Enterprise AI was supposed to cut costs. Instead, for many companies, it is blowing budgets.
Uber CTO Neppalli Naga recently revealed that Uber had burned through its 2026 AI budget in just a few months. And this isn’t an isolated case. Arvind Jain, CEO of enterprise AI company Glean, recently told CNBC that companies are exhausting their annual AI budgets in one or two months.
AI offers a genuine breakthrough in automation, particularly for companies that are laden with process-heavy, manual work. Its ability to handle complex reasoning, reference and apply business logic, interpret unstructured information, and act autonomously means that, for the first time, large parts of office work can be fully automated. In theory, that should break the link between company growth and operational headcount, allowing businesses to scale without costs rising at the same rate.
The economic case, however, only works if the technology costs less than the work it replaces. If companies replace human labour with AI workflows that are more expensive to run, they have not solved the cost problem – they’ve only swapped salaries for tokens. So how can companies capture the benefits of AI automation without letting costs spiral?
5 steps to controlling your AI spend
1. Use the right model for the job
The most powerful AI models can handle very complex reasoning, but they’re also very expensive. One of the fastest ways to inflate an AI bill is to default to frontier models for every task, regardless of whether the work actually requires that level of capability. Routine summarisation, classification, extraction, routing, formatting, and simple comparison often do not need the most advanced model available.
Using a frontier model for basic admin work is like asking a senior executive to do data entry. They can do it, but it is a terrible use of money. The better approach is to match the model to the task. Use the most powerful models where the work is genuinely complex: ambiguous decisions, nuanced reasoning, edge cases, or tasks where the cost of an error is high. Use smaller, cheaper models where the task is narrow, repeatable, and well-defined. This can dramatically change the unit economics of an AI workflow.
2. Use software where software is enough
AI should be used to extend what software can do, not replace it. Somewhere along the way, companies started treating AI as the answer to every automation problem, but traditional software is still better suited to many parts of a workflow. Rules, routing, permissions, integrations, status updates, reporting, data validation, and deterministic logic often do not need AI at all.
For these tasks, software is usually faster, cheaper, more reliable, and easier to control. The mistake is building workflows where AI handles everything simply because it can. A more cost-effective system uses software as the backbone, with AI applied only to the parts that genuinely require judgement, language understanding, or reasoning. That is how companies can get the power of AI without paying for it at every step.
3. Design for token efficiency from the start
AI costs are not just about which model you use. They are also about how much context you send to it, how often you call it, and how much unnecessary work you ask it to do. Agentic AI can consume far more tokens than a simple chatbot because it often involves multiple steps: reading documents, calling tools, checking outputs, retrying failed attempts, and generating intermediate work the user never sees.
That means cost control cannot be treated as an optimisation exercise after launch. It needs to be designed into the product from the beginning. Teams should be asking whether they are sending the model more context than it needs, whether documents could be filtered before being passed to AI, whether the same AI call is being repeated unnecessarily, whether outputs could be templated instead of generated, and whether they are measuring cost per completed workflow rather than just cost per model call.
Without this discipline, token usage can quietly become one of the biggest hidden costs in the system. A workflow may look efficient from the outside, but if it relies on multiple expensive model calls, repeated retries, and excessive context windows, the economics can quickly fall apart.
4. Measure AI by unit economics, not novelty
AI demos often look impressive when they show capability, but enterprise automation has to prove economics. Teams should only be using AI when it can do a task reliably, and at a lower cost than the current process. If a workflow saves ten minutes of human time but costs more than ten minutes of human labour to run, it’s not efficient automation, it’s expensive theatre.
Every AI workflow should be measured against the cost of the work it replaces. That means looking at the full cost: model usage, implementation, review, errors, retries, maintenance, and human oversight. The companies that win with AI will not be the ones that deploy it most aggressively, replacing whole teams with AI. They will be the ones that understand where it creates real operating leverage.
5. Keep humans in the right places
The goal of AI should not be to remove humans from every process. It should be to remove repetitive, manual work from people’s plates so they can focus on the judgement-intensive work that actually requires human expertise. That is not only a better vision for AI adoption; it is also a better way to control costs.
Human review should be used where it adds value: high-risk decisions, exceptions, ambiguous cases, quality control, and human interaction.
A smarter way to build AI automation
At Unitary, these principles are foundational to how we build. Our automation platform uses AI to build the automation, not to run it. That means a one-time cost to turn a process into code, rather than an ongoing AI bill every time the workflow executes. Then we use AI only where judgement, language understanding, or reasoning are genuinely needed, matching the model to the complexity of the task. For US MGA Attune, we automated one workflow that was 50% cheaper than their previous solution — outsourcing the work to a BPO. That is what keeps unit economics viable at scale.
AI’s real value is not that it can replace human work at any cost. It’s that it can remove repetitive, manual work in a way that is faster, cheaper, and more scalable. But that only happens when AI is deployed deliberately.
The companies that keep their AI costs under control will be the ones that make smart architectural choices early: use software where software is enough, use smaller models where smaller models are enough, use frontier models only where frontier models are justified, measure cost per workflow rather than model capability, and design every system around the most cost-effective, reliable way to get the job done.
AI can absolutely transform operations. But only if companies stop treating it like magic, and start treating it like infrastructure.
