Four in five call it dependence
A Coddy Tech survey asked 305 developers where their AI tools sit in the working day. Of those, 80 per cent said the use has felt more like dependence than an advantage. Some 43 per cent keep coding after hours even when they meant to stop, 32 per cent have put off sleep to keep going, and 39 per cent say the tools make it harder to switch off from work.[1]
The same group answered two further questions. Heavy AI use made a raise or promotion more likely for 74 per cent of them, and made burnout more likely for 51 per cent. Those answers describe the same behaviour, and nothing in the Coddy Tech figures shows an employer putting a price on the second reading.[1]
What makes an agent reliable is procedure someone has to write
In a study run across 8,135 controlled trials, researchers from Princeton and UC San Diego and other institutions separated out where an agent's gain from skills actually comes from. Procedural grounding accounts for 65.7 per cent of it: setup steps, tool ordering, intermediate checks. Direct knowledge transfer accounts for 4.5 per cent. In 10 per cent of cases the agent applied a playbook that did not fit the task.[3]
That places the reliability work on a person. Someone writes the procedure, someone keeps it current, someone decides which one applies. Precision falls from 29.6 per cent with 5 skills to 3.3 per cent with 100, so the curation labour grows as the library grows and shows up on no token bill. A better retrieval layer could absorb that growth without adding human labour, and the same study leaves the possibility open.[3]
The instruments an employer reaches for measure tokens
Maxio chief executive Branden Jenkins found that a weekend session with a coding tool had charged 1,000 dollars to a wallet that tops itself up in steps of 1,000 dollars. He called the bill annoying and put his real worry elsewhere: employees anxious about falling behind the tools. The company's answer was a cost-improvement programme and charts that show employees alongside the AI agents they manage. Gartner estimates agentic token use at 5 to 30 times a standard chatbot exchange, and the finding that 68 per cent of US companies overrun their AI budgets comes from a WitnessAI survey.[2]
This column argued on 21 August that oversight duty raises a worker's power only when it arrives with real authority and allocated time. Today's figures test the second half of that condition. Nothing in the Coddy Tech survey or in Maxio's account measures hours: the survey shows people working past the point they meant to stop, while the company's response measures spend and places agents on a chart. Authority and budget are being written down. Time is not.[1], [2], [4]
There is a measurable version of this gap. When a comparable developer survey next repeats these questions, the number to watch is the 43 per cent who keep going after hours. If that share stays at the same level while the 74 per cent who expect a raise also stays, the extra hours will have been absorbed as capacity nobody priced. A smaller share would show that an employer has begun writing review time into the work schedule, and that is the outcome worth looking for before the middle of 2027.[1]