A cheaper licence does not necessarily produce a cheaper IT service. Replacing it can require migration work, new skills and more administration. Keeping a product unchanged can also waste resources or leave subscriptions that no longer match usage.
The first decision is therefore to identify the expense an intervention can affect. The organisation then compares keeping and changing the service over the same period, with equivalent requirements.
Connect spending to its purpose
The invoice shows licences, hosting and consumption. Total cost adds internal work, support, security, backups and interruptions. Some expenses appear in contracts; others appear only in the hours teams spend.
Attributing cost to a service makes the figures useful. More storage may reflect valuable growth, forgotten copies or an excessive retention period. Each cause calls for a different decision. Removing a resource without understanding its role can turn a saving into an incident.
Relative size helps prioritise work. Cutting a category by 80% when it represents 3% of the budget reduces the total by 2.4%, before implementation cost. That may be worthwhile, but the amount must justify the work and associated risk.
For metered services, measurement distinguishes volume from price and usage behaviour. For per-person licences, it checks whether active accounts need the features of the subscribed plan.
Reduce AI spending while preserving quality
An AI application can pay repeatedly to prepare the same context or use an expensive model for a simple operation. Usage records should connect the expense to a task and an accepted result. Request counts alone miss input length, output length and retries.
Reuse stable input
Prefix caching can reduce the cost of identical material sent to a model. Stable instructions or a catalogue may qualify. Conditions vary by provider, including minimum length, exact matching and duration. A changing timestamp at the start of every request can prevent the expected reuse. The team measures actual cache use and its share of the total bill.
Batch work that can wait
Document extraction and overnight summaries do not always require an immediate response. Some APIs offer discounted asynchronous processing. The application must account for deadlines, volume limits and failed requests, and be able to retry only the affected items.
Choose a treatment for each step
Calculations, sorting and deduplication may need deterministic code. Interpreting an ambiguous document may justify a model. Model comparisons use representative cases and count the corrections and retries required. The useful cost is the cost of an accepted result, not just the price per million tokens.
These changes need before-and-after measurement. A lower rate can lose its advantage when the application doubles its attempts or users spend more time correcting the output.
Compare keeping and replacing the service
The same reasoning applies to virtualisation and collaboration tools. A different platform may reduce fees but require new backup integration and training. Self-hosting may suit a team with the operating skills; another organisation may prefer a subscription that includes that work.
This fictitious three-year comparison illustrates the calculation. Amounts are in Swiss francs and are not product prices. Costs identical in both scenarios are omitted.
| Three-year cost | Keep | Replace |
|---|---|---|
| Licences and subscriptions | 18,000 | 3,000 |
| Infrastructure | 4,500 | 6,000 |
| Administration and support | 9,000 | 18,000 |
| Migration | 0 | 4,000 |
| Training | 0 | 2,000 |
| Compared total | 31,500 | 33,000 |
Here, the licence saving does not cover the change. A different administration estimate could reverse the decision. Low and high scenarios help show how much the conclusion depends on those assumptions.
Requirements matter as well as price. For electronic signatures, the required evidence and legal form depend on the document. OFCOM explains the Swiss framework. Two tools can only be compared fairly after checking that they cover the same requirement.
Verify the saving afterwards
The team keeps the scope, volumes and reference period, then follows invoices and operating effort. It also examines errors and effects on users. Growth can hide a lower cost per transaction; retiring a service can lower the bill without making the architecture more efficient.
The most useful intervention has an understood cause, implementation cost and measurable effect. Those facts let the organisation decide when to optimise its current system and when a replacement warrants the investment.
Sources and assumptions
Rates depend on the model, contract and date. The table contains calculation assumptions only; an actual comparison uses the costs of the service under review.