The Line Missing From Your Quote: What AI Features Cost After Launch
When you commission a system, the quote usually lists design, development, testing and hosting. If the system includes AI features, such as replying to customers, summarising documents or generating content, one cost is often left off: AI running costs. Every time a user presses the button, the system pays the model provider, and the more the feature is used, the more you pay. A survey published by research firm Futurum in September 2026 found that 46.9% of enterprises are already over budget on AI. This article explains why the number is hard to estimate, and four things to ask your developer to put in writing before you sign.
Why AI running costs are different from ordinary software
Traditional software has fairly fixed running costs: hosting, a domain, a maintenance contract. Add ten users and the bill barely moves. AI features are priced by usage. The provider charges for the amount of text the model processes, counted in units called tokens.
That creates a counterintuitive pattern: the unit price falls while the total bill rises. eeNews Europe reported that Goldman Sachs Research estimated in May 2026 that agentic AI could drive a 24-fold increase in token consumption by 2030, while chip improvements cut the cost per token by 60% to 70% a year. The unit gets cheaper, and usage grows faster.
Futurum surveyed 1,636 enterprise technology decision makers worldwide. The results:
- 46.9% say AI spending is over budget, including 11.3% who are substantially over
- 31.8% are roughly in line with plan
- 10% have no formal AI budget at all
Large companies are not exempt. The same eeNews Europe report notes that Uber reportedly used up its 2026 AI budget within the first few months of the year.
How the bill grows: four common causes
In our development experience, four things usually decide what an AI feature costs:
- Number of uses: the better the feature, the more people use it. That is good news, and the cost rises in proportion.
- Length of each request: to get accurate answers, systems often send conversation history and related documents to the model. The more that travels with each request, the more each one costs.
- Number of steps: an AI agent may call the model a dozen times to complete one task, which costs far more than a single question and answer.
- Model tier: the most capable models can cost several times to tens of times more than lightweight ones. Using the most expensive model for every task is the most common waste.
None of this shows up in a prototype, because only a few people are testing. The bill appears after launch, when the whole company or every customer starts using it.
Four things to have in writing before you sign
1. Estimated cost per action
Do not just ask "roughly how much a month". Ask for the cost of one customer reply, or one document summary, then multiply by the usage you expect. With a unit cost, you can work out the total for different scenarios yourself.
2. Usage caps and alerts
The system should let you set a daily or monthly spending cap, notify the owner as it approaches, and have a defined response when it is exceeded, such as switching to a cheaper model or pausing non-essential features. An AI feature without a cap is a credit card without a limit.
3. Model tiers and caching
Use lightweight models for simple classification and summaries, and premium models only where deep reasoning is needed. Answers to repeated questions can be stored instead of recalculated. These choices belong in the build. Retrofitting them costs far more.
4. A cost dashboard
You should be able to see at any time which feature costs the most, which team uses the most, and whether the trend is rising or flat. A cost you cannot see is a cost you cannot manage.
All four belong in the requirements. We covered what a specification should spell out in why the brief matters most now that AI writes code, and running costs deserve their own section in it.
AI does the computing, people decide whether it is worth it
Cost control is ultimately a question of trade-offs, not technology. eeNews Europe quotes Uber president and COO Andrew Macdonald as saying the company does not yet see a clear link between higher token consumption and more useful consumer-facing features. Spending more does not mean building better.
The division of labour looks like this:
- AI handles: large volumes of repetitive work, choosing a model according to the rules you set, and logging every use.
- People handle: deciding which features deserve AI, how much each is worth, where the cap sits, and reviewing regularly what the money has bought.
Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, with escalating costs among the main reasons. Settling running costs early is the most direct way to keep a project from ending up there. If you are still weighing a custom build against an off-the-shelf tool, see custom software or SaaS: how to decide in 2026.
Conclusion: a good quote tells you what happens after launch
A quote that lists only development fees tells half the story. The value of an AI feature starts to build after launch, and so does the cost. TitApeX believes "AI handles perfection, people handle what moves us". On budgets, people have one more job: knowing where each dollar goes and whether it was worth it.
If you are planning a system with AI features and want the running costs estimated before work begins, take a look at our services or get in touch.
Comments