When to switch off the old system after go-live

You switch off the old system on a date you set before go-live, with a written exit criterion, and not when everyone feels comfortable. While the old system stays available, it competes with the new one on every entry, and a study of a system adoption at a telecommunications company suggests that preference for the new system falls the more useful the old one looks.
Why does the old system keep competing after go-live?
Because the operator already knows the old system and is still learning the new one. On a Friday at five, with the queue full, many people take the path that does not require thinking.
The new system arrives with a learning cost. For the operator, the old one carries no learning cost. Every entry made in the old system is a small vote, and the tally accumulates: part of the operation is recorded in one place, part in the other, and anyone who needs a complete number has to add the two. That reading is ours, from architecture, not the result of a study.
In practice, go-live stops being the day the new system takes over. It becomes the day the operation gets two records. The new system usually works. What is missing is the architectural decision about which record counts, and that decision is human: no tool, new or AI-powered, makes it for you. It is the same reasoning as in system adoption: why go-live is not the end: a login to the new system does not prove the work happens there.
What does the research say about running both systems?
There is little direct research on this, and what exists has limits. Two studies help explain the mechanism, and a third serves only as a sense of cost.
Ridings and Geffen, in the Journal of Information Technology Theory and Application in 2000, studied a large telecommunications company that adopted a new customer service system while keeping the old one running in parallel. Users' preference for the new system rose with the usefulness they perceived in it and fell with the usefulness they perceived in the old one. The published abstract does not state the sample size, and it is a single company. On that basis the study indicates a direction, not a magnitude: while it exists, the old system enters the user's comparison.
Polites and Karahanna, in MIS Quarterly in March 2012, address inertia. According to the paper, habitual use of the current system, perceived transition costs and sunk costs feed inertia, and inertia explains the effect of those factors on acceptance of the new system. The empirical results, in the authors' words, largely support the proposed relationships. We cite the direction of the effect and not its size, because the abstract does not quantify it.
As an order of magnitude for the weight of operations and maintenance, the GAO reported that the US federal government spent about 75% of its fiscal year 2015 IT budget on that category, and that 5,233 of roughly 7,000 IT investments spent all of their funds on it. The report covers every system in operation, not overlap after go-live, and its scale is far larger than any mid-sized company's. It only serves as a reminder that a system left on has a continuing cost.
What changes when the overlap has an end date?
What the operator understands as a choice. With a date set, the old system stops being a permanent alternative and becomes a temporary resource, and the team's question changes from which of the two to use to how to reach the date.
None of the studies above tested a switch-off date. What follows is architectural design, derived from the mechanism they describe, and not a measured result. Four pieces support that design.
The first is the window. It is short and defined before go-live, with the date written in the plan and communicated by the line manager of the people who operate. Whoever is asked for results also owns the date, as we showed in why the line manager decides adoption.
The second is the exit criterion, written per workflow. Instead of "the team is comfortable," something verifiable: every order opened in the last two weeks was created in the new system, and no parallel control holds a record the new one lacks.
The third is a source of truth per data item. Each piece of information has an owning system from day one, and the other only reads. Without that, reconciliation becomes routine and the window never closes. Migrating off spreadsheets follows the same rule, detailed in migrating off spreadsheets without stopping the operation.
The fourth is the old system in read-only mode before switch-off. It stays searchable for a while, for history, but nobody enters data in it. This is the step that makes the date real, because it removes the option of going back out of habit.
What does this look like in a mid-sized company?
Imagine, as a hypothesis, a regional trucking company with 90 people that tracks workshop work orders in a spreadsheet and moves to an internal system. The numbers below are design parameters, not measurements.
The plan sets 30 days of overlap. In the first week, every new entry is created in the system, and the spreadsheet only receives corrections to what already existed. The exit criterion is two consecutive weeks with no work order opened outside the system. The workshop supervisor, who runs the Monday meeting, builds the agenda from the system from day one. From day 16 the spreadsheet is read-only. On day 31 it is retired, provided the criterion has been met.
If a mechanic is still logging in the spreadsheet in week three, the plan treats that as information, not indiscipline. Some step in the system's workflow does not solve what he needs, and that is where the design will be fixed. This is the pattern described in process exceptions: the path nobody logs.
What if the exit criterion is not met by the date?
You postpone, but with a rule: a new date, the cause recorded and an owner to remove it. What does not work is postponing without a new date. That path returns the operation to the state where the old system is permanent.
There are also cases where a longer overlap is right. A financial close that has to be reconciled between the two systems is one example. Even there, the overlap has a limit and ends when the difference between the two reaches zero for two closing cycles, or whatever criterion finance sets.
Frequently asked questions
My team does not accept change. Won't switching off the old system cause pushback?
The studies cited do not measure pushback. By design, the owner of the date should be whoever answers for the operation's results, and innovation joins as the one who writes the exit criterion together with them. A deadline announced before go-live gives the operator time to prepare. The date alone does not solve adoption: the per-workflow criterion and fixing what the new system still does not cover do that work.
Isn't running both systems safer, and doesn't it take less of the team's time?
By design, it takes more. Every duplicated entry is operator time, and reconciling the two becomes recurring work. Overlap is safer only when it is short, has an owner for each data item and ends on a known date.
How do I know whether the switch-off date is working?
Measure per workflow. Count how many orders run end to end in the new system and list the parallel controls still in circulation. If the number of outside controls falls to zero and the time to information shrinks, the design is working. If it does not fall, the workflow design still needs adjusting, as measuring real use shows.
The next step
If your new system has lived alongside the old one for months, the Operational Architecture Diagnostic, a 30-minute conversation, is meant to assess with you whether the switch-off date and exit criterion are well designed: book the diagnostic.
Related case studies

Did you recognize your operation in this article?
Book the Operational Architecture Diagnostic: 30 minutes to map where your operation's bottleneck is. No proposal in the first conversation.
Book a diagnostic