Press Enter: Can AI Do Work Unasked?

As AI becomes a bigger part of the workplace, it is expected to do things without explicit instruction. In this study, we measure how good AI is at noticing and doing tasks on its own.

Existing research looks at the relationship between AI and work in three ways.

1. Exposure studies such as Eloundou et al.’s “GPTs are GPTs” [1] estimate which tasks AI could do.
2. The Anthropic Economic Index, introduced by Handa et al. [2], records what people actually use AI (Claude) for.
3. Benchmarks such as GDPval, by Patwardhan et al. at OpenAI [3], grade AI on tasks it is handed.

Our focus is on unprompted AI work. An assistant watches a professional’s week unfold in real time. During the simulation, it drafts proposals based on the environment. If a draft is acceptable, the worker presses enter to accept it.

We measure how much work AI can do that is directly acceptable or acceptable with slight edits. All code and data produced for this study can be found on GitHub and Hugging Face.

Approach

We simulate one work week across eight different professions, chosen to cover the full range of observed AI use in the Anthropic Economic Index. Each week is a realistic representation of a job, consisting of meetings, tasks sourced from O*NET [4], breaks, untracked in-person work, and traps such as phishing emails or fraudulent payments.

We call the work an assistant could help with in the week planned work. Each piece is tied to an O*NET task. Planned work makes up between 26% and 59% of the worker’s labor hours. Planned work mostly corresponds to tasks done at the desk and on a computer.

The work week runs in a sandbox where work unfolds over simulated time. Email and chat messages appear when they are sent. A phone call can only be followed while it is live. At the start of an event or every 5 minutes, the model can look through anything the human worker can see, propose a piece of work, ask a question, or do nothing.

A simulated worker (Gemini 4 Argon) answers the model’s questions in character, in at most three sentences. Each question costs the worker the time to read it, at 238 words a minute, plus 30 seconds to answer, valued at their loaded wage. A question asked outside working hours goes unanswered and costs only its reading time. Gemini’s tokens are not charged to the tested model.

Our test spans 20 models, four from each lab. It includes the models each lab had in force on 1st January 2026, 30th April 2026, 31st July 2026, and the latest release at the time of writing.

The table below describes the simulations we ran.

Measuring AI’s Work

To measure AI’s work in the same unit as a person’s we introduce a new primitive, finished hours.

For each piece of planned work, we developed an estimate of how long a human worker would need to complete it. When a model’s output for that task is good enough to accept unchanged, we want to credit the model with that time saving.

To do this, our grading system scores the probability of an AI proposal being accepted unchanged. The tested model is credited with that share of time. Thus, a finished hour is a labor hour that AI takes over.

Each proposal is scored by five AI judges from a panel of six, leaving out the judge from the tested model’s own lab. We also graded 92 proposals by hand and fit the panel’s score to the hand grade (Calibration).

For example, below is Claude Opus 5.5’s finished hours in the claims adjuster scenario:

Over the course of the week, Claude Opus 5.5 finished 6.2 of the 23 planned hours.

Using finished hours, we build four measures to understand how much work AI does, what it is worth, what it costs, and whether it is safe.

AI Can Do 20% of Work Unasked, Up from 7% in January

With the latest models, 20% of a worker’s planned hours qualify as finished hours. In other words, AI can claim 20% of the time a worker in our 8 professions spends working at their desks (90% confidence interval [13%, 27%]).

In January 2026, this metric was 7%. In just 10 months, it has nearly tripled.

The five-lab average climbed at every date, to 12% in April and 17% in July. Comparing within each lab, the press-enter rate rises by about 16 percentage points a year. Leaving out any one lab keeps that between 14 and 18 points, so no single lab drives the trend. Every model ran each week once, so a gap of less than about 5 points between two models is within noise.

Latest Models Do More Work for Less

Inference cost per finished hour tells us how much AI costs for each hour of work it completes. The human equivalent of this is the cost of a labor hour, which averages $58 across the 8 selected professions.

In January 2026, a finished hour cost $68 in inference. With the latest models, it costs $7.

In 10 months, AI went from costing 18% more than a labor hour to costing 87% less.

Two things drove the fall. First, newer models notice more pending work and thus finish more tasks. On average, the latest models record 4.3 finished hours a week against 1.6 finished hours by the January models. The section The Gain Is in Noticing discusses this in greater detail.

Second, newer models spend less on inference per week. With the latest models, the average cost of inference for the whole work week was $31. For January models, it was $105, 3.4x more. Together, a finished hour has gotten about 9x cheaper.

Within each lab, the inference cost of a finished hour halves about every 3 months. A notable result here is DeepSeek V4.1 Flash 1017, which delivers a finished hour for $0.33.

AI Help Is Worth $432 a Week and Rising

To calculate the net value of AI help, we consider the following metrics:

• Time saved. For each piece of planned work, the minutes a person needs times the chance the worker accepts the proposal, either unchanged or after a small edit.
• Time the help costs. We assume a worker reads every AI proposal at 238 WPM plus a few seconds of attention. Drafts that need a small edit also cost the time to fix them. Proposals that do no planned work still cost their reading time, and so does answering the model’s questions.
• Inference cost. What the model costs to run at list prices.

We convert the first two metrics into dollars by valuing the minutes at the worker’s loaded wage, as reported by the BLS [5].

Because time saved counts drafts that need a small edit, it covers more than the finished hours. For the latest models, 49.5% of planned hours are usable, against the 19.6% accepted unchanged.

With the latest models, AI help is worth $432 per week. In January it was worth -$10.

Efficiency is the key driver here. DeepSeek V3.2 was the most eager model in our tests, making 557 to 1,143 proposals a week. However, most of them were not worth accepting. Consequently, it lost $485 a week.

Claude Opus 5 lost $259 a week because its drafts were long. A worker would spend 9.8 hours a week just reading them. Its successor, Claude Opus 5.5, takes only 2.1 hours a week to read. This results in an average net value of $476 a week.

AI Value Follows Task Size

Jobs that assign fewer but larger tasks see more gain from AI help. The latest models are worth $935 a week to software developers and $906 to market research analysts. On the other hand, the net value of AI for customer service reps is -$59.

When tasks are small, like a two-line reply to a customer, reading a draft takes almost as long as writing one, resulting in negative net value.

Today’s inference is too slow for LLM assistance in live calls. In the customer service simulation, 30% of responses from the latest models arrived too late. This gap is likely present due to our usage of standard LLM APIs. Dedicated voice agents are designed for use in live calls. We expect such tools won’t run into these issues.

The Gain Is in Noticing

The press-enter rate can rise for two reasons. A model may notice more of the work that needs doing, or it may do the work it notices better.

To tell the two apart, we split model actions into the following categories:

1. Proposals a worker would accept unchanged
2. Proposals a worker would accept with light changes
3. Proposals a worker would reject

Most of the gain comes from newer models noticing and tracking their environment better.

Models from January 2026 attempted only a quarter of the possible tasks they could take on (time-weighted). Meanwhile, the latest models attempt more than half.

The quality of attempts also increased, although much more modestly. Of the work the models attempted, the share a worker would accept unchanged rose from 28% to 36%, while the amount attempted doubled.

Measured Inputs Exceed Acemoglu’s Assumptions

Daron Acemoglu’s “The Simple Macroeconomics of AI” [6] estimates AI’s effect on productivity from two numbers.

1. AI affects 4.6% of US tasks. That is the 19.9% of tasks exposed to AI, weighted by wage bill (Eloundou et al. [1]), times the 23% of those that are cost-effective to automate (Svanberg et al. [7]).
2. AI saves 27% of the labor cost of each.

Together, they imply a 1.24% saving in labor costs. He expects total factor productivity to rise by no more than 0.66% over 10 years.

Our 8 jobs are not a sample of the economy. We chose them from occupations whose work runs through email, chat, calls and documents, which are more exposed to AI than most. So we also apply his method to the jobs we study. By his rule, a task is exposed when Eloundou et al. judge that AI can do more than half of its activities. That makes 39.3% of these jobs’ tasks exposed, twice the economy’s 19.9%. With his 23% and 27%, AI takes on 9.0% of their work and saves 2.4% of their labor cost.

In our experiments, we measure both numbers directly. For the latest models:

• Usable work is 20.4% of all labor hours (90% interval [16.4%, 23.4%]). For each piece of planned work, we take its minutes times the chance it is accepted unchanged or after a small edit. We sum those and divide by every hour at work in the simulated week, 45 on average and between 42.5 and 48. Finished hours count unchanged work only, so they come to 7.7% of the same hours.
• Net value is 12.4% of weekly labor cost ([9.0%, 15.0%]). It is the net value above divided by the week’s labor cost.
• That is a 61% saving on the hours AI touches ([54.6%, 65.3%]), or 12.4% divided by 20.4%.

All three are averaged across the 8 professions, weighted by each job’s US labor cost.

In the same 8 jobs, that is 5x the 2.4% Acemoglu’s method implies (90% interval: 3.7–6.2x). Against his economy-wide 1.24% it would be 10x, but a factor of 2 in that comes from the jobs we chose.

The rest splits evenly. AI takes on 2.3 times as much work as his method allows, 20.4% against 9.0%, and saves 2.3 times as much on it, 61% against 27%. The first part comes from his 23%, Svanberg et al.’s estimate of which tasks are cheap enough to automate, which they measured on computer-vision tasks. Our net value already subtracts inference cost, and it is positive in 7 of the 8 jobs. Without the 23% filter, his inputs give these jobs a 10.6% saving, 39.3% times 27%, close to our 12.4%.

Note that there are some key differences between the methodology behind the two estimates. His 27% of labor cost savings are derived from people working faster with AI [8], [9], while we consider work AI does itself. Eloundou et al.’s share counts each task in a job equally, while ours weights work by its hours. Most importantly, his estimate is about adoption over 10 years. Ours measures capability in a simulation where every worker has an ambient assistant available 24/7.

Capability Outpaces Usage

The Anthropic Economic Index records which O*NET tasks people use Claude for. Since we source our planned work from O*NET as well, we can compare usage from the Economic Index against capability recorded by our study.

Of the 137 tasks in our weeks, 98 have no recorded AI use. They make up nearly two-thirds of the planned hours.

Capability-wise, there is no difference between tasks with recorded AI use and tasks without. Both show a press-enter rate of ~20%.

Whole jobs show the same pattern as well. By Massenkoff and McCrory’s observed exposure measure [10], 8% of claims adjusters’ work is done with AI, versus 70% of customer service representatives’ work. In our study, both jobs reach about the same press-enter rate, 18% versus 17%. Capability is spread more evenly than use.

Despite Doing More Work, AI Is Now Less Harmful

Our simulation features traps like phishing emails or fraudulent payments. The goal is to measure whether relying on ambient AI increases the risk of harmful actions.

As models took on more work, they set off fewer traps. The five-lab average fell from 5.1% of the week’s traps in January to 0.85% for the latest models, or from 1.57 harmful proposals a week to 0.25. Over the same period the press-enter rate rose from 7% to 20%. The latest OpenAI and Anthropic models set off no traps.

Conclusion

Across the 8 selected professions, the latest models can take over 20% of a worker’s planned computer work while delivering a positive economic value and reducing safety concerns. This is a far cry from the start of the year, when the average January model was worth nothing, and two of the five lost money.

If this pace of improvement holds, AI’s share of work will quickly exceed current estimates. The consequences this will have for the value of human labor are outside the scope of this post but important to think about next.

@note{nigam2026pressenter,
  title  = {Press Enter: Can AI Do Work Unasked?},
  author = {Nigam, Shreyash and Feyn Research},
  year   = {2026},
  venue  = {Feyn Field Notes}
}