Generic AI tools promise to solve every problem under the sun, from running enterprise CRM to analysing why your cat isn’t drinking enough water during a heatwave. And when you try to use one tool with no defined role for every task, results are disappointing and erode trust in technology.
Multiply that pattern across hundreds of employees, and you get tokenmaxxing: prompt after prompt, refining AI output until something usable comes out.
CEO and co-founder of StackBlitz, the parent company of Bolt.new.
Some organizations, by instinct, default to blaming the employees for tokenmaxxing, or double down on training on ‘responsible use of AI’.
Latest Videos FromTechRadar
Tokenmaxxing is a symptom of capable people having to create solutions in an environment that lacks clear shape, scope, guardrails – and ways to analyze spend and outcomes.
Locking it down is not the solution
Traditional procurement systems were designed for traditional business software cycles. If building an enterprise system takes years, it’s not a big deal if procuring the right vendor takes years, too.
Routing AI through procurement to evaluate all options and keep the spend predictable initially felt to organizations like a safe and responsible choice. But software creation is very different today than it was even 5 years ago. Standard IT procurement still takes months, while opening a public LLM chatbot and asking it to generate a landing page is a matter of five seconds.
Policies tend to live their happy lives somewhere in the corporate portals, unbothered and forgotten unless an incident calls for a formal investigation. And when the approved systems are slow or restrictive, employees find workarounds. They turn to unvetted shadow AI tools for daily tasks and put proprietary corporate data into public models.
Look at the app prototypes that come out of it: the bad code is one problem, but the interfaces are off-brand, ignore corporate design systems and WCAG 2.2 accessibility rules, and lack basic authentication patterns.
Restricting AI use for employees in reality means they pay the £18 monthly subscription out of pocket and still use it. It moves the spend somewhere finance can’t see it, at least not immediately and not until it’s too late, adding a data governance problem on top.
The other consequence is harder to put on a balance sheet. A company where only engineers can create software moves at the speed of engineering’s queue, but opening options up to a wider group – product management, analysts, marketers – means a company can speed up innovation while having greater visibility into it.
The cost of ungoverned building
The pressure is already on. Teneo’s Vision 2026 survey of more than 350 public company CEOs and 400 institutional investors found that 53% of investors expect to see a positive ROI on new AI initiatives in six months or less, while only 16% of large-cap CEOs believe they can deliver on that timeline. The same CEOs report that fewer than half of their current AI projects are ROI-positive. Investors want returns on a clock that executives don’t think they can meet.
Volume is not the problem. Goldman Sachs Research forecasts that token consumption will multiply 24 times, to 120 quadrillion tokens per month, between 2026 and 2030, while semiconductor providers are already delivering 60% to 70% annual declines in cost per token. Tokens are getting cheaper, and we are going to use far more of them.
But who – or what – uses tokens is going to change. Organizations are building and deploying AI agents who act on their own. And agentic AI burns far more tokens than good old human prompting, because the queries repeat in sequence: a chatbot request multiplied by 20, or even 50, running unattended. Send an agent to work on an unscoped problem, and it will keep going long after a human would have stopped to ask whether this was worth it.
A lot of those tokens go into prototypes that never get to see the real world. Prototyping and sifting through ideas is how software evolves, but building with no eye on production drains resources. Gartner found that at least 50% of generative AI projects were abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value. If a prototype works in isolation but has to be scrapped because engineering would need to rebuild it from scratch, that’s pure waste.
Then organizations have to pay the bill twice – for inflated tokens and lost engineering hours. When teams build with generic LLMs and no guardrails, senior engineers spend hours cleaning up the output. That deepens the trust gap in technical teams. Stack Overflow’s 2025 developer survey found that 46% of developers distrust the accuracy of AI output. “Almost right” (but not quite) AI answers are the biggest frustration for 66% of developers, while 45% report that debugging AI-generated code is more time-consuming.
Access for everyone, accountability from everyone
The speed of code generation isn’t the problem. Lowering the technology barrier for non-engineers produces better software, software shaped by the people who understand the business need. AI building is an edge, and an edge works best when everyone has it.
But alongside democratized access to technology, we need new ways to think about accountability for how you use it and the associated costs. Modern corporate culture can promote unnecessary complexity in how we think and speak about our work. Being able to explain in simple words how you spent your day is a useful exercise, and we should apply the same thinking to AI spend.
So, stated plainly: tokens should be used for a job somebody actually needed. In turn, the user needs to be able to explain, beyond the buzzwords, what the purpose was behind a particular token spend.
To work, rather than become another source of corporate frustration, this accountability needs to be part of the environment offered. One way to start is to retire the do-everything chat box as the default. Instead, organizations should start thinking about how to give people task-specific agent workflows tailored to their roles, and build execution limits into the software environment.
This way, non-technical staff get brand guidelines, fonts, security rules, and components pre-coded into the environment, so the apps they build are compliant and production-ready by default. They don’t need to file separate requests because the shape of the environment already includes the right guardrails on security and spend.
What it looks like when it works
Bounded workflows pay off in ways you can measure. In a recent mid-sized company implementation, a 3-person data analytics team built an agent that analyses data across all their systems; it saved each of them 12 to 13 hours of manual work a week.
Three analysts now produce what once took a team ten times their size, and their daily job also got way more interesting. They spend their time on high-value strategy instead of gathering data by hand.
That agent has a clearly scoped job and executes it in line with pre-defined parameters – it is not a chat box with a company credit card.
Bringing together abundance and accounting
Uncontrolled AI costs and falling trust in technology have the same solution, and it is not rationing of tokens. Tokens will be getting ever cheaper, and their abundance is the point of this technology.
The answer is in defining and understanding the scope of AI. Give everyone the tool but make it purpose-specific from the start rather than a generic catch-all.
Every company I talk to can tell me what it spent on AI last quarter, but when I ask what they wanted to get in return, the answer is murky.
This lack of understanding of what job AI should do for an organization is the problem, and the surprising bill is only its symptom. The first step is for the organization to articulate, in the plainest words possible, what the tools are for before the tokens start moving.
Store your data in the cloud with the best cloud storage.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
https://cdn.mos.cms.futurecdn.net/DVffQnnibMWmNpx2Wfb5Se-1920-80.jpg
Source link
Eric Simons
