AI uptake amongst businesses has moved beyond ambition to reality, and fast. At the start of 2025, 52 percent of UK organisations had adopted AI. Today, that figure has risen to 64 percent – the equivalent of one new business every 40 seconds. In fact, one study found that almost all (94 percent) IT decision-makers consider AI to be an integral part of their corporate strategy.
At face value, this increase in uptake makes sense. AI promises organisations a wealth of opportunities with the ability to do more with less, faster and more easily than before.
Yet, in the rush to embrace AI, many organisations have paid little attention to its wider implications and, crucially, how its use should be controlled. As a result, they are blindly adding AI tools without fully understanding the risks, potentially increasing their exposure while quietly draining IT budgets.
The complexities of AI models
While AI may be easy to deploy, beneath the surface lies a web of complexity that is hard for most organisations to understand, let alone manage. Left unchecked, these can impact the potential value that can be derived from AI. Some of these key pitfalls are as follows:
1. Overlapping services and the rise of shadow AI
Part of the complexity surrounding AI has arisen from the way in which it has been rolled out within organisations. Rather than implementing an overarching organisational strategy or indeed training, many organisations have seen individual departments operating independently and deploying a variety of tools and applications that best meet their individual needs.
On the other side, there are organisations that are so focused on AI that they’ve rolled out a multitude of platforms to best help their employees. This, however, risks creating a library of different platforms, all of which overlap and do similar things to each other – creating additional expense where it is not required and requiring additional management around each tool.
2. The use of tokens
Services, such as ChatGPT, are typically accessed by consumers for free or through a monthly subscription service. However, at an enterprise level, organisational use of AI can be measured through “tokens”.
At their most simplistic level, tokens are units of data or text that AI models process. Companies pay per token used – which is usually a word or a part of a word. Using a Large Language Model (LLM) to answer simple, everyday questions will only use relatively few tokens, however, asking an AI platform to develop long-form content or to execute a complex task will use considerably more.
While this appears relatively straightforward so far, not all text and data is created, or importantly charged, equally.
Take the word “appearance”. In one LLM, if this word is used in a user’s input query, it will be classed as one token. However, if the user inputted the word “happen”, despite using less letters, it would be classed as two tokens. The use of a capital letter can add extra tokens again, while mis-spellings also lead to more tokens and higher bills.
This lack of predictability around token use only makes it more problematic for organisations looking to foresee and manage budgets around AI.
3. AI is prone to hallucinations
AI platforms present users with the ability to process complex queries quickly. However, they are prone to error or “hallucinations”. In fact, a study by Open AI found that GPT-3 provided inaccurate information in approximately 15 percent of cases.
This is problematic from a number of perspectives. Not only are organisations spending money and time on inaccurate information, but more worryingly, that misinformation could result in poor decision making, something that could easily lead to costly organisational mistakes and even reputational damage.
Indeed, there have been several instances of AI hallucinations making their way into legal arguments in court proceedings. One recent example involved claims that the court had been misled by the use of references, which included hallucinated case law, and the use of AI to provide a letter of explanation to the court, which again misapplied the law.
The need for user training and corporate governance to prevent, or at least identify, such errors must be a key part of AI adoption.
4. Licensing complexity and the lack of predictability around costs
When it comes to the licensing models surrounding AI, the complexity only intensifies. Each AI platform licenses differently. LLMs such as Open AI in Azure and Amazon Bedrock in AWS, for example, charge users per token, while user-based SaaS AI such as Microsoft’s Copilot and Anthropic’s Claude charge a monthly per-user subscription fee.
4.1 Seat + Consumption billing
More and more, AI vendors are moving towards a “seat + consumption” billing model – where the monthly license only covers part of the usage and additional work is charged separately.
Microsoft recently announced this price model for Copilot Cowork where a 60-user company could easily spend over $100,000 a year on the additional AI usage and 30,000 users could see a bill over almost $5,000,000 per month.
Corporate governance, usage policies, user training, and a focus on ROI become even more important here.
4.2 Tokenomics
Within LLM models and tokens, there are different pricing structures based on which model is used – for example, ChatGPT 5.2 has a cheaper input price than ChatGPT 4.0 – meaning its cheaper for you to ask it a question – however, it is more expensive for it to give you an answer. Which model is cheaper overall will depend on how it is used – large prompts and short answers or vice versa. Behaviours such as caching and batching must also be considered when assessing AI costs.
With unclear pricing and a complex token structure, it is easy to see how costs escalate. Uber reportedly burned through its entire annual AI budget in less than four months after staff rushed to adopt Claude Code. Meanwhile, Peter Steinberger, whose AI tool OpenClaw was sold to OpenAI, reportedly spent $1.3m (£1m) in a single month on tokens – costs which he passed on to his employer.
With many organisations unaware of the need to set guardrails around AI usage, costs can accelerate rapidly if left unchecked. And with organisations often lacking a dedicated team to manage their AI portfolio, it’s an issue that is largely falling through the cracks.
Increased exposure to risk
If left unmanaged, AI can significantly increase an organisation’s exposure to risk. Something that the rise of shadow AI is only exacerbating further.
A recent US study by Thomson Reuters revealed that “a third of lawyers, accountants and compliance professionals are using AI their organisation has not approved, rising to 41 percent among those who say their organisation is moving too slowly on AI.” Of those surveyed, while 96 percent say that their AI must safeguard confidential data, 41 percent lack access to professional-grade tools that meet those standards.
Shadow AI risks exposing any sensitive company or client information to a third party, leaving organisations vulnerable to data breaches and cyber-attacks, and falling foul of compliance regulations.
Establishing guardrails around AI
Organisations looking to see the most value from AI need to put in place effective parameters to govern its use. For organisations today these include the establishment of the following:
- Workplace AI policy and with assigned ownership
It is essential that all organisations have a corporate AI policy in place that details expectations for AI use and user conduct, including a requirement to use only approved corporate tools. This should also outline parameters for approved use cases and users so that all parties understand from the outset whether they are eligible to use it.
Secondly, a department must be assigned to oversee the execution of the AI policy. AI adoption has been on such a rapid growth trajectory that its management has largely slipped through the net – with teams across IT Asset Management (ITAM), FinOps, security, and IT, unaware of where responsibility lies – particularly for cost.
Responsibility could legitimately sit across any of these departments, or even warrant its own dedicated team, however it is essential that there is a clear body responsible for AI’s management. If internal expertise is lacking, external and specialist consultants can be brought in to help. Wherever the responsibility falls, it is essential that it drives consistency across the organisation with the tools used and ensures that approved users are trained in how to get the most out of their AI systems used.
- Targeted use cases
Organisations need to be more selective in their application of AI if they are to realise its true business value. Rather than merely experimenting with the technology, they should query every use case, asking “Will deploying AI here offer better value than a human?” In some cases, that answer will be no.
For Uber, in the aforementioned example, its Chief Operating Officer later acknowledged the challenge of proving a return on investment on its AI strategy after blowing the budget in a matter of months. This is a problem faced by many.
A study of 951 global companies from Bain & Company found that while 37 percent of those surveyed targeted cost reductions of 11 to 20 percent from using AI, almost 40 percent of those who measured outcomes, landed in the 0 to 10 percent bucket instead. While the technology might work, the value doesn’t always arrive.
AI can make a meaningful difference to an organisation – but that doesn’t mean that it should be used in every circumstance. Ultimately, organisations must establish clearer and more defined user cases and success metrics, so that spending on AI doesn’t outpace the potential value it is meant to deliver.
- User profiling
Likewise, not every person in an organisation will benefit from, or needs, blanket access to AI. Its use should be limited to those departments and individuals that will truly see a valuable return on its use and those who have a viable use case.
Ensuring that individuals establish clear outcomes and KPIs for their AI projects from the outset will help here and ensure that AI’s use is targeted to delivering the best value for an organisation.
- AI training programmes
An AI training programme will fix this at the point where the money actually gets spent, not after the invoice arrives. That means giving staff a working sense of how consumption-based pricing behaves, when a cheaper model is genuinely fine for the task, and what internal approval routes exist before someone commits the business to a new tool.
The finance team needs a different layer of training again: how to read usage data, what a normal month looks like, and which numbers should prompt a question rather than a shrug.
Done properly, this training is ongoing and sits alongside the governance framework and the spend visibility work. This means the people making day-to-day decisions understand the guardrails they’re operating inside, and finance can trust that the numbers they’re seeing reflect informed choices rather than guesswork.
Embracing the value, avoiding the pitfalls
With complex licensing models, confusing pricing structures, and a workforce eager to embrace AI, it’s easy to see how AI budgets can spiral out of control, without any consideration of demonstrating a return on investment.
To truly unlock AI’s potential, organisations need to move past the curious, experimentation phase and move towards a more considered approach to AI adoption. That means defining approved use cases, aligning them to clearly profiled users and implementing clear governance frameworks that control and create accountability around AI spend.
Crucially, organisations must also question how they evaluate value – moving away from asking what can AI do, towards, what value will AI bring.
Those that can strike an effective balance here will deliver meaningful impact and a true return on their investment. Those that don’t, may find their IT budget unknowingly drained with little value to show for it.
