# Four problems enterprises hit with AI. AI is moving from chat to agents that run real work. That exposes four problems: the top models are priced for the hardest work, most tasks do not need them, budgets overrun, and privacy is the top worry. This page explains each one in plain words, with its source. ### 1. Top models are priced for the hardest work. The top AI models list at up to $50 per 1M output tokens. That price makes sense for the hardest tasks. It does not make sense for every request. #### What is going on A token is a small piece of text that a model reads or writes. A chatbot answers once and stops. An AI agent does not. It plans, calls tools, reads the results and tries again, and every step carries the growing context with it. One task becomes many model calls. #### How big the effect is - About 4 times and 15 times. Anthropic reported that an agent typically uses about 4 times the tokens of a chat, and a system of several agents about 15 times. - Up to about 1,000 times for coding. A research paper from April 2026 measured agentic coding tasks at about 1,000 times the tokens of code reasoning and code chat. It found that input tokens, not output tokens, drove most of the cost. #### What helps - Count cost per task, not price per token. Price per token is a rate. Cost per task is the bill. - Send each request to the level of model it needs, not every request to the top one. - Look at input tokens as well as output tokens, because an agent re-reads its context at every step. The Decision Layer picks the model for each request, so you pay the top price only for work that needs it. - Decision Layer (https://avelin.ai/decision-layer) - Public prices (https://avelin.ai/pricing) - How cost per task is counted (https://avelin.ai/method#cost) Sources: top-model list prices, read 24 Sep and 1 Oct 2026. Anthropic, "How we built our multi-agent research system", 13 Jun 2025. Bai et al., arXiv 2604.22750, 24 Apr 2026. ### 2. Most tasks do not need the top model. Most AI work is routine: summarize, extract, classify, draft. A cheaper model can do much of it. Teams still send it all to the top model. #### What is going on Accenture's 2026 research found that 54% of AI requests went to a more capable and more expensive model than the task needed. In the same research, 90% of more than 9,000 tasks did not need a frontier model. #### Why it happens Teams do it on purpose. A top model is the safe default, and the cost of a failed workflow is higher than the cost of a few extra tokens. Nobody risks a cheaper model. The workflow could break. So the default stays, and nobody can show that a cheaper model would have been enough. #### What helps - Decide per request, not once for the whole app. An easy request and a hard request should not get the same model. - Keep hard tasks on a top model. The aim is the right level, not the cheapest one. - Check quality on your own tasks, side by side, before you change anything. The Decision Layer reads each request and picks the level it needs. A POC shows the result on your own workload, beside your current setup. - How the Decision Layer picks (https://avelin.ai/decision-layer#how) - What a POC looks like (https://avelin.ai/poc#request) Source: Accenture, The CIO's Guide to AI Tokenomics, 10 Sep 2026. Survey of 750 senior executives in 17 countries, plus Accenture's own task tests. ### 3. Budgets overrun, and cost is not tied to a task. Most teams see an AI bill. Few can say which task, agent or team caused it. #### What is going on In a 2026 survey, 79% of 500 finance leaders in the US and UK reported AI cost overruns in the past 12 months. Accenture found that only 35% of companies can calculate the cost per business outcome of their largest AI use case, and that less than one dollar in five of enterprise token spend can be traced to a quantified financial outcome. #### Why it happens A provider bills by token and by account. It does not know which team, agent or task made a call. Every new agent adds calls, so the bill grows with each one, while the link between spend and result stays missing. #### What helps - Give each team, agent or app its own key, so spend has an owner. - Set a budget and rate limits on each key, with an alert before the limit. - Report cost per task, summed over every model call in it. The AVELIN API platform shows cost by team, key and model, with budgets and caps on each key. - API platform (https://avelin.ai/api-platform) - Price a task yourself (https://avelin.ai/pricing#estimate) Sources: DoiT, survey by Sapio Research, Feb 2026, 500 finance leaders at companies with 1,000 or more employees, US and UK. Accenture, The CIO's Guide to AI Tokenomics, 10 Sep 2026. ### 4. Privacy is the top-ranked AI worry. For banks, hospitals and governments, the first question about AI is often not cost. It is where the data goes. #### What is going on In AvePoint's State of AI 2026, 76% of 750 IT, security and AI leaders said they were highly or extremely concerned about data security and privacy. It ranked first among the AI concerns in the survey. AvePoint sells data-governance software and commissioned the research. #### Why it matters more now Agents make many requests for one task, so more of your data passes through a model. The question is not only who can read a prompt. It is where the prompt runs, how long it is kept, and whether it can leave. #### What control means - What do you keep? Is a prompt stored, for how long, and is it used to train a model? - Where does it run? Which machines and which regions process it, and who runs them? - Can it leave? Can the request go on to another model or provider, and who decides? Banning AI does not remove the risk. Control does: you decide what is kept, where a request runs and whether it can leave. AVELIN answers at three levels: cloud with zero retention by default, hybrid with a privacy check you set, and on-premise on your own GPUs. - Privacy, in three levels (https://avelin.ai/privacy) - Hybrid and the privacy check (https://avelin.ai/deployment#hybrid) Source: AvePoint, State of AI 2026, research by Osterman Research, 750 respondents. ### What AVELIN does about them. Nobody risks a cheaper model. The workflow could break. So the top model stays the default. AVELIN has two products. The Decision Layer picks the right model for every request, which answers problems 1 and 2. The AVELIN platform gives you cost control and an inference layer on your own GPUs, which answers problems 3 and 4. You can start with a POC on one workload. - Decision Layer (https://avelin.ai/decision-layer) - Sovereign AI platform (https://avelin.ai/sovereign-ai-platform) - What is AVELIN (https://avelin.ai/what-is-avelin) - Request a POC (https://avelin.ai/poc#request) - Try the live demo (https://arena.avelin.ai) ### FAQ: Questions about the four problems **Why do AI agents use so many more tokens than chatbots?** A chatbot answers once. An AI agent works in a loop: it plans, calls tools, reads the results and tries again, and every step carries the growing context. Anthropic reported in 2025 that an agent typically uses about 4 times the tokens of a chat and a system of several agents about 15 times. A research paper from April 2026 found about 1,000 times the tokens of code chat for agentic coding tasks, driven mainly by input tokens. The top models list at up to $50 per 1M output tokens, so the bill grows fast. More: Problem 1 in full (https://avelin.ai/enterprise-ai-problems#token-price) **Do most AI tasks need the most expensive model?** No. In Accenture's 2026 research on AI tokenomics, 54% of AI requests went to a more capable and more expensive model than the task needed, and 90% of more than 9,000 tasks tested did not need a frontier model. Teams do this on purpose, to be sure a job finishes, because nobody risks a cheaper model and breaks a workflow. The Decision Layer picks the level each request needs, and hard tasks still go to a top model. More: Problem 2 in full (https://avelin.ai/enterprise-ai-problems#wrong-model) **Why do AI budgets overrun, and how do you see cost per task?** Providers bill by token and by account, not by task, team or agent, so a bill does not say what caused it. In a 2026 survey by DoiT, 79% of 500 finance leaders in the US and UK reported AI cost overruns in the past 12 months, and Accenture found that only 35% of companies can calculate the cost per business outcome of their largest AI use case. Cost per task adds up every model call a task makes. The AVELIN API platform shows cost by team, key and model, with budgets and caps. More: Problem 3 in full (https://avelin.ai/enterprise-ai-problems#budget) **Why is privacy the top AI worry, and what does control mean?** In AvePoint's State of AI 2026, 76% of 750 IT, security and AI leaders were highly or extremely concerned about data security and privacy, the top-ranked AI concern. Agents send many requests for one task, so more of your data passes through a model. The question is where a prompt runs, how long it is kept and whether it can leave. Banning AI does not remove that risk. Control does, and AVELIN gives three levels of it: cloud, hybrid and on-premise. More: Problem 4 in full (https://avelin.ai/enterprise-ai-problems#privacy-worry); Privacy (https://avelin.ai/privacy) **What is a token?** A token is a small piece of text that an AI model reads or writes, often a short word or part of a word. Providers bill by the token, so the length of a prompt, the context an agent carries and the length of an answer all change the bill. **What is cost per task, and why not compare price per token?** Cost per task is what one finished piece of work costs, summed over every model call in it. An agent makes many calls per task, so a low price per token can still give a high bill. Price per token tells you the rate. Cost per task tells you the bill. More: How cost per task is counted (https://avelin.ai/method#cost) Related: Decision Layer (https://avelin.ai/decision-layer); Privacy (https://avelin.ai/privacy); Pricing (https://avelin.ai/pricing); Method (https://avelin.ai/method); What is AVELIN (https://avelin.ai/what-is-avelin) ### Bring one workload. See its cost per task. Start with a POC and see the cost per task next to your current bill. - Request a POC (https://avelin.ai/poc#request) - Try the live demo (https://arena.avelin.ai) - Start building (https://avelin.ai/docs/api/quickstart) URL: https://avelin.ai/enterprise-ai-problems Updated: 3 Oct 2026