Will AI steal our jobs? Gartner's Token Paradox

Posted by

   

If you hang around in the world of software development, over the last two years you’ve surely heard the usual apocalyptic prophecy: "Within a few years, programmers will be obsolete, replaced by AI agents."

Well, put down the popcorn and get ready for a plot twist that no one saw coming, but that the numbers are already confirming. According to the latest Gartner report (released at the end of June 2026), the real problem won't be developer unemployment, but the bill.

It’s a prediction that will make CFOs around the world jump out of their chairs: By 2028, AI computation costs for writing code will exceed the average salary of a human programmer.

Yes, you read that right. Having an AI do your programming could cost the company more than paying you. How is this possible? The answer lies in a magic little word that is pushing corporate budgets into the red: Tokens.

🔗 Do you like Techelopment? Check out the website for all the details!

The problem no one calculated: The hunger for Tokens

In the beginning, it was all wonderful. We tried GitHub Copilot, Cursor, or various software agents while paying a flat subscription (the classic $20 or $30 per month "per seat" license). A massive bargain for companies.

However, LLM providers are changing the rules of the game, switching to a pay-per-use model based on token consumption (the unit of measure for text processed by the AI). And this is where the trouble begins.

As Nitish Tyagi, an analyst at Gartner, explains, we programmers have a "flaw" when we use AI: we optimize for speed and convenience, not for cost-efficiency.

Think about it:

  • We copy and paste entire log files or mega-classes into the prompt ("bloated context windows").
  • We let AI agents run infinite loops of trial and error to fix a bug.
  • We cyclically ask to refactor code without thinking about how much data we are exchanging with the servers of OpenAI, Anthropic, or Google.

The result? The transition from simple "writing assistants" to autonomous agents (that act on the code by themselves) is causing token consumption to skyrocket. Budgets allocated by companies for the fiscal year are being burned through in the first few months.


Why are costs taking off?

Gartner identifies three main problems in how we are using AI in engineering teams:

  1. Autonomy without control: We hand over entire repositories to AI agents without placing limits on the attempts they can make to resolve a ticket.
  2. Giant contexts: Instead of isolating the portion of code in question, we feed the AI enormous contexts. The larger the context, the more tokens you consume with every single interaction.
  3. Lack of vendor transparency: Many commercial tools do not clearly show how many tokens you are consuming in real-time or how the bill is calculated. It’s like driving a car without a speedometer and only finding out the cost of the fuel at the end of the month.

In addition, the hardware infrastructure to run these models costs billions, and vendors are raising prices to recover their investments.


How to survive the "Token-geddon"? (Gartner's recipe)

To avoid AI becoming an unsustainable luxury, the way we develop software will have to change radically, introducing what we can define as a true "token discipline." Gartner suggests 5 guidelines for lead developers and managers:

  • Define autonomy levels: Don't leave complex tasks 100% to the AI. Tasks need to be categorized into three modes: Human Only, Human + Agent, or Agent Only (only for repetitive, low-cost tasks).
  • Intelligent Model Routing: You don't need to use GPT-5 or Claude 3.5 Sonnet to write a regex or a simple validation function. Simple tasks should be delegated to small, local, or open-source models (SLMs - Small Language Models), reserving "Frontier" models only for complex architecture.
  • Context Engineering: We must learn to summarize prompts. Providing the AI with only what is strictly necessary will eliminate superfluous tokens without losing quality.
  • Automated spending thresholds: Automated blocks and alerts if an AI agent exceeds a certain token cap on a single branch or ticket.
  • Token review in Sprint reports: Just as we analyze code performance or bugs during retrospectives, we will have to analyze which workflows consumed the most tokens to understand where to optimize.

Considerations

This scenario opens up a huge debate. If AI ends up costing more than a developer's salary, will we see a return to the past? Will companies cut back on AI usage, or will they prefer to pay for computation rather than the benefits and management of human resources?

The truth is that knowing how to code will no longer be enough: the new fundamental skill will be knowing how to manage AI computational resources efficiently.


References



Follow me #techelopment

Official site: www.techelopment.it
facebook: Techelopment
instagram: @techelopment
X: techelopment
Bluesky: @techelopment
telegram: @techelopment_channel
whatsapp: Techelopment
youtube: @techelopment