Before you post the next job, ask whether an agent should do the work

Stanford’s 2026 AI Index came out last week, and one number in it should change how you write job descriptions. Employment among US software developers aged 22 to 25 has dropped nearly 20% since 2024, according to Unite.AI’s summary of the report. The same report found that AI agent adoption inside businesses is still in the single digits in nearly every department.

Put those two facts side by side. Companies are hiring fewer junior people for work AI can now do, but very few of them have put an agent to work in a department on purpose. Most are making the hiring decision by default.

I think every open role should go through one test before it’s posted: could an agent do a real share of this work, with a person checking it? If the answer is yes, the job you post should look different. Sometimes you shouldn’t post it at all.

By the numbers

20%

drop since 2024 in employment of US software developers aged 22 to 25 (nearly)

Source: unite.ai

Shopify asked this question a year ago

In April 2025, Tobi Lütke told Shopify teams that before asking for more headcount they had to show why they couldn’t get the work done with AI (TechCrunch). He asked them to picture their area with autonomous agents already on the team.

A lot of people read that memo as a hiring freeze. I read it as a process change. It forced managers to describe the work before asking for a person to do it. Most hiring requests skip that step. A manager is busy, the team is behind, and the fix that comes to mind first is another person.

A year later, large companies are making this call in public. Nikkei counted 78,557 tech layoffs from January into April 2026 and attributed 47.9% of the cuts to a reduced need for workers because of AI and automation, as reported by Tom’s Hardware. Some of that is companies blaming AI for cuts they wanted to make anyway. Sam Altman has called that “AI washing,” per the same article. Some of it is real.

A 50-person company should ask the question more carefully than a giant does, because each hire is 2% of its headcount. Get one wrong and you feel it for years.

Break the job into tasks before you decide

A job description is a bundle of tasks. Say you’re about to hire a customer success coordinator. In a typical week she would answer account questions, update the CRM after calls, chase renewals, build a monthly usage report and handle the two unhappy customers who need a real conversation.

Take the role you’re about to post and list every task in it. Next to each one, write how many hours a week it takes and whether it has a clear right answer.

Then sort them:

Tasks with a clear input, a clear output and written rules are agent work. Drafting replies to common questions, updating records from call notes and pulling the weekly report all fit here.
Tasks that need judgment on cases the rules don’t cover belong to a person, with the agent preparing the file.
Tasks that depend on trust built over time, like a renewal negotiation with a key account, stay with a person.

Get the hours from the people doing the work today. Ask them to look back at their last two weeks of calendar and email and write down where the time went. Job descriptions are written to attract candidates. They rarely match the week someone actually has.

Now add up the hours. Say 25 of the 40 land in the first bucket. You’re not hiring a coordinator anymore. You’re hiring someone to run the accounts that need a human, with an agent doing the typing behind them.

Before and after

The job as posted

Answers routine account questions
Updates the CRM after calls
Builds the monthly report
Handles escalations

The job after the test

Agent drafts routine answers
Agent updates the CRM
Agent builds the report
Person reviews and handles escalations

The same role, split into agent work and the judgment work a person should own.

The AI Index shows where the hours come from

The productivity data in the report points the same way. It cites gains of 14% to 26% in customer support and software development, and up to 72% in marketing teams (Unite.AI). Those figures come from many studies and they won’t match your company exactly. They do tell you where the time goes: the repeatable parts of knowledge work.

The drop in young developer employment worries me for a different reason. Entry-level jobs are where people learn how a business works. If agents take every routine task and you stop hiring juniors, in five years nobody on your team will know how the work actually gets done, and nobody will be able to tell when the agent is wrong.

So the test has a second half. Once an agent has the routine work, decide what the person you hire will spend their time on. Often the answer is to hire the person anyway and give them a job that’s mostly judgment, with an agent handling the rest. That’s a better job. It’s also a harder hire, because you need someone who can read an agent’s output and catch its mistakes.

The test ends in one of three places

After you sort the tasks, you’ll land in one of three places.

Don’t post the role yet. The routine tasks are most of the job and the rest can go to people already on the team. Put an agent on the routine work with a named person checking its output, measure it for a month and revisit. If the agent can’t do it, you’ve lost four weeks and learned exactly what the hire needs to cover.
Post a different role. Rewrite the description around the judgment work. Make “reviews and corrects work drafted by AI” a stated part of the job, and interview for it.
Post the role as written. The work is mostly relationships, physical presence or decisions an agent can’t make. Plenty of roles land here, and that’s fine.

What you shouldn’t do is skip the test because the team is underwater. That’s exactly when it pays off, because the team is too busy to notice which half of its work a machine could take.

One warning. Don’t run this as a cost-cutting exercise in secret. Your people will see the job descriptions change. Tell them what you’re doing and why, and tell them the plan includes training them to work alongside the agents.

What to do this week

Pull every open requisition and every role you plan to post in the next 90 days.
Ask each hiring manager to list the tasks in their role with weekly hours for each. It takes 30 minutes.
Mark each task as agent work, agent-prepared or person-only.
If agent work is more than half the hours, hold the posting for two weeks and test an agent on those tasks first.
Rewrite the descriptions you do post so they say plainly that the person will review work drafted by AI.