people, and the question they will ask first is which system to start in.
It’s the right question, and nothing in the stack answers it. The order of operations is not a decision any of these systems surfaces, none of them knows what the others are about to be told, and nothing anywhere signals that the order matters. That is where the cost comes from.
We get paid to untangle this, which is our interest in writing about it.
Nobody agreed what the word means
Whatever your database calls them, the record on your screen is one copy. There are others. The email tool has its own, with its own identifier and its own subscribe state. The payment side has one, holding the card and the schedule. The events tool, the petition or advocacy system, the peer-to-peer pages, the spreadsheet someone keeps: each learned about this person once and has kept its own version since.
Here is the part that costs money. Every one of those systems was built by a different company, and each of those companies decided on its own what the word “delete” should mean inside their product. None of them coordinated. There’s no standard, no shared definition, and no reason there would be.
So “delete this donor” is not one instruction. It’s one instruction per system, each with a different outcome, and the differences between those outcomes are the whole problem. Deleting is not the risk. The disagreement between the definitions is the risk, and it exists in every stack, including yours.
The questions that separate a cleanup from an incident
The definitions differ along the same axes wherever you look, because a record carries the same kinds of thing everywhere: money, consent, history, its relations to other records, and copies of itself elsewhere. Five questions come out of that, and they are the five that decide most cleanups rather than the only axes there are. Put them to each system you run, and expect the answers not to match.
| Ask each system | Why the answers differ | What it costs if you assume |
|---|---|---|
| Does deleting stop any money? | Systems that hold a card or a schedule may treat removing the person as cancelling the arrangement | Recurring gifts end silently, with no failed payment and nothing in a churn report |
| Does the record of their consent survive it? | Opt-out and unsubscribe state is kept on the record, so removing the record can remove the evidence that they asked not to be contacted | A later import mails a person who unsubscribed, and the first to notice is them |
| Can it be undone, and by whom? | Some keep a recovery window, some are immediate and final, and some can only be reversed by the person themselves | A donor removed in error becomes unreachable, permanently |
| What travels with the record? | Related records may be deleted alongside the parent, orphaned, or block the deletion entirely | History disappears, or a bulk job half completes and leaves no list of what got through |
| Does the deletion propagate, or get undone? | Integrations carry some operations and not others, and a record missing on one side is often treated as one to recreate | The record returns without its history, and now there are two of them |
The answers will not match across your systems. That mismatch is the thing to plan around.
You won’t find these answers in one place. They live in each vendor’s own documentation, and the wording is usually precise once you go looking, which is the good news: it’s checkable in an afternoon rather than a project.
What the spread actually looks like
Four products, taken one at a time. This is not a stack and not a recommendation: four unrelated data points, chosen only because each publishes its behaviour plainly. Yours will be different products with different answers, which is the point.
Money can stop silently. Stripe’s developer documentation says deleting a customer “permanently deletes a customer. It cannot be undone. Also immediately cancels any active subscriptions on the customer.” A monthly gift is an active subscription. It stops that moment, with no failed payment, no lapse notice and nothing in a churn report, because from the system’s point of view nothing failed. Stripe also keeps the history deliberately, so the reporting still looks right while the income has gone.
Consent can be deleted along with the person. HubSpot documents two kinds of contact deletion, which answer two of the five questions differently from each other. After the restorable kind, “if a contact is re-created or re-imported using the same email address, bounces, spam reports, subscriptions, unsubscribes, and quarantines will be restored.” After the permanent kind, “subscription preferences, including unsubscribes, aren’t restored.” One product, one person, two buttons: the first keeps the record that they asked not to be emailed, the second does not. Delete the second way, let that address arrive again in a later import, and nothing left in the system knows to stop the next appeal.
Some doors only open outward. Mailchimp distinguishes archiving from deleting, and deleting is one-way: “You can’t add a deleted contact back to your audience. The only way to add that contact back is if they rejoin through the hosted Mailchimp form.” Not from a backup, not from your database. If someone’s removed in error, the organization has no route back to them that doesn’t run through that person deciding to sign up again.
Related records can travel with the parent. Salesforce’s developer documentation: “If you delete a parent object, you delete its children automatically, as long as each child object can be deleted.” Which of your records are children rather than neighbours is a configuration question specific to your setup, and it’s worth answering before a bulk job rather than after, because that is the difference between removing a person and removing their history.
Swap all four for whatever you actually run and the answers change completely. What doesn’t change is that they differ, that the differences are documented, and that nobody has ever written them down in one place for your organization.
The half-deleted state is worse than either end
Do it in one system and not the others, and your systems now disagree about who exists.
The person is gone from the database and still on the mailing list, so they keep receiving appeals from an organization that believes it no longer knows them. Or they are gone from the mailing list and still live on the payment side, giving every month to an organization with no record of them, until the card expires and there’s nobody to contact about it.
Then the syncs run. A person present in one system and missing from another is exactly the condition an integration exists to resolve, and plenty of them resolve it by writing the record back. It returns without whatever used to hang off it, and now you have two.
None of this was decided. It happens because the instruction was treated as one action, and the systems were never in the same conversation.
“We have done this before and nothing broke”
Worth taking seriously rather than waving away.
Two reasons it holds. Deleting a record that lives in one system is a different operation from deleting one that lives in five, and only the second kind has anything to disagree about, but both are called the same thing and neither announces which one you are doing. And the failures here are quiet by construction: a cancelled recurring gift doesn’t raise an error, a resurrected record looks like a new donor, and an unreachable person doesn’t complain. Nothing broke, and nothing would have told you if it had.
The test is not whether past cleanups caused visible trouble. It’s whether anyone checked afterwards, in the other systems, and could have found it.
What it is worth doing first
Before the sequence, do the arithmetic that decides how carefully to handle this.
Take the number of people on the delete list who have an active recurring gift, and multiply by their annual value. A list with 30 people giving $25 a month carries $9,000 a year. That’s the amount at risk from one afternoon’s work, and it’s the figure that decides whether this gets an hour of care or a week of it. A recurring gift can also read as inactivity in a database that counts one-off gifts, which is how a live donor ends up on a delete list to begin with.
- 01List every system that holds a copy of a donor, and which way each sync runs. Not which systems you own. Which ones hold people, and whether the arrow points in, out or both. If nobody can answer this, stop, because nothing after it is safe.
- 02Answer the five questions above for each one, in writing. It is an afternoon in vendor documentation, and it is the only version of this that exists for your organization.
- 03Take anyone with a live recurring gift off the list entirely, until someone decides deliberately to end that gift. This one step protects the money.
- 04Prefer whatever your systems call archiving, suppressing or marking inactive. These keep the record and its history while taking the person out of view, and they can be undone. That is the difference from the door that only opens outward.
- 05Do one batch of 50 and stop. Wait a full sync cycle, then look at every system again and check they still agree. That is the test, and it costs a day.
- 06Write down what was decided and who signed it, so the next person inherits a rule rather than this question.
That’s an afternoon, and it’s the difference between a cleanup and an incident you find in October.
Where we fit
Working out what one decision does across systems that were never designed together is most of what we do.
Most of the list above you can run without us, and you should. What we would actually give you is the version of this article written for your systems specifically, with the five answers filled in and the order set, which is a conversation rather than a project. It’s worth having when nobody can answer the first question.