Cloud migrations rarely stall because the technical plan is wrong. They stall because the business case was built by engineers for engineers, and the finance director asked a question nobody had modelled.
We have reviewed enough migration proposals to recognise the shape of one that will not survive scrutiny. It compares a current monthly invoice to a projected cloud invoice, subtracts one from the other, and calls the difference savings. That comparison is almost always incomplete in both directions.
The four numbers that decide the case
A defensible business case stands on four figures, each with a stated assumption set.
1. Current cost of running the estate. Not just hosting. Add the hardware refresh cycle, data centre space and power, software licensing tied to owned hardware, backup and disaster recovery provision, and the proportion of internal engineering time spent on keeping the lights on rather than on work the business asked for.
2. Target run rate in the cloud. Priced at realistic utilisation, not at a theoretical right-sized state nobody will actually operate. Include the managed service layer, egress, storage growth and the observability tooling that replaces whatever came bundled with the hardware.
3. One-off migration cost. Assessment, engineering time, parallel running, data transfer, retraining, and the disruption buffer for the workloads that take twice as long as planned. If this line is understated, the payback period is fiction.
4. Risk-adjusted cost of doing nothing. The figure finance rarely sees. Remaining hardware support costs, the probability-weighted cost of an outage on the current platform, the premium paid for emergency capacity, and the staff retention risk of asking a team to maintain infrastructure it no longer believes in.
When all four are on one page, the conversation changes from "is cloud cheaper?" to "which of these options carries the least risk for the cost?" That is a question a finance team can actually answer.
Where the savings really come from
Savings do not arrive when workloads start running in the cloud. They arrive at decommissioning.
Parallel running is the single most common cause of a migration that costs more than it promised. The old platform stays live "just in case", the hardware contract renews because nobody cancelled it, and the projected saving never lands. Every migration plan should name the decommissioning date per workload before the project starts, with an owner who is accountable for it.
Beyond decommissioning, the realistic savings levers are:
- Right-sizing against measured utilisation rather than peak guesses.
- Storage lifecycle policies that move cold data automatically.
- Reserved capacity or committed use for the stable, predictable baseline.
- Retiring duplicate tooling that accumulated across on-premises and cloud estates.
- Reducing unplanned work by making failure modes observable before they are incidents.
A migration that leaves the old estate running has not saved anything. It has added a second estate.
Sequencing the move
A sequence that keeps both the board and the engineers comfortable looks like this:
- Assessment. Inventory workloads, dependencies and data gravity. Identify what must move, what should be retired, and what should be rebuilt rather than lifted.
- Foundation. Landing zone, networking, identity, logging and cost controls in place before the first workload arrives. Retrofitting governance after migration is significantly more expensive.
- Pilot. One representative, low-risk workload moved end to end, including operations, backup verification and support handover. The pilot exists to test the operating model, not the technology.
- Wave migration. Grouped by dependency, in order of business risk, with a defined rollback point per wave.
- Decommission. The least glamorous phase and the only one that produces the savings in the business case.
Estimation errors worth pre-empting
- Data transfer and egress treated as a rounding error. For data-heavy workloads it can exceed compute.
- Latency-driven redesign discovered late, forcing re-architecture mid-project.
- Licence portability assumed rather than confirmed with the vendor in writing.
- Skills gap ignored: the target platform needs people who can operate it on day two, not just consultants who can build it on day one.
A migration is a change of operating model, not a change of address. Build the case around that, and the approval conversation becomes straightforward. Build it around a hosting invoice comparison, and it will be sent back for rework within a week.