As generative AI becomes widely embedded across business operations, we’re seeing the conversation change from what AI is capable of to what it actually costs to maintain long-term.
According to a recent report identifying barriers to GenAI implementation, among respondents whose organisations had fully integrated GenAI into regular processes, 41% cited prohibitive large language model (LLM) costs as an implementation barrier and 45% reported below-expected return on investment.
This highlights a clear gap between expectations and reality, with many organisations still struggling to translate AI adoption into measurable business value. This isn’t just a line item anymore. Markets took notice too, wiping hundreds of billions off tech valuations in a single session this summer amid fears that AI spending has outrun its returns.
The organisations we work with that have moved GenAI into production aren’t just asking us whether it works anymore. They’re asking how they can ensure they have control over costs.
The economics of GenAI are changing. While the cost of individual models is falling, businesses are spending more overall as GenAI becomes embedded across their operations. As organisations move beyond pilots and increase usage across workflows, the sheer scale of consumption can outweigh falling model costs, creating a new challenge: controlling the total cost of GenAI at scale.
The technology isn’t the main problem anymore. Deployment discipline is. Many organisations underestimate the cost implications of scaling GenAI across their operations and lack the right strategies to manage consumption effectively.
As a result, businesses across industries, including financial services, insurance and public sector often lack a clear strategy for scaling GenAI economically and sustainably. As these costs continue to mount, we’re increasingly speaking with customers about the true cost of this lack of preparedness and how organisations can address it before GenAI budgets spiral out of control.
The hidden cost of financial unpreparedness
We frequently see companies waste thousands of pounds on AI inefficiencies – from unmonitored redundant models to overly broad context windows running in the background. It’s possible for these specific deficiencies to go unnoticed for months, by which time they may have already made a significant dent in a company’s finances.
Now, more than ever, as agentic workloads increase alongside heavy compute demands, these pressure points will intensify. An unmonitored agent doesn’t know when to stop. Left ungoverned, it can burn through millions of tokens overnight without producing a shred of business value.
It’s a particularly common problem in industries like financial services where high-frequency queries are standard practice. Without budgeting and execution limits, unmonitored API calls can quickly escalate, forcing pilot projects to be cancelled due to a lack of budget.
Additionally, deploying AI fast without correct governance can lead to severe operational and technical risks. Moving at speed to keep up with the market or meet arbitrary deadlines can contribute to data breaches or compliance problems. A business that can’t control its AI spend is usually a business that can’t fully understand its AI risk either – the two rarely travel alone.
Typically, these vulnerabilities can manifest as shadow compute in which independent business units deploy unmonitored APIs without oversight. Businesses that bypass IT procurement to solve these bottlenecks create tools that are fragmented and not optimised. It becomes a challenge to audit risk and evaluate cost attribution because of this lack of visibility.
Scaling GenAI sustainably
Creating a sustainable future for GenAI does not mean stopping innovation or slowing down technical experimentation. It means governance first, cost control second, then scale. Innovation needs measurable and defined parameters.
The first element to successful GenAI implementation is establishing spending limits and cost controls. Even though most companies issue their employees with cards that have spending limits, many give their staff access to unmonitored and high-cost LLMs. Another aspect of setting clear guardrails is making engineers aware of problems with software well before they have damaging effects.
Using intelligent model routing is another part of efficient adoption of GenAI across software architecture. Not every task needs the smartest model in the room – knowing when to use a smaller one is where the savings live. Most importantly, using these alternatives should not compromise the performance of these systems.
Financial operations should be directly embedded into a company’s culture. Each department using GenAI on a widespread scale needs visibility and awareness of its individual daily spend. Not only does this allow teams to evaluate business value against consumption, it also corresponds with direct accountability across corporate teams.
The future
It’s no longer sustainable to prioritise unconstrained GenAI experimentation and pilot projects. Moving forward, success means placing the same level of importance on implementation alongside financial predictability and technical capability.
Organisations that use this software long-term must prioritise solid financial operations, model routing and governance in the running of their companies.
Businesses that build this discipline now won’t just control costs, they’ll be the ones still scaling when everyone else is still figuring out their overspend.
Encouraging this structural stability within companies is essential for growth. Ultimately, governance isn’t what slows AI down. It’s what makes it affordable enough to keep running. With the further integration of agentic workflows, financial guardrails will help protect businesses’ money.
