When to hand an agent a whole job

For a couple of years the transaction had one shape. You asked, it answered, and you decided what to do with the answer. The exposure was a paragraph wide. If the answer was wrong you deleted it, and the worst case was five wasted minutes and a slightly worse mood.

The same subscription now accepts jobs. You describe an outcome, the thing goes away for forty minutes, and it comes back with a finished piece of work you did not watch it produce. Nothing about your billing changed to mark the moment. What changed is the size of the unit you are buying, and every decision you used to make in passing, about whether the output was any good, has been pushed into one review at the end. This guide is a way to decide which of your jobs are ready for that and which are not. It is written for someone running a small business or working alone who already pays for an assistant. It is not a rollout plan for a team, and it is not a security guide, though the last section touches the edge of one.

The unit of AI work is getting longer, and that is the whole story

There is now published evidence for what the change looks like at scale. In June 2026, researchers at OpenAI, Columbia Business School, Duke’s Fuqua School and Wharton released “The Shift to Agentic AI: Evidence from Codex”, filed as an economics paper and built on OpenAI’s own usage data [1][2]. They took a sample of user prompts and estimated, with a model, how long an experienced human would have needed to do each task without AI. In December 2025, 35.4% of active individual users had sent at least one prompt worth an hour or more of human work. By May 2026 that was 70.2%. Since December 2025 the share sending at least one prompt worth eight hours or more has risen from 2.1% to 25.6% [1].

That is more than a tenfold rise in five months in the fraction of people handing over something they would have had to block out a working day for. The number of weekly active Codex users grew more than fivefold between 1 January and 1 June 2026 [1]. The direction is corroborated from outside the vendor. METR, which benchmarks how long a task a model can finish on its own, put the best measured 50% time horizon at 320 minutes in January 2026, with the length of task models can handle doubling roughly every 196 days on their overall trend and every 89 days on the trend since 2024 [3].

Take both as shape rather than schedule. The Codex figures are estimates produced by a model reading prompts, not stopwatch times, and the paper says so plainly [1]. METR measured real human baselines for only 5 of its 31 long tasks and gives a confidence interval on that 320 minutes of 170 to 729 minutes, which is wide enough to drive a business through [3]. The durable finding is not any single number. It is that the thing you are buying keeps getting bigger while the price of the subscription stays where it is.

Verifiability decides what you can delegate, not difficulty

The obvious explanation for coding going first is that engineers were early adopters. The better explanation is that code tells you when it is wrong. There is a compiler, a test suite, and a diff you can read. Anthropic’s own documentation for Claude Code makes this the advice rather than the side effect: give the model verification targets, include test cases, define expected output, so that it can catch its own mistakes before you have to [5].

So sort your work by how expensive it is to check, not by how hard it is to do. Reconciling a month of transactions against a bank export is tedious and slow, and the check takes ten seconds because two totals either match or they do not. Drafting a proposal to a client you have known for six years is quick for you and unbelievably expensive to check, because the only real check is reading every line with the relationship in your head. The second job is easier. The first job is the one to hand over.

Applied honestly, this rule kills some of the delegation you most want to do. Anything whose correctness lives in your judgement rather than in the artefact is a bad candidate, however boring it feels. Anything that ends in a number, a file that either parses or does not, a list of links you can open, or a document you can diff against a source is a good one. The test to run before you start is a single question with a time attached. What will I look at to decide this is right, and how many minutes will that take.

Review is the line item nobody puts in the model

The review is not free and you will underestimate it. The strongest evidence on this is a randomised controlled trial METR ran in 2025 with 16 experienced open-source developers across 246 real issues, in repositories they had contributed to for multiple years [4]. The developers expected AI tools to make them 24% faster. They were 19% slower. After finishing, having lived through the actual slowdown, they still estimated that AI had made them 20% faster [4].

Read the caveats, because METR wrote them and they matter. The sample was small, the developers were experts in codebases they knew intimately, the tools were the early-2025 generation, and METR states explicitly that the study is not evidence that AI fails to speed up most developers, or that better ways of working with these tools do not exist [4]. What travels is not the 19%. What travels is the gap between the felt speedup and the measured one, in the direction you would least like.

The practical consequence is that you cannot run this on impressions. Keep a plain log of the jobs you delegate: what it was, how long the review took, and whether you shipped the output unchanged or rewrote it. Ten entries will tell you more about whether delegation is working for your business than any benchmark will, and it is the only way to catch the case where you are enthusiastically paying to produce work you then redo.

Price the job, not the seat

Subscription prices are set up to make the shift invisible. Claude Pro is $20 per month, or $17 per month billed annually at $200 up front, and includes Claude Code [7]. ChatGPT Plus is $20 per month, and Codex is included in every ChatGPT plan from Free upward, with ChatGPT Business at $20 per user per month billed annually [6]. Standard team seats on Claude are $20 per seat per month billed annually, or $25 monthly [7]. None of those numbers change when you stop asking questions and start handing over jobs.

The bill changes underneath them. On both platforms the included allowance is measured in usage inside rolling windows rather than in tasks, and once you pass it you are into paid credits or metered tokens [6][5]. A long job spends more than a long chat for a structural reason: the agent sends your whole accumulating context back with every step, so file contents, tool results and multi-step reasoning are re-read repeatedly [5]. Anthropic’s documentation is unusually direct about the resulting order of magnitude in enterprise coding deployments, at around $13 per developer per active day and $150 to $250 per developer per month, with costs staying below $30 per active day for 90% of users [5]. Running several agents in parallel multiplies it again, at roughly 7 times the tokens of a standard session when a team of agents runs in plan mode [5].

Those figures describe engineers on a vendor’s own tool, not you, and you should not budget from them. The useful part is the shape. Cost tracks active days and task length, not headcount, which means it does not scale with the size of your business and cannot be forecast from your seat count. Before committing to a job, price it the way you would price a contractor.

The delegation is already landing outside engineering

The part of the Codex research most relevant to a non-technical operator is what happened to non-technical departments. Inside OpenAI, Codex accounted for more than 90% of the average engineer’s AI usage by March 2026. In functions such as legal and recruiting, use was close to zero in January 2026, climbed gradually to about 20% of output tokens by early April, then rose from roughly 20% to 75% within a month [1][2]. Across the wider user base, the paper finds developers remain an important share of users, but growth is faster among non-developers [1].

Legal is the interesting one, because it is a function where being wrong is a professional problem rather than an inconvenience, and where caution is the trained instinct. Compliance-sensitive people crossing over that fast suggests the constraint was never appetite. It was whether the work had a shape an agent could take.

Hold the paper’s own warning firmly while you read that. It states that OpenAI usage is not representative of the typical organisation today, and offers itself instead as a view of what agentic use may look like in the future, when adoption frictions are minimal [1]. A company that builds the tool, sells the tool and organises itself around the tool is the fastest possible case, not the median one. What you can take from it is the ordering, not the pace: the work that goes first is the work with a mechanical check, in any department.

The paper’s own summary of what happens to intensive users is worth sitting with. They appear to manage portfolios of agentic work, with their role shifting toward delegation, supervision and integration [1]. That is a different job from the one you have been doing, it is mostly reviewing other people’s work, and nobody is much good at it on the first try.

What still goes wrong

The measurement underneath all of this is soft. A model estimated how long a human would have taken on those Codex prompts, and nobody timed a human doing them [1]. METR’s long-task figures rest on 5 measured baselines out of 31 tasks and a confidence interval running from under 3 hours to over 12 [3]. “An eight-hour task” is a label produced by a language model reading a sentence, and treating it as a stopwatch reading is the first way to talk yourself into delegating something you should not.

The second failure is the one the METR trial exposes, and no checklist fixes it. Delegation feels fast in a way that is almost independent of whether it is fast, because the waiting happens while you do something else and the review does not feel like work [4]. Only a written log settles that argument, and most people will not keep one, which is a reasonable prediction about you as well.

The third is quieter and does not appear in any research here. A delegated job that comes back 90% right is more dangerous than one that comes back obviously broken, because it survives a skim. The skim is exactly what you will do on the fourth job of the day, the one you delegated because you were behind. That is the real limit on how much of your business you can hand over, and it is set by your attention rather than by the model. Prices, limits and plan structures cited above were read from the vendors’ own pages on 5 September 2026 and move without much notice.