{"id":1355,"date":"2026-10-19T06:00:00","date_gmt":"2026-10-19T06:00:00","guid":{"rendered":"https:\/\/nexpert.fi\/?p=1355"},"modified":"2026-10-19T13:14:08","modified_gmt":"2026-10-19T13:14:08","slug":"the-real-cost-of-ai-part-2-ai-project-budgeting-and-ai-finops","status":"publish","type":"post","link":"https:\/\/nexpert.fi\/en\/tekoalyn-todellinen-hinta-osa-2-tekoalyprojektin-budjetointi-ja-ai-finops\/","title":{"rendered":"The real cost of AI, part 2: AI project budgeting and AI FinOps"},"content":{"rendered":"<p class=\"wp-block-paragraph\">The monthly cost of an AI service is usually easy to find. The harder question is what it costs to integrate it into everyday work. Someone has to organize the data, define the workflow, test the results, and help people use the solution. These tasks may not be explicitly listed in the software vendor\u2019s quote, but the organization still pays for them.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In the first part of the series, I discussed tokens, credits, and artificial intelligence pricing models. In the previous article, we examined the selection of the first development target. Once the target is defined, the next task is to assess what it costs to implement and use, as well as under what conditions the investment is worthwhile.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Why is a mere license budget not enough?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A user-specific monthly license can be an absolutely effective solution for an expert\u2019s daily work. When artificial intelligence is connected to other systems or tasked with performing tasks in the background, additional services that are billed based on usage can be added. Fixed fees therefore may not disappear, but new expenses will accumulate on top of them.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.finops.org\/wg\/finops-for-ai-overview\/\" target=\"_blank\" rel=\"noopener\">The FinOps Foundation\u2019s overview<\/a> describes AI costs as a combination of several pricing models. Charges can arise from, for example, licences, processed tokens, API calls, computing capacity and storage. The budget must therefore separate fixed fees from costs that vary with usage.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">A cheaper model call does not mean a cheaper whole<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">There have been major changes in the operating costs of artificial intelligence. <a href=\"https:\/\/hai.stanford.edu\/ai-index\/2025-ai-index-report\" target=\"_blank\" rel=\"noopener\">Stanford\u2019s 2025 AI Index report<\/a> shows that the inference cost of producing GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. The figure describes the cost trend for one level of performance, not a price drop of the same size for all AI solutions.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The total costs of the organization may still increase if the number of materials to be processed, users, or automation workflow steps increases. In addition to the price of a single model call, the cost of the entire task should also be considered. In a design office, a useful review unit could be, for example, one processed document whose result has been reviewed by an expert.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Budget the entire work, including the time of your own staff<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">There is no universally applicable percentage allocation for the costs of an artificial intelligence project. The implementation of the finished service, the construction of document search, and the development of one\u2019s own model are different projects from each other. The budget should be built based on the necessary work and the quantities of them.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For practical budgeting, I would use the following breakdown:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>Cost item<\/th><th>What is included in that?<\/th><\/tr><tr><td>Defining the problem and the process<\/td><td>Determining the current situation, objectives, responsibilities, and defining the acceptable outcome<\/td><\/tr><tr><td>Preparation of the material<\/td><td>Data collection, cleaning, updating, and access rights<\/td><\/tr><tr><td>Implementation and testing of the solution<\/td><td>Choosing a service or model, guidance, workflows, and evaluating the quality of the results<\/td><\/tr><tr><td>Integrations and implementation<\/td><td>Interfaces, user management, data security, documentation, and moving to production<\/td><\/tr><tr><td>People's work and learning<\/td><td>The involvement of experts, training, practice and changing working practices<\/td><\/tr><tr><td>Continuous operation and maintenance<\/td><td>Licenses, consumption fees, monitoring, support work, repairs, and testing of changes<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Each row requires an assessment of workload, cost, and the person responsible. Additionally, it is necessary to distinguish one-time work from recurring tasks. For example, the initial cleaning of the data belongs to the onboarding process, but maintaining its timeliness continues after the project ends.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.rand.org\/pubs\/research_reports\/RRA2680-1.html\" target=\"_blank\" rel=\"noopener\">A study published by RAND in 2024<\/a> interviewed 65 experienced AI experts about why projects fail. Key problems included an unclear problem definition, insufficient data and a weak link between the solution and practical work. For budgeting, this means that real working time must be set aside for defining the problem and for the experts\u2019 involvement.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Hidden costs are often found between the work stages<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Consider the example of an AI system that compiles a list of defects from design documentation. Before processing, someone must ensure that the correct document versions and agreed inspection criteria are included. After processing, an expert evaluates the findings and separates the real defects from erroneous interpretations. The workload involved in these steps determines much more than the price of just one model call.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In the budget, I would like to highlight the following points in particular:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>The work of one's own staff:<\/strong> searching for, defining, testing and providing feedback on the materials.<\/li>\n\n\n\n<li><strong>Inspection and repairs:<\/strong> checking results, handling exceptions and re-running failed runs.<\/li>\n\n\n\n<li><strong>Parallel work during commissioning:<\/strong> The old and new ways of operating may have to be used at the same time for a short period.<\/li>\n\n\n\n<li><strong>Changes:<\/strong> The documentation, instructions, interfaces, and models may change, in which case the solution must be tested again.<\/li>\n\n\n\n<li><strong>Increased usage:<\/strong> A larger number of documents or longer datasets may change the original cost estimate.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">It is not worth giving these a general buffer percentage without justification. A better way is to name the uncertainty and assess its impact: what does it cost, for example, to build one additional integration or to reprocess the data? This way the allocation is based on the identified work.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Make three versions of the budget<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Instead of a precise final figure, I would first make a conservative, likely, and high consumption estimate. In these, we change concrete assumptions, such as the number of documents being handled, the time spent on inspections, and the proportion of re-processing. Simply increasing the final figure does not yet tell us where the risk comes from.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Below is <strong>a hypothetical calculation example<\/strong>, not a market price estimate or a promise of what a particular solution costs. Let\u2019s assume that 6,000 euros are reserved for the implementation of a limited document workflow, including the employee\u2019s own work. The ongoing costs consist of the following:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>Monthly expense<\/th><th>Example assumption<\/th><th>Cost<\/th><\/tr><tr><td>Fixed service fees<\/td><td>Total \u20ac150\/month<\/td><td>150 \u20ac<\/td><\/tr><tr><td>Processing to be invoiced based on usage<\/td><td>1,000 documents \u00d7 0.10 \u20ac<\/td><td>100 \u20ac<\/td><\/tr><tr><td>Inspection of results<\/td><td>1,000 documents \u00d7 2 minutes, cost of the work 60 \u20ac\/hour<\/td><td>2 000 \u20ac<\/td><\/tr><tr><td>Maintenance and support<\/td><td>3 h \u00d7 60 \u20ac<\/td><td>180 \u20ac<\/td><\/tr><tr><td><strong>Total<\/strong><\/td><td><\/td><td><strong>2,430 \u20ac\/month<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">With these assumptions, the costs for 12 months of operation are 29,160 euros. When the commissioning costs are added, the total cost is 35,160 euros. The calculation does not take into account VAT.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In this example, the largest ongoing cost is the human inspection time. If the inspection time were one minute longer, the monthly cost would increase by 1,000 euros. Therefore, it is essential to carefully measure the time spent on the inspection in the pilot.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">AI FinOps means the joint monitoring of costs and benefits<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">In practice, AI FinOps means managing the costs of artificial intelligence and the value derived from it in collaboration with business, finance, and technical management personnel. <a href=\"https:\/\/www.finops.org\/wg\/finops-for-ai-overview\/\" target=\"_blank\" rel=\"noopener\">FinOps Foundation<\/a> emphasizes the visibility of costs, the definition of responsibilities, and the integration of usage into business goals. It is not just about reducing the bill; it is about understanding what can be achieved with the money.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In a small organization, I would start with a simple monthly report. The same line would record the number of tasks processed, total costs, the proportion of acceptable results, and the time the person spent working. In addition, a person would be appointed to investigate any discrepancies and update the forecast.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The unit to be monitored could be <strong>the cost approved per end result<\/strong>. If a cheap solution requires a lot of repairs, its actual unit cost may be higher than the cost of a more expensive service. Therefore, price and quality must be considered together.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Optimise the workflow first, then choose the model<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Reducing costs starts with eliminating unnecessary work. Should the same material be processed again without any changes? Should the entire document collection be sent to the model even though the task only concerns one part? Do we need artificial intelligence at all, or can the process be carried out in a clear, regular manner?<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Model routing sends each task to a model suited to it. <a href=\"https:\/\/arxiv.org\/abs\/2406.18665\" target=\"_blank\" rel=\"noopener\">In the RouteLLM study,<\/a> routing cut costs to less than half in some tests while response quality was maintained. The result does not mean that every organisation\u2019s total AI costs would be halved.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Routing, cache, and a centralized management layer also require implementation and maintenance. In a small application, their cost may exceed the savings achieved. It is advisable to implement a technical additional layer only when there is a need for it and sufficient usage volume.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">When does an investment pay off?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">There can be no promise of a general three-year return rate for AI investments. Profitability depends on the initial situation, usage volume, quality, and whether the freed up time can be put to useful use. A comparison must be made between the same workload and the same quality level.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Let us continue with the previous hypothetical example. If the current work takes five minutes per document and the cost of the work is 60 euros per hour, processing a thousand documents costs 5,000 euros per month. The monthly cost of the new operating method would be 2,430 euros under the assumptions of the example, so the calculated difference would be 2,570 euros.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In this case, the 6,000-euro rollout would pay for itself in about 2.3 months of use <strong>at the full assumed usage level<\/strong>. If the current workload only took three minutes per document, the monthly difference would be 570 euros and the corresponding time would be approximately 10.5 months. Delays in implementation and the gradual increase in usage would prolong the repayment period.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Freed-up working time does not automatically reduce the payroll either. The benefit may be seen in the fact that the same staff members are able to handle more material, reduce overtime, or focus on more demanding specialist work. These benefits should be distinguished from direct financial savings.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What should be included in the budget before a decision is made?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Before making a decision on implementation, I would ensure that the following questions are answered. You don't need to know everything exactly, but you need to identify uncertain assumptions. The pilot's role is to provide better information about them.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Which defined task changes, and what does it cost to do it now?<\/li>\n\n\n\n<li>What are the one-time setup costs and ongoing monthly costs?<\/li>\n\n\n\n<li>How much time is needed for your own staff?<\/li>\n\n\n\n<li>At what usage rate and quality level is the solution worth it?<\/li>\n\n\n\n<li>Which assumptions can change the calculation the most?<\/li>\n\n\n\n<li>Who monitors costs and decides whether to continue, repair, or terminate?<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">A good budget makes assumptions visible and helps decide what is worth trying next. However, it also needs evidence of what can be achieved with the money. In the following article I will discuss measuring AI transformation: how to distinguish the increase in usage from real benefits and track the change in everyday work.<\/p>\n\n\n\n<div class=\"wp-block-buttons is-layout-flex wp-block-buttons-is-layout-flex\">\n<div class=\"wp-block-button\"><a class=\"wp-block-button__link wp-element-button\" href=\"https:\/\/nexpert.fi\/en\/ajanvaraukset\/\">Ask for more information<\/a><\/div>\n<\/div>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>","protected":false},"excerpt":{"rendered":"<p>Teko\u00e4lypalvelun kuukausihinta l\u00f6ytyy yleens\u00e4 helposti. Vaikeampi kysymys on, mit\u00e4 maksaa saada se osaksi arjen ty\u00f6t\u00e4. Jonkun pit\u00e4\u00e4 j\u00e4rjest\u00e4\u00e4 aineistot, m\u00e4\u00e4ritell\u00e4 toimintatapa, testata tulokset ja auttaa ihmisi\u00e4 k\u00e4ytt\u00e4m\u00e4\u00e4n ratkaisua. N\u00e4m\u00e4 ty\u00f6t eiv\u00e4t v\u00e4ltt\u00e4m\u00e4tt\u00e4 n\u00e4y ohjelmistotoimittajan tarjouksessa, mutta organisaatio maksaa niist\u00e4 silti. Sarjan ensimm\u00e4isess\u00e4 osassa k\u00e4sittelin tokeneita, kreditej\u00e4 ja teko\u00e4lyn hinnoittelumalleja. Edellisess\u00e4 kirjoituksessa tarkasteltiin ensimm\u00e4isen kehityskohteen valintaa. &#8230; <a title=\"The real cost of AI, part 2: AI project budgeting and AI FinOps\" class=\"read-more\" href=\"https:\/\/nexpert.fi\/en\/tekoalyn-todellinen-hinta-osa-2-tekoalyprojektin-budjetointi-ja-ai-finops\/\" aria-label=\"Read more about The true cost of artificial intelligence, part 2: Budgeting for an AI project and AI FinOps\">Read more<\/a><\/p>","protected":false},"author":6,"featured_media":1397,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[33],"tags":[236,224,238,237],"class_list":["post-1355","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-kayttoonotto-projekti","tag-ai-finops","tag-kustannusoptimointi","tag-smart-routing","tag-tekoalybudjetti"],"_links":{"self":[{"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/posts\/1355","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/comments?post=1355"}],"version-history":[{"count":6,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/posts\/1355\/revisions"}],"predecessor-version":[{"id":1466,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/posts\/1355\/revisions\/1466"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/media\/1397"}],"wp:attachment":[{"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/media?parent=1355"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/categories?post=1355"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nexpert.fi\/en\/wp-json\/wp\/v2\/tags?post=1355"}],"curies":[{"name":"WP","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}