AI experts have warned that many UK businesses will fail because they are investing heavily in, and leveraging themselves on, AI agents “whose future pricing and operating costs they barely understand”, as current AI agent usage is being massively subsidised by venture capital (VC) investment.
Each time a business uses an AI agent, whether Claude or ChatGPT, it is burning through significant numbers of ‘tokens’, the true cost of which — ‘tokenomics’ — is being concealed by billions of pounds of VC money.
A token is the fundamental computational resource that powers AI agents, specifically the massive GPU (graphics processing unit) and electricity costs required to process inputs and generate outputs.
Each word inputted, each punctuation mark and even the space between words is a token, and that token costs money, as it involves energy and compute.
So a prompt or series of prompts, a reply, an uploaded file, a generated image or video, dataset analysis or automated flow, is burning huge volumes of tokens that — for now at least — aren’t being paid for by most businesses.
Tokenomics
Colette Mason, AI Ethics Consultant at London-based Clever Clogs AI, who works with companies of all sizes on their AI adoption daily, said: “In my experience, the vast majority of UK SMEs are trying to build an economy on a technology whose true costs they simply do not understand.
“Today, tokenomics, or the true cost of compute, doesn’t mean much to most SMEs but in the not-too-distant future it could be their epitaph if they don’t get to grips with it, fast. Failing to model the costs of AI correctly could see many businesses fail.
“Only businesses using pay-as-you-go (API) billing understand the real cost of AI. Flat-rate subscriptions are heavily subsidised to bring in users, effectively hiding the true cost of compute.
Mason gives an example of a 10-person firm where each member of staff summarises just one 10-page PDF daily, which would equate to roughly 220 PDFs per month.
She says the true cost of this one simple task carried out by one member of staff is roughly the same as one person’s current monthly subscription to Claude Opus 4.6 or GPT-5.5, adding that “this assumes the output is right first time and doesn’t need a rerun”.
Mason continued: “Add in 10 more tasks with reruns and it will cost a small fortune. PAYG reveals the true AI price tag. Scale this across dozens of daily workflows and the real cost of enterprise AI quickly becomes impossible to sustain.”
Overly dependent
Mitali Deypurkaystha, Human-First AI Strategist at Newcastle upon Tyne-based Impact Icon AI, said “many British businesses are behaving like they’ve been handed free samples without realising the dealer eventually comes back with the bill”.
She continued: “As an AI strategist, one of the biggest parts of my job is often explaining where businesses should not use AI. Too many owners see cheap subscriptions and assume automation belongs in every workflow whether the return on investment exists or not.
“I’ve had to push back on companies wanting to force AI into processes where the gains simply weren’t there, especially when today’s bargain pricing is heavily subsidised and may not reflect the true long-term cost.
“AI will absolutely transform the economy, but too many SMEs are rebuilding their operations around tech whose future pricing and operating costs they barely understand. The danger is that by the time those costs rise sharply, businesses may already be too dependent to walk away.”
Dependency risk
Katrina Young, Chief Technology Officer at KYC Digital, an AI consultancy, said “the operational economics of large-scale AI adoption are still being tested in real time”.
She added that many businesses are not just underestimating the true cost of AI but also the true cost of monitoring the AI they use: “Many organisations are scaling AI into workflows before governance, tracking, oversight and operational resilience are fully established.
“ROI calculations often focus on productivity gains while excluding verification labour, escalation handling, accessibility, compliance exposure and long-term dependency risk.
“The hidden risk is removing too much human verification in pursuit of speed. That is where trust failures, bias risks, errors, and operational fragility begin surfacing.”
