Work & worth

Am I only worth 2.2 million AI tokens?

Treating people’s worth as a token cost is a serious business mistake. The comparison misses human capital and the long-term value people create.

By Adrian Dunkley · Monday 2 November 2026 · Opinion

A cream paper fingerprint stands above small blue and yellow tokens on a dark tabletop. — Beyond the Meter · AI-generated artwork by AI Boss Ltd

A business that treats a person’s worth as a quantity of AI tokens is making a serious mistake. The price of generating text cannot tell you what an employee contributes, whether their role can be replaced, or what the organisation stands to lose if they leave. Using that comparison to cut staff gives an incomplete cost estimate authority it has not earned.

Companies should examine where AI can improve work and reduce costs. They should also demand evidence before treating a tool’s output as equivalent to the human capability behind a service.

How the 2.2 million AI tokens were estimated

The estimate counts selected text outputs, their input material and drafting. Its processing cost depends on which models handle which tasks. Read How I calculated 2.2 million AI tokens for the assumptions, task-by-task dollar costs and subscription comparison, or change the estimate in the calculator.

What an AI token measures

For text models, an AI token is a unit the system processes: a character, part of a word, a whole word or punctuation. Input tokens represent material supplied to the model; output tokens represent what it generates. Some models also bill for internal reasoning. Counts vary with the model, encoding and language, as OpenAI’s explanation describes.

Account credits belong to the payment arrangement. They may be a prepaid monetary balance or a product-specific allowance. With OpenAI’s prepaid API billing, usage draws down purchased credit at the applicable rates. There is no universal exchange rate of one credit to one token.

Compare the ability, output and reasoning behind the work

Years of relevant experience help identify the difficult cases and judgement expected of the person whose work you are comparing with AI. Years alone do not establish competence, and multiplying a token allowance by tenure would invent a relationship rather than measure one.

Ability comes first: can the tool complete the task with the information, permissions and tools available? For completed cases, assess whether the output meets the same requirements for accuracy, coverage and usefulness as that person’s work. Similar length or polished formatting is insufficient.

Reasoning equivalence needs a defined work standard too. Check whether the approach respects the evidence and constraints, handles exceptions, tests alternatives and recognises when to defer. This is a comparison of observable decision quality, not proof that a model thinks like a person. Anthropic’s research on reasoning explanations shows why a model’s account of its reasoning cannot simply be taken as a faithful record of its internal process. A larger reasoning-token bill does not resolve that question.

The updated calculator records relevant years, task ability and observed output and reasoning checks for the same test cases. It uses the cases that passed both checks together, rather than multiplying separate pass percentages. Without a record comparing the same cases under agreed conditions, the cost of producing work to that standard remains unestablished. The 2.2 million figure remains the reference text workload.

Past work does not settle future replacement

A person’s accumulated output over their working years is a different quantity from this monthly estimate. Counting an archive can describe the text someone produced. It cannot count everything they learnt, the mistakes they prevented or how they will respond to a situation they have never encountered. Relevant years help define the comparison standard; multiplying one assumed month by a whole career would require evidence that the work volumes and conditions were comparable.

Fine-tuning adapts a model using selected training examples, as Google’s explanation of supervised fine-tuning describes. Learning patterns from someone’s past documents does not establish that the resulting system can take over their future responsibilities. That requires separate tests on unseen work, including exceptions, changing circumstances and the judgement needed to challenge a request. Training costs, future operating costs and a person’s long-term contribution answer different questions.

A generated document is only part of the job

Consider a board paper. A model may draft its paragraphs quickly. The business still needs to know whether the evidence is sufficient, the assumptions fit the company and the recommendation survives scrutiny. Someone must understand the consequences and accept responsibility for the advice. Counting the words in the finished paper leaves much of that work out. Nor does completing several isolated tasks establish that a tool can handle the combined role through exceptions and changing demands.

The same problem appears when an organisation converts a salary into a token allowance. A change in the provider’s price changes the number of tokens that budget buys without changing the employee’s expertise. A shorter, better report can require fewer tokens while becoming more useful.

NIST’s AI Risk Management Framework warns that mathematical representations of complex human phenomena can remove necessary context. Applying a text-processing price to an entire role makes precisely that omission.

Human capital has value beyond today’s output

By human capital, I mean the knowledge, skills and judgement people develop and apply. Some of its value appears in completed work. Some appears when an experienced colleague spots an unusual case, helps someone learn or remembers why an earlier approach failed.

If a company removes experienced staff because AI can reproduce their documents, it should ask who will assess unfamiliar situations and teach the next cohort. If it removes junior work, it should ask how people will gain the experience needed to take responsibility later.

Where AI can meet the required service standard at a lower total cost, a company has a reason to consider automation. But the investment case needs to account for the service it must sustain, the knowledge it needs to retain and the capabilities it wants to build.

Amazon’s announcements do not establish that AI can do the work

In his June 2025 message to employees, Andy Jassy said he expected AI-driven efficiencies to reduce Amazon’s total corporate workforce over the following few years. The company’s January 2026 announcement of approximately 16,000 role reductions also described reducing layers, increasing ownership and removing bureaucracy.

Management’s stated reasoning does not show which duties a model can reliably perform. A layoff announcement, a forecast of capability and a tested replacement process provide different kinds of evidence. Some work can be automated; a company still needs to demonstrate the case for the particular change it proposes.

Measure what the business gains with its people

Could employees use AI to examine more evidence, respond more quickly or spend more time with customers? Does the improvement survive checking and correction?

Brynjolfsson, Li and Raymond’s study of 5,172 customer-support agents found that AI assistance increased issues resolved per hour by 15% on average, with substantial differences between workers. A separate METR experiment with early-2025 tools found that experienced open-source developers took longer on their repository tasks when AI was allowed. Neither result supplies a universal replacement rate.

Choose a service or output and set its acceptance standard before testing a change. Measure the time and cost of reaching that standard, including supervision, errors and rework. Examine what happens to customer outcomes and staff learning. Then decide how any released time will be used.

Sources and further reading

Sources checked 1 November 2026. Prepared with AI assistance and editorial review.