· AI · 6 min read
What Switching AI Solutions Really Costs
The simple switching calculation compares a price per user with a price per token. Those are two different things. An honest calculation has five cost blocks and no fixed saving.
Anyone calculating a switch of AI solution needs five cost blocks and one fixed time frame, not a price per user set against a price per token.
A saving in percent, or a fixed user count for break-even, can only be derived from the concrete case. Anyone who names one in advance has not calculated. They have sold.
The simplest switching calculation is a comparison: on the left the licence price per user and month, on the right a price per million tokens. The right-hand number is small, so the switch looks cheap. The calculation is still incomplete, on both sides.
Two things that are not the same
A token price is a rate per quantity. What a model request costs only follows from the quantities consumed, split by input and output. A licence describes a working environment: interface, connection to documents and mailboxes, permission management, logs. Comparing the two compares an ingredient with a dish.
The model cost is only one of five items. On its own it says nothing about total cost. What counts is everything around the model that has to be paid for and delivered.
Five blocks that belong in the calculation
| Block | What belongs in it |
|---|---|
| Current solution | Prices actually paid rather than list prices, term, usage-based items, required add-on products |
| One-off switch | Setup, integrations, data preparation, rollout and your own effort in-house |
| Ongoing operation | Scope per area and knowledge source, agreed operating services |
| Model consumption | Measured or explicitly estimated volumes, multiplied by the prices of the intended models |
| Cloud and additional services | Base cost of the cloud account, licences that stay, and custom services |
The second block easily drops out of the comparison, because it occurs only once and appears in no monthly bill. The fourth is easy to overestimate, because it is so convenient to pour into a single number.
Three calculation errors that skew the comparison
Licences that stay. According to Microsoft’s pricing page, Microsoft 365 Copilot requires a separate licence for a qualifying Microsoft 365 plan. Anyone who replaces only the AI function saves the Copilot licence, not the base package. The calculation has to know which licences keep running after the switch.
Deducting a credited pilot price twice. If a pilot is later credited against operation, its price is paid only once. The credit may therefore appear in only one place in the calculation: as a reduction of the operating cost or as a deduction from the pilot, never both. Apply it twice and the saving comes out too high by exactly that amount.
Different time frames. A switch has costs at the start and running costs afterwards. The cost of the first year of operation is therefore higher than that of a later year, provided running costs stay the same. Summed over the period, only the total counts. Both variants must be calculated over the same period and with the same start of operation, otherwise the one whose costs fall later wins.
Two cases that end differently
Two patterns show the difference without concrete numbers. They are thought models, not customer cases.
In the first case many people use AI occasionally, mostly for routine work. Model consumption is low and spread widely. Here several quantities decide together. How much of the licence goes unused today? How high is the one-off switching effort relative to the term? And how high are the running costs for operation and cloud, which arise even at low consumption?
In the second case a few people use AI daily for demanding tasks. Model consumption is high and concentrated, and more expensive models stay sensible for part of the tasks. Here a per-head licence can be cheaper than billing by consumption.
Which case applies is decided by usage in your own organisation. That is why there is no user count above which switching always pays off.
What you should have in hand before deciding
Prices actually paid and their term. A list of the functions you really need, and a check that both variants cover them. Measured or clearly marked estimated usage volumes. A shared time frame. And the rule that a saving is named only once these four points stand.
Where usage volumes are missing, the first step is not a calculation but measuring your own usage. Without it, every saving remains an assumption.
Frequently asked questions
Which costs belong in the calculation for switching an AI solution? Five blocks: the current solution, the one-off switch, ongoing operation, model consumption, and cloud plus additional services. If one block is missing, the comparison is incomplete.
Is there a user count above which switching always pays off? No. The break-even point depends on scope, usage patterns and time frame. A fixed number without those details is a claim, not a calculation.
Why are token prices not enough for the comparison? A token price is a rate per quantity; the cost of a request only follows from consumption. A licence describes a working environment with integrations, interface and administration. Comparing the two without matching the scope of functions distorts the result.
Sources
- Microsoft 365 Copilot requires a separate licence for a qualifying Microsoft 365 plan (retrieved 24 September 2026): Microsoft, Microsoft 365 Copilot pricing page
Related: Paid per licence, used per task, Copilot: What the Admin Center Shows, What happens when the provider drops out and What share of your AI tasks needs the most expensive model.
