Finance teams usually approve software by looking at a license line. The invoices that cause trouble tend to come from somewhere next to it: an AI feature whose monthly charge tripled without any new users, or a backup product whose storage bill quietly passed the cost of the software itself.
Both come from pricing models where the headline number measures something other than what actually drives spend.
What a Token Actually Measures
AI model providers almost all use token-based pricing. A token is a chunk of text, roughly three quarters of an English word on average, and providers charge separately for tokens sent in and tokens generated. Output tokens usually cost several times more than input tokens.
The part that surprises people is what counts as input. A model has no memory between calls, so every request resends the full instructions, any reference documents, and the conversation so far. A chat that runs twenty turns pays for the first message twenty times.
Languages matter too. Text in Hindi, Japanese, or Arabic often breaks into more tokens than the same meaning in English, so a product serving multilingual customers can cost more per conversation without anyone changing the feature.
Where Token Bills Actually Come From
Consider a customer support assistant. Its instructions run about 2,000 words. Each question pulls in three help articles to ground the answer. Before the customer types anything, each call might carry 5,000 input tokens. At 100,000 conversations a month with several turns each, that’s billions of tokens, most of them the same instructions sent again and again.
There are practical levers, and some matter much more than others.
Prompt caching, which many providers now offer, charges a reduced rate for repeated content at the start of a request. For an assistant with long fixed instructions, it’s often the single biggest saving. Batch processing, where requests don’t need answers immediately, typically comes at a steep discount and suits jobs like tagging tickets overnight. Routing simple tasks such as classification or language detection to a smaller, cheaper model saves money without affecting quality much.
Trimming conversation history helps as well, though cutting too much makes the assistant forget what the customer already said, which brings its own costs in repeat contacts.
The mistake to avoid is estimating cost from a demo. Demos use short prompts and short conversations. Production traffic doesn’t.
Backup Software Is the Smaller Number
When people price backup, they usually start with the cost of Veeam or a similar product. Veeam licenses most protected workloads through its Universal License, counting instances such as virtual machines, servers, or cloud workloads, typically on subscription and usually bought through resellers rather than from a public price list. Higher editions add features like more advanced recovery orchestration and security capabilities.
That license is rarely the biggest number over three years. Storage is.
Retention policies multiply data in ways that aren’t obvious at first. Keeping 30 daily restore points, 12 monthly, and 7 yearly ones means storing far more than the size of production, even with deduplication and incremental backups. Ransomware protection usually requires immutable copies that can’t be deleted for a set period, which means those copies stay on the bill whether or not anyone needs them. An offsite copy in cloud object storage adds per-gigabyte monthly charges on top.
Then there’s the repository hardware for on-premises copies, replaced every few years, and the staff time to monitor jobs that fail at 2 a.m.
Restores Have a Price
Backup costs are usually modeled for the steady state. Recovery gets less attention.
Pulling data out of cloud object storage incurs egress fees, which become noticeable when restoring many terabytes after an incident. Archive tiers are cheap to store in but charge for retrieval and can take hours before data is available, which may break a recovery time target the business has already promised. Regular test restores, the only real evidence that backups work, consume compute, bandwidth, and someone’s afternoon.
A company that discovers these costs during a real incident is in no position to negotiate.
Questions Worth Asking at Renewal
For AI features: what’s the average token count per request in production, how much of it repeats, and is caching turned on? For backup: what does retention actually multiply the storage to, what would a full restore from the offsite copy cost, and when was the last test restore?
Pricing the Thing That Changes
Both models reward the same habit. The license or unit price stays fixed for a year, while conversation length, retention periods, and data volume change every month. The companies that avoid surprises track those moving quantities as carefully as the contract price, because that’s where the next invoice is actually decided.
