The cost of using AI is about to skyrocket, but do we have the appetite to pay?

We have been gorging ourselves at an all-you-can-eat buffet, but the menu is about to change. AI tools that have been discounted will soon start charging based on your usage, raising serious questions about our appetite to pay.

The answer could have big implications for the future of artificial intelligence (AI).

Millions of businesses worldwide use software that now contains AI models. If you use services like ChatGPT, Claude, Workday, or anything from SAP’s Business AI solutions, you’re probably one of them. So far, AI software companies have subsidised the use of their models because they want people to use them. But this is starting to change. Software companies are ending some free or low-priced AI offers and are now starting to charge businesses based on usage, which pushes employers to reassess the cost of tools already built into their daily work.

According to Harvard Business Review, suppliers of AI technology – such as OpenAI and Anthropic – have so far absorbed computing costs to attract customers before moving towards charges based on consumption.

For businesses, it’s been an all-you-can-eat buffet. For a monthly subscription fee, they’ve got AI to do writing, analysis, summaries and coding. Staff have been empowered to experiment without every request they send to AI companies adding to the bill. Over time, these optional extras have become part of how business works.

But the restaurants are about to start charging by the plate, which means the cost of using AI is about to skyrocket.

And the cost of using an AI agent will be hard to predict. If you use a simple AI model to do a basic web search, it might not cost that much. But if you give it a complicated task that involves coding, or acting autonomously, then it costs a lot more, and might not be able to budget for how much work it’ll go away and do without human oversight.

Mistakes and repeated attempts can also incur charges.

Costs also depend on which tools employees choose to use, and how often they use it. Even if the per-use charge is small, it can really add up if an entire workforce uses it every day.

Under the current subsidies, AI software companies bear more of the cost of heavy use. For example, OpenAI incurs approximately two dollars in costs for every dollar of revenue generated by its AI services. Now that they are going to start charging on a per-use basis, more of that cost will be passed onto the customer.

There are plenty of examples of this happening right now. In June 2026, GitHub started charging businesses that use its Copilot coding assistant. The company said it had absorbed rising computing costs and the previous system was “no longer sustainable”. Monthly subscription prices stayed the same, but use beyond agreed allowances now attracts additional costs.

Anthropic’s usage-based Claude Enterprise plans have two costs: a fixed fee for each user’s account and separate usage charges. The account fee lets users access Claude, but includes no usage. So even after paying the subscription, businesses still have to pay for the text Claude reads and generates. Depending on the plan, companies buy credits upfront or receive a monthly bill based on how much they have used.

More and more providers are changing their pricing models. Harvard Business Review reports that business software supplier Workday’s waiver of charges above certain software connection limits ends on 31 January 2027.

And business software supplier SAP’s pricing guide says that, from the end of 2026, customers will have to pay to run AI agents they build themselves.

The shift follows a period in which AI companies strongly encouraged businesses to use their products. In turn, business executives urged employees to use the technology and spend freely on experiments. However, businesses quickly reached the limits of their AI allowances. Uber exhausted its annual AI coding budget within just four months.

The Federal Reserve Bank of Atlanta, part of the US central banking system, also tracks how businesses are spending on AI. Its survey found that firms planned to spend US$2,068 per employee on AI in 2026, up 52 per cent from 2025. That includes equipment, training and support. The increase measures spending, which can rise through greater use even when price per request falls.

All of this has significant implications for the cost of using AI. If it costs more to get the job done, then will it really make sense for businesses to replace staff with AI? The evidence so far suggests that the panic about AI replacing jobs is overblown. Even an executive of one of the world’s leading chip manufacturing and AI companies, NVIDIA’s Bryan Catanzaro, said his team’s use of AI computing costs more than paying its staff . Although he was only talking about one team within the company, it shows that AI is not automatically cheaper than paying human staff. Employers may find themselves in a position where they have to pay for both the AI technology and the people who check, correct and maintain the work it does.

The longer-term risk is that businesses build operations around prices that are only going to increase. If experienced staff are pushed out because essential work is done by AI , it will be hard to go back – a discount can expire a lot faster than it takes to rebuild a workforce.

As discounted offers end, employers will need to reassess the costs of using AI. This includes not only the cost of the software, but the cost of paying a human to check and correct its work.

AI may still deliver worthwhile gains. The test is whether those gains survive the end of the unlimited buffet.