AI and a person edit the same record. Whose save just won?
Test concurrent human and AI edits, stale writes, dependent fields and downstream effects before trusting the final record.
Article language: English

In this article
At 10:00 an assistant reads an order: collection is at 18:00. At 10:02 the account manager changes the time to 20:00 after speaking with the customer. At 10:03 the assistant saves a prepared summary and, in the process, restores 18:00. Both parties did their work. The customer is left with two possible arrival times, one of which is no longer valid.
This is a fictional situation at Aster, a decoration studio. The AI was meant to complete the collection instructions; a person was agreeing the conditions. The problem arose between reading and writing: the tool treated an earlier snapshot as if it were still current. A similar collision could concern an address, a quantity, an approved design version or a decision to pause an order. A well-formed sentence does not protect the other data.
Before the trial, establish when a person's change takes effect. Typing a new time into an unsaved form is not the same as a confirmed update to the record. Keep the response to the save or another form of evidence provided by the system. Otherwise the team may call it a conflict when the person was only preparing a change and the two tools never actually worked from different saved versions.
When accepting such a solution, check what happens to another person's change when AI finishes work it started earlier. It is not enough to look at a screen showing two names. You need a reproducible sequence: a known starting state, the person's save, the AI's later attempt to save and a final read. Carry out every trial on authorised test material, away from an active customer order.
First write down what the assistant is allowed to do. At Aster it prepares collection instructions from current arrangements. It does not set the time itself or approve a change of address. If an integration sends the entire record back even though it only needs to change the instructions, the operation is broader than the task. Those extra fields are where the pre-conversation state may return.
Prepare a copy of the record with a clear identifier and recognisable values. Use 18:00, the location "studio" and the note "collected by the event coordinator". A test change should be visible without interpreting writing style. A typo in a long comment is a weak first case: the team may miss its loss even while looking at the result.
Ask the provider to pause the flow after it reads its input and before it saves. During that pause, the account manager sets the time to 20:00. Then let the assistant finish. If the tool cannot be paused that way in the test environment, agree another controlled method for forcing the order of events. Pressing buttons simultaneously can be dramatic, but does not prove that the gap you care about was actually tested.
The expected result does not have to be an automatic merge. A solution may reject the stale save, preserve the proposal and send the conflict to a person. It may instead reread the record and prepare new instructions under agreed rules. It should not silently restore 18:00 or announce success if the instructions remain out of date.
Rereading does not close the window between reading and saving
A provider may suggest checking the record again immediately before saving. That is useful, but a moment still exists between the check and the write. A person could save a change in that moment. Move their action into that gap during acceptance testing. You will test the write condition, not merely how often the data refreshes.
Microsoft Graph's documentation for updating a list item describes the If-Match header. If the supplied eTag does not match the item's current tag, the request receives a 412 response and the item is not updated. This applies to the documented interface when that condition is used; it is not an automatic property of every AI integration.
Here, an eTag is a version marker for a resource, not a judgment of content quality. It checks a technical condition: does the version used for the work still apply when the save is made? It does not decide whether two changes can sensibly be combined. A separate process rule must decide that. At Aster, a new collection time affects the instructions even if the assistant saves them in a completely different field.
Assess both the data sent and its dependencies. A narrow update to one field reduces the chance of overwriting others, but it can still preserve an incorrect description. "Collect after work ends at 18:00" contradicts the time 20:00 even if the time field itself survived. Acceptance does not end with finding the right number in the right column.
Also ask what the integration does after rejecting an operation. Fetching a new version marker and resending old content does not resolve a conflict of meaning. In our case it might simply make it easier to save instructions for a time that no longer applies. Check whether the new data is actually assessed before a retry, rather than merely replacing the technical condition. Keep the inputs to both attempts so they can be compared.
For comparison, the Confluence Cloud REST API v2 description of updating a page provides for reconciling submitted content with a draft. The documentation warns that, when the versions have diverged significantly, the submitted content may entirely override the earlier draft. The word "merge" in a tool description is therefore not proof that every other person's change survives.

These examples show why a report must name the specific product, operation and configuration. Do not transfer the behaviour of one interface to a company CRM merely because both tools have a change history. Ask for evidence from the flow being tested. With AI batch jobs (in Polish), the time between retrieving a record and receiving a result can be particularly easy to overlook.
Three collisions that call for different responses
The first trial involves independent changes. A person corrects an internal organisational label while AI tidies packing instructions. Establish why you consider them independent and which part of each version must survive. If the tool rejects every change made by someone else to the whole record, it may require a retry here too. That is an operating cost worth knowing instead of promising seamless collaboration.
In the second trial both sides change the same information. The account manager enters 20:00 while the assistant proposes 18:30 based on old correspondence. The order in which the writes arrive does not by itself establish which time applies. A rule may give priority to the account manager's confirmed arrangement. In that case the AI should preserve its proposal as rejected or needing review, rather than overwrite the field on another retry.
The third trial is less obvious: the fields differ, but their meanings are connected. A person changes the collection location to the warehouse while AI adds directions for entering the studio. Automatically combining the fields creates an internally contradictory record. Write that dependency into the acceptance condition and check the whole result from the perspective of the person who must carry out the task. A mechanical absence of field conflicts does not mean the arrangements are consistent.
For each trial, keep the base version, the person's change, the AI proposal and the state after the operation. Include the write result and whether someone had to resolve a dispute. Using AI to compare document versions (in Polish) may help prepare a comparison, but the final check should include the actual values, not only a summary of differences generated by the same tool.
Check the effect outside the record too. If the process of sending a message or preparing a document began before the conflict was detected, a correct final record does not erase the earlier mistake. At Aster the team must establish whether the customer received instructions for 18:00. The report should identify when downstream actions begin and whether they use an approved state or only a proposal.
Cancellation of the entire order while AI is working is a separate case. Completing the description should not restore the order to active fulfilment. The rule here concerns an allowed state transition, not only protection of one field. A similar mistake can come from a delayed automation event (in Polish). In both cases a newly received result may describe the past.
Add a change and reversal to the harder cases. The account manager moves collection to 20:00 and then, for a new reason, sets it back to 18:00. The final time matching the AI's input does not by itself show the whole history. Check whether the collection conditions changed along the way, for example the responsible person or the permitted entrance. Comparing one visible field may miss a new decision even though the numbers look familiar.
A rejected save has a next step
A conflict should not disappear into a generic "try again" message. The operator needs to know what was preserved, what was not saved and where their proposal can be found. Designing a concurrent-edit warning (in Polish) helps separate those details in the interface. An audit also checks whether the visible message matches the actual state of the data.
Test a rejected save with long instructions that would take work to prepare again. Can the person view them and compare them with the current record? Can they safely return to resolving the conflict later? A message is not enough if the proposal then disappears and the employee must recreate it from memory. Protecting another person's change does not require wasting all the parallel work.
If the solution regenerates the instructions, inspect the new input. It should contain the current time and location, not merely the earlier text with a request to revise it. A change to a confirmed arrangement must also be reflected in approval. Rules for human approval of AI actions (in Polish) tie consent to a particular version; an old approval does not cover any future revision.
Also test what happens if the account manager changes the record again while the first conflict is being resolved. The system cannot assume that every subsequent save is unconditional after one warning. Show both conflicts and the final decision in the report. When changes keep coming, establish a route to a designated person instead of leaving the automation in an endless series of retries.
Restoring an earlier version needs testing too. Rolling back the entire record could remove a person's correction unrelated to the AI error. Establish the scope of repair, the person authorised to make it and the material needed to reconstruct the situation. Change history helps explain what happened, but seeing it does not prove that the available rollback operation preserves the other arrangements.
When the record and the outgoing message are in different tools, identify them in the evidence as well. A matching time in two logs is not enough to link operations to a particular version of an order. The person reconstructing the trial should be able to find the read, proposed change, write result and any message without guessing from similar case titles. That does not require copying all customer content into every report.
Add an explicit acceptance condition to the record of the trial: in the tested cases, another person's confirmed change survived, or the write was stopped and the conflict resolved before further action. List the operations not tested. The scope of single-record trials can be prepared with Aeneia Audit using material from the actual workflow. Protecting one record does not establish consistency across all related orders, stock and documents; their concurrent update needs separate examination. At Aster the result can be checked concretely: the record, the instructions and the message sent should all show the same agreed time.







