Most of us by now will have tried using an AI chatbot like ChatGPT or Claude or Perplexity or Grok. They are hard to avoid, with, for example, Google now putting AI-generated answers at the top of its queries, with sometimes amusing results. However, if you are to implement one of these generative AI tools in your business, what are the actual costs?
Firstly, there are explicit costs. You can play with tools like ChatGPT and Claude for free, but there are limits on the number of interactions per day, for example 40 short messages per day for Claude at present. For any serious usage then you need to buy a license. Claude charges $200 a month at present for example. Enterprise licences can be negotiated if there are many users. This quickly adds up. At present OpenAI’s ChatGPT enterprise licence works out at $144,000 a year for 200 employees, or a list price of over $14 million for an organisation with 200,000 employees, though doubtless there is scope for negotiation for larger companies.
Once you have your licence for your chosen AI, you need to factor in the processing costs. There will be charges from your cloud provider for compute costs, bandwidth and storage for things like conversation logs. A large part of any computer project is staff, whether the cost of employees or of external consultants to help you. Depending on the scale and complexity of the project, building an enterprise chatbot might cost anything in labour costs from a few thousand dollars to over $1 million. As well as the hourly costs of people, there may be costs of integration into existing corporate infrastructure such as Slack, or possibly purchasing a vector database if you want to do retrieval augmented generation (RAG) to customise the chatbot with company-specific files or data. You will need to do testing, and carefully consider security and privacy. For example, if a chatbot is to have access to a corporate database, such as a customer order database, then you need to make sure that it is secure. This may include encrypting data in transit and at rest, and conducting penetration testing, as well as considering GDPR or similar compliance. Chatbots have their own security issues, such as the risk of prompt injection attacks, data leaks or social engineering attacks.
Once a chatbot is up and running there are ongoing costs to consider. Every time a customer asks a question there is a cost somewhere, and the chatbot will need monitoring and support. There will be new releases of software and bug fixes to deal with, just as with any software project. There may be a need for marketing or brand review to ensure that the chatbot application is in line with corporate brand standards, and possibly a legal review. For example in one famous case an Air Canada chatbot gave poor advice to a customer, resulting in a court case which Air Canada lost. If the chatbot is successful then it will engage well with customers, which in itself has a cost, with every customer interaction incurring a small cost in processing. A chatbot question costs around ten times as much processing as a typical Google query, as there are computationally complex. All this ignores indirect costs such as the need for training of staff in AI, or the cost of handling any incidents that occur such as due to hallucinated output from the chatbot. The content of the chatbot conversations will need to be logged, monitored and analysed. These cost categories all assume using an off-the-shelf chatbot. If you intend to carry out customisation such as a customised user interface, then you need to factor this in too.
Any business investment of any size should have an economic case, often called a cost/benefit analysis. You need to estimate the benefits of a proposed project and weigh against that the costs, including the ongoing costs over time as well as the original one-off investment. With these figures, you can calculate the estimated economic return of the project, including things like the project payback period, the net present value (NPV) of the investment, and the internal rate of return (IRR) of the project in order to determine whether there is a good return on investment. Not all companies take such a rigorous approach, but large corporations do, at least in the cases of the ones that I worked at (Exxon and Shell). At the time of writing, the return on investment for AI projects seems distinctly mixed. An IBM study of 2000 CEOs conducted in 2025 study reported in Fortune Magazine, found that just a quarter of their AI projects had shown a return on investment. AI is novel and has generated considerable excitement in business over the last couple of years, but it is just another technology tool, and needs to be subject to the same investment rigour as any other technology investment.
