Data is worse than anyone admits
Every programme assumes the source data is broadly fine. It is not, and the discovery usually lands in the final third when there is no room left to react. Duplicate records, missing dimensions, historical entries that predate a policy change and balances that have never reconciled: all of it is normal and all of it costs time.
The prevention is cheap and almost nobody does it. Profile the data in week two, not week nine. You either find out it is fine, which costs you a few days, or you find out it is not, while there is still time to choose between cleaning it, cutting it, or moving the date.
Customisations nobody can explain
Legacy estates carry modifications whose original purpose left with the person who requested them. The programme assumption is that they must be rebuilt. Often they should not be, but nobody will make that call without an owner to sign it off.
Inventory them early and classify each one: load-bearing, habit, or unknown. Unknowns need an owner found or a decision to drop them, taken by a sponsor rather than by a developer at midnight during cutover weekend.
Cutover was never rehearsed
A cutover nobody has practised has an unknown duration, which means the window cannot be planned and the rollback point cannot be identified. That is how a weekend go-live becomes a Tuesday incident.
At least one full rehearsal, with the reconciliation run and the timings recorded. Two is better. The rehearsal is also the only way to find out whether the rollback plan works, and a rollback that has only been written down is not a rollback plan.
The people you need are busy
Programmes are planned as though the business will provide its best people on demand. Those people have day jobs, and month end does not move for a project plan.
Named availability, agreed in the plan, against the specific weeks it is needed. If your financial controller cannot be in user acceptance testing, that is a fact to schedule around rather than discover.
The plan was optimistic on purpose
The uncomfortable one. Competitive selection rewards the shortest timeline, so timelines compress to win work rather than to reflect the work. Everyone involved knows it, and the slip is priced into nobody's expectations except the sponsor's.
The defence is to ask every bidder what their plan assumes, and to treat a plan with no stated assumptions as incomplete. The firm with the longest credible timeline is frequently the one telling you the truth.