Take a copy before you change anything

You ask the AI to tidy up duplicate customers. It suggests a filter. The filter reads sensibly, and it is only a small change anyway, so you say go. On Monday somebody opens a record and finds two different people inside it, with both sets of notes. The filter matched on name and postcode, not on email. There is no copy. The only way back is people re-typing notes from memory and their sent items, one record at a time, for a week.

Small changes with a wrong filter are how you learn the restore procedure under pressure.

A copy is not the backup somebody told you exists. Not last spring's export. Not an undo button you have never pressed. It is today's records, the exact ones this change will touch, with their IDs, saved where the change cannot reach them.

Step three of Give your AI the keys says to take a copy before every change, without judging whether the change deserves one. This step is how, for one change, in a system you run.

The rules

Copy the exact records the change will touch, found by the same filter. Not a similar filter, and not the AI's description of one. The same rule the change will use, run the same way. If the change also moves linked records, such as notes or attachments, copy those too. A full export of everything is a fine extra. It is not a replacement, because you cannot count it against the change.

No record ID, no copy. Names repeat. Emails are often the very thing you are matching on. The ID is what lets a row go back where it came from, so it goes in every row.

Tag the records as well as copying them. The copy sits outside the system. The tag sits inside it, on each record, naming the change and the month. If the copy goes missing, or somebody asks next spring which records the merge touched, one search answers it.

Three counts, one number. The filter count, the rows in the copy and the records carrying the tag. If any of the three differ, you have copied or tagged a different set from the one you are about to change. Stop and find out why.

Keep the copy where the change cannot reach it. A read-only folder, outside the system, somewhere the AI cannot write. Name the file for what it holds and the date. A copy that the next change can overwrite is not a copy for long.

Do it now

1. Write the change in one line. What it does, to which records, found how.
2. Ask the AI for the filter and the count, without changing anything. It needs read-only access for this, and step one of Give your AI the keys sets that up.

I am about to make a bulk change in <system>. Do not change anything.
The change: <your one line>.
1. Write the filter that finds every record this change will touch, in plain words and in the form <system> uses.
2. Run it read-only and give me the count.
3. List every other kind of record the change will move or edit, such as notes or attachments, with a count for each.
If you cannot do this without writing to the system, say so and stop.

3. Run the same filter yourself in your system's own search. Export the result with every column, and check the ID column is there before you save. If the search cannot express the filter, export every record with its ID and apply the filter in a spreadsheet instead.
4. Export the linked records the AI listed, for the same records, into files of their own.
5. Name each file for what it holds and the date, such as candidates_before_merge_2026-10-05. Save them in a folder the AI cannot write to, and make the files read-only.
6. Add one tag to every record the filter found. Name the change and the month in it.
7. Fill in this sheet and keep it with the copy.

Change:        <verb> <which records> in <which system>
Found by:      <the filter, in plain words>
Filter count:  <what the system's search shows>
Copy:          <file name>, <rows> rows, record ID in every row
Linked rows:   <file name and row count for each kind>
Tag:           <tag name>, on <number> records
Stored:        <folder>, read-only
Counts agree:  <yes or no>

A worked example

Say you run a twenty-person recruitment agency. Eleven years of candidates sit in one database. The same person often appears twice, because they came back through a different job board. Nadia, the operations manager, wants the duplicates merged, and the AI has offered to do it. Before it touches anything, she takes the copy.

Change:        Merge candidate records that share an email address
Found by:      email lowercased and trimmed, on more than one record, blanks left out
Filter count:  1,311 records in 604 groups, out of 18,412 candidates
Copy:          candidates_before_merge_2026-10-05.csv, 1,311 rows, ID in every row
Linked rows:   notes_before_merge_2026-10-05.csv, 4,870 rows
               applications_before_merge_2026-10-05.csv, 1,036 rows
Tag:           dedupe-oct, on 1,311 records
Stored:        the directors' shared drive, read-only, no AI access
Counts agree:  filter 1,311, copy 1,311, tag 1,311

Nothing is merged yet, and nothing is at risk.

The failure you will hit

The export without the IDs. You will export from the screen you always use. It gives you the columns on that screen: name, email, phone, consultant. It looks complete. On the day you need it, you will have a thousand names. You will have no way to say which record each came from. Names repeat, and the records have since been merged. Beat it by opening the file before you save it anywhere, finding the ID column and running the count.

Under that sits a quieter mistake. On Wednesday the AI suggests a better filter, and you accept it. Monday's copy now covers a different set of records. Any change to the filter means a new copy and a new count.

Not for you if

If you will not open the file and look for the ID column, stop here. You have a file, not a copy. And if your plan for a bad day is an undo button you have never pressed, carry on without this step and learn the restore procedure under pressure.

Step two is the fence: one sentence naming what this change touches, and everything else reported rather than fixed. The Vault is the paid library, and Data & Analytics is one of its categories.