AI adoption is not leading to a job apocalypse or saving many firms real money. Make haste slowly? A reality check

Recent comments by some Big Tech top brass on Artificial Intelligence adoption are providing a reality check moment: Is AI adoption really helping companies cut costs, bring in efficient processes, and improve bottom lines? The ‘token burn rate’ of AI tools — particularly agentic AI models like Claude — is rising faster than the return on investment (ROI) for the firms using AI. In simplest terms, tokens are the total input-output measurement of computing required to answer your queries, prompts, or chats with an AI interface. OpenAI CEO Sam Altman has admitted that the “job apocalypse” that he and many other AI advocates predicted is not happening as fast as he thought. Why is AI still not able to bring the expected rewards for companies that adopted it? Why are firms like Starbucks re-hiring humans after having adopted or integrated AI into their workflows?

AI adoption meets reality

After a couple of years of hype, enterprise AI adoption is entering a more sceptical and realistic phase. Companies that adopted AI saw an explosion in token consumption, the main criterion used by AI tools for billing. But firms like Uber are now realising that this is not translating into productivity gains, shipped features, or workforce reductions. Some are now asking the question: Is AI spending economically justified as a cost-cutting and productivity-increasing measure? Some are realising that the “AI apocalypse” may arrive much before the long-predicted jobs apocalypse, because AI stocks could fall if adoption is not in tandem with the billions of dollars spent on computing power to answer user queries.

Uber president’s warning: Astronomical token growth meets unclear business value

Uber President and COO Andrew Macdonald recently said AI investments are becoming “harder to justify”, as companies cannot yet connect rising AI usage to concrete business outcomes. Speaking on the Rapid Response podcast, Macdonald talked of a widening gap between AI-assisted coding activity and actual delivered value. “You talk to your senior engineering leaders, and you’re saying, ‘Okay, how many projects that were on the cutting room floor got moved above the line because of the productivity gains, because 25 per cent of our code commits were via Claude Code last quarter?’ That link is not there yet,” he said. “If you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify.”

Macdonald noted that token usage and AI-generated code commits are going up “in a really astronomical direction”. Yet Uber still struggles to tie those metrics to measurable gains in productivity, feature delivery, or user outcomes, he said.

Firms are burning through their AI budgets but have little to show for productivity or cost gains

Uber reportedly exhausted its entire 2026 AI coding tools budget within roughly four months because of extremely high employee adoption — around 95 per cent in some teams. This means that, when given a chance, employees would use AI for most of their tasks, driving up the costs. In Uber’s case, the AI budget costs were partially offset by reduced hiring of humans.

Macdonald’s reality check came weeks after his own company’s CEO, Dara Khosrowshahi, talked of AI as a way to dramatically increase throughput. Uber was targeting between 20 and 100 per cent productivity gains per employee while slowing down hiring. But that doesn't seem to be happening.

The basic problem with enterprise AI: Usage is easy to measure, but value gains are not

Macdonald’s remarks are an example of growing doubts across several AI-user industries about whether spending on AI is generating returns on investment fast enough. AI coding tools are usage-based systems where costs rise as usage increases. Those using AI — such as developers — consume more tokens very quickly.

AI token costing is different from traditional software licences. The costs can rise exponentially, particularly with agentic workflows and large-context usage. Companies can easily measure and track token counts, code generation volume, AI-assisted commits, or tool adoption rates. But in many cases, this is not going hand in hand with increases in shipped features, faster delivery cycles, better products, or higher revenue.

Some firms are moving from ‘AI pilot project’ hype to more disciplined use of AI: Enter the token throttling era

Several enterprises expected that AI would help cut down large engineering teams while producing equal or greater output, but they are unable to prove this in real life. The causes are many: AI integration challenges, organisational friction, and quality issues that require human testing, monitoring, and control.

Reports indicate that some of these companies are now transitioning from the exuberant “pilot and hype” phase of AI adoption to a more financially disciplined era focused on accountability. Both AI providers and companies that adopt AI are increasingly relying on “token throttling”: methods by which caps are introduced on how many tokens users or teams can consume within specific time windows.

AI providers such as Anthropic and OpenAI use throttling to manage GPU capacity and stabilise services in order to limit queues and “Too Many Requests” errors. They are trying to ensure fair access and control compute costs. Limits include tokens-per-minute caps, rolling usage windows, weekly caps, and peak-hour restrictions.

On the other hand, companies using AI tools are also implementing their own throttling systems and governance policies to prevent uncontrolled spending from heavy users or autonomous coding agents. These controls reflect how quickly costs can spiral in AI workflows, particularly agentic coding systems that repeatedly process large codebases and contexts.

How much are companies spending on AI tokens?

It depends on usage intensity, which AI models are used, and how much automation is introduced using AI. Anthropic, the firm behind Claude, estimates that average Claude Code usage costs roughly $6 per active developer per day. It said that 90 per cent of users stay below $12 daily. This translates to approximately $100–$250 per developer per month for moderate-to-heavy usage. For context, Anthropic’s enterprise plans roughly cost $200 for coding workflows of power users, around $85 for typical users, and $25 for light users per month. But all this depends on a specific company’s usage patterns. One company with 100 engineers might spend only about $2,500 monthly overall, while extreme outliers can consume $500–$1,400 worth of tokens in a single day when running agentic workflows with massive contexts.

Some startups have reportedly seen AI spending grow 15 times within months, jumping from roughly $200 per developer monthly to nearly $3,000 as adoption deepened and workflows became more AI-dependent.

This is why human governance, throttling, and budget controls are becoming standard enterprise practices.

Claude vs OpenAI: The pricing arms race

Token costs are central to enterprise budgeting for AI adoption. In flagship models, Anthropic’s Claude Opus 4.7 charges approximately $5 per million input tokens and $25 per million output tokens. Note that one million input tokens are roughly equal to 750,000 words. Output tokens are charged at a higher rate than input tokens.

OpenAI’s GPT-5.5 is priced at around $5 input and $30 output. Anthropic is therefore slightly cheaper for high-end, output-heavy workloads. Most coding and production use of AI is happening in the mid-tier market. Here, OpenAI is generally cheaper. Claude Sonnet 4.6 costs roughly $3 input and $15 output per million tokens, while OpenAI’s GPT-5.4 costs about $2.50 input and $15 output. OpenAI’s GPT-5.4 Mini is significantly cheaper at $0.75 input and $4.50 output. At the budget tier, OpenAI has a stronger cost advantage. GPT-5.4 Nano costs around $0.20 input and $1.25 output per million tokens, while Claude Haiku 4.5 costs roughly $1 input and $5 output.

But remember, you get what you pay for.

Enterprises are rethinking AI deployment versus human engineers

It is emerging that heavy AI coding usage could end up costing more than simply paying human engineers. Businesses are looking at cost-cutting methods while adopting AI. They often mix AI providers depending on the workload. Claude is popular for coding, reasoning, and long-context tasks despite higher costs, while OpenAI is frequently favoured for broader cost efficiency. Regardless of the AI provider, real-world developer costs commonly remain in the $100 to $400 monthly range for serious AI-assisted engineering work. The cost differential is leading some companies to mix and match AI and humans.

Microsoft, which gave thousands of engineers access to Anthropic’s Claude Code in late 2025, saw AI usage costs rise so sharply that it started cancelling many direct Claude licences in divisions like Experiences & Devices. The company is shifting employees towards its own cheaper tool, GitHub Copilot CLI, mainly to control token spending.

Starbucks, which rolled out its AI inventory system called “Automated Counting” across North American stores to reduce manual work and improve efficiency, had to shut it down after nine months. The AI system frequently made mistakes, often miscounting or mislabelling items like milk and syrups, forcing employees to spend extra time double-checking inventory manually. In this case, instead of saving labour, the AI tool created additional work and operational problems. Starbucks has since shifted focus back towards consistency and human staffing, including hiring more baristas.

Companies like Meta and Amazon encouraged heavy internal AI use, but are finding out now that newer agentic AI systems consume vastly more computing power than normal chatbot use. In some cases, agentic workflows can use up to 1,000 times more tokens than standard AI queries, leading to increase in AI bills faster than companies can measure real productivity gains.

A reality check from OpenAI CEO: Sam Altman admits the jobs apocalypse hasn’t arrived

It is in this context that we have to read Sam Altman’s recent comments acknowledging that AI has displaced far fewer jobs than he and other AI evangelists expected just one or two years ago. The OpenAI CEO said at the Commonwealth Bank of Australia conference that he was “delighted to be wrong” about the pace of white-collar job elimination. “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened.”

He said OpenAI had been “roughly right” about technological progress, but “pretty wrong” about the social and economic implications. “I don’t think we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about.”

Why AI adoption isn’t translating into mass job cuts

Altman thinks the slower-than-expected displacement is due to the enduring importance of human interaction, organisational inertia, and the complexities of AI integration into workflows. Acknowledging the persistent “human part” of work, he said, “We really do care about our interactions with people.” People still value trust, judgement, relationships, and authentic communication, he noted, while sharing how his own attempts to outsource his email and Slack communications to AI made him realise that meaningful interactions could not simply be automated away.

Why enterprises will move only slowly on AI adoption

Even when AI systems are technically far more capable, organisations face institutional friction, risk aversion, AI integration challenges, oversight requirements, and structural complexity that delay real transformation. This leads to a paradoxical situation: AI adoption and token usage are growing extremely quickly, but businesses are struggling to convert that activity into measurable reductions in headcount, faster shipping cycles, or clear bottom-line improvements.

This is why some firms are moving from an “AI experimentation” phase to a “scrutiny and governance” phase, where spending controls, throttling, ROI measurement, and human oversight are becoming central priorities.

AI washing and re-hiring of humans: How organisational reality is outpacing AI hype

From the point of view of these companies, the economic and organisational limits of AI adoption are arriving faster than mass labour displacement. There are two things happening. One is “AI washing”, where layoffs are blamed on AI, but the real causes are elsewhere. The second is re-hiring after premature job cuts. These two aspects are creating scepticism about exaggerated AI automation narratives.

There is no doubt that AI is improving productivity in targeted workflows. But these gains often lead companies to pursue more ambitious goals rather than simply reduce staff. AI-led efficiency increases total consumption instead of lowering it, and it creates the need to increase AI budgets.

A more sober consensus appears to be emerging in at least some companies that, while AI capability is advancing rapidly, organisational absorption is much slower.

For now, the immediate crisis for many enterprises is not mass unemployment — it is runaway AI spending without clear accountability. Big Tech companies like Microsoft, Meta, Google-Alphabet, and Amazon are pumping massive capital expenditure into AI infrastructure, estimated to total up to $700 billion. But investors and executives are seeking clearer monetisation and productivity links.

The AI apocalypse, financially speaking, may arrive before the jobs apocalypse if and when AI provider companies fail to scale adoption in tandem with the computing costs involved in answering queries.