Don't Cut Jobs When an AI Agent Goes Live

When an AI agent goes live in one of your teams, say the billing desk, it arrives with a number: the hours it will save, from the vendor's proposal or your own builder's estimate.

Then comes the staffing question, and the arithmetic looks easy: if the agent takes 40% of the team's hours, the team can be 40% smaller. Get it wrong and the same desk pays: a backlog, overtime, and the people who knew the odd cases gone.

We build AI agents for companies of a few hundred people, and I'd answer that question differently.

Don't cut anyone on the strength of the proposal. Run the agent for a full cycle of the team's real work, a quarter for most back-office jobs, and count only what it finished that nobody had to redo.

Then take the saving out of your hiring plan first: leave jobs open as people leave that team, and offer its people the openings you were about to recruit for.

Cut only where the record adds up to whole jobs that departures won't absorb in time, and never the people who handle what the agent hands back.

The hours on a proposal are a forecast. Cutting jobs is the hardest step of a rollout to undo, and a hiring freeze is the easiest. And people leave every team, so a freeze collects much of the saving at the pace they go.

This page shows why the proposal is a forecast, what a wrong cut costs, how a freeze collects the saving, and how Qvery builds the agent with a weekly count to decide on.

If the business is shrinking for other reasons, decide on those numbers; the agent changes nothing about them.

The Hours on the Proposal Are a Forecast Until the Agent's Record Confirms Them

A proposal's hours are a multiplication: items a month, times the minutes the agent saves on each, at the error rate it showed on the documents it was tested on.

What your team gets back is smaller, by three things the multiplication leaves out: the minutes someone spends checking each output, the items that come back wrong and get done twice, and the exceptions the agent hands to a person, which were the slow ones to begin with.

The gap shows up where finance looks, once a vendor's hours are held up against payroll and revenue:

A study of Danish workers using AI chatbots, a lighter tool than an agent, found the same gap across 11 occupations: people reported saving about 3% of their hours, and the national registers showed no measurable change in their recorded hours or earnings.

Hours saved are capacity. They turn into money only when someone decides where they go: a backlog cleared, a check that never got done, a job you don't have to fill.

So count the agent the way your accountant would: the items it finished that nobody redid, confirmed in the system where they land (the payment released, the file closed), times the minutes each used to take, less the minutes people now spend checking it and fixing what it got wrong.

A two-week pilot tells you whether the agent works on your documents.

How many people the team needs takes a full cycle: three month-ends, a bad week of scans and the customers who always send the odd ones.

A Cut Made on the Forecast Is the Hardest Step to Take Back

In July 2025, Commonwealth Bank of Australia announced it would cut 45 customer-service jobs after putting an AI voice bot on its phone lines, which it said was bringing calls down.

The Finance Sector Union said the bank put the drop at 2,000 calls a week and wouldn't show the numbers behind it.

The union's members said the opposite was happening: calls were rising, the bank was offering overtime, and team leaders were answering phones.

In August the bank reversed. A spokesperson said its first assessment "did not adequately consider all relevant business considerations, and this error meant the roles were not redundant," and the bank apologized and offered the 45 their jobs back.

Klarna took the other route. In February 2024 it said its AI assistant was doing "the equivalent work of 700 full-time agents", a month after launch.

It had stopped hiring, its chief executive said in December 2024, and its workforce had shrunk 22%, mostly as people left.

In May 2025 he said cost had weighed too heavily in the plan: "what you end up having is lower quality," and Klarna began recruiting customer-service staff again.

Two cases set no base rate, and both companies are far bigger than yours. What they show is the price of each correction.

Klarna's was a recruiting drive, after a stretch of service its own chief executive called lower quality. The bank's was an apology, a dispute before Australia's workplace tribunal, and weeks in which 45 people didn't know whether they'd keep their jobs, and it was caught before anyone had left.

Once people have gone, what leaves with them is the work the agent doesn't do. The agent takes the routine items and hands back the rest: the page it can't read, the customer whose documents are always wrong, the file that needs someone who remembers last year's dispute.

The people who handle those sit on the same team a cut is drawn from, and in the week an agent goes live, nobody has a record yet of which ones they are.

But the Work Does Shrink, and Paying People to Wait Costs Money

C.H. Robinson's North American freight brokerage has 35.58% fewer employees than at the end of 2022, while its shipments per person rose more than 60%.

In ADP's payroll records for millions of U.S. workers through June 2026, Stanford's Erik Brynjolfsson, Bharat Chandar and Ruyu Chen find that employment of 22- to 25-year-olds in the jobs most exposed to AI stands 19% below where it would be had it kept pace with their less-exposed peers.

The authors call their findings descriptive, not proof of cause.

So every month you wait, the objection goes, you pay salaries for work that may already be gone.

In those payroll records, though, the gap comes mostly from hiring: fewer young people taken on in those jobs, not more people let go. Separation rates fell in exposed and unexposed jobs alike, and experienced workers in the most exposed jobs show no comparable gap.

That is how a hiring freeze works too. It has a cost, and the same study shows who pays it: young people who never get the job.

I'd still take that over a list of names drawn up on a forecast, and I'd keep hiring juniors into the work the agent leaves to people, because that's where your next exception handlers come from.

A Hiring Freeze Collects the Saving at the Speed People Leave

People leave teams on their own, at a pace you can look up in the Bureau of Labor Statistics' count of job turnover. Even finance and insurance, the slowest of these four industries, saw 15.6 quits for every 100 jobs in 2025.

Your own team's record of who left in the last two years is a better guide than these averages. In a small team nobody may leave for a long time; there, the openings elsewhere matter more than the freeze.

Freeze the job, not the person.

When someone leaves the team the agent works in, leave the job open, unless the person leaving is the one who handled the exceptions. That job you fill, which a freeze lets you do and a list of names to let go doesn't.

Meanwhile the rest of the company keeps hiring: at the 2025 rate in finance and insurance, a company of 200 fills about 50 jobs a year, and the ones that suit a loan processor come with someone who already knows your customers and systems.

Where it's been asked, companies using AI rarely let anyone go over it. Among the AI-using service firms in the New York Fed's August 2026 survey of New York and northern New Jersey:

Take one team: ten people who key and check loan documents at a lender. The agent's proposal says it takes 40% of their hours, and a quarter of counting confirms 20%, about two people's worth, once the checking and the redone files are in.

The wrong cut's price is yours to fill in: what it costs you to hire and train two people in that role, plus the months the team runs short.

How Qvery Builds the Agent and Counts Its Hours Every Week

Each week an agent of ours runs, we report the record this page asks for: the hours it handed back after the checking and the redone items, and the errors it stopped before a customer saw them, such as a figure keyed wrong or a file closed with a page missing.

The agent works through one team's routine items, say the loan documents your processors key and check, on autopilot in the systems that team already uses, so each item it finishes shows up there as a file closed or a payment released.

What it can't finish goes to whoever owns that kind of case, marked with why it stopped.

Every agent we build belongs to one company and one team. We start by timing the job as that team does it today, and by the end of the second week a first working version is handling its real documents. Then we train whoever checks its work and takes its handoffs.

It is the count your accountant would make, kept from the agent's first week.

Turn the Record Into Whole Jobs Before Anyone Talks About Cuts

So the week an agent goes live in one of your teams:

1. Write down the work it takes before it starts

   Items a month, minutes on each, who does them, and the same count for your busiest week. Every later number is measured against this.

2. Tell the team the plan on day one

   Nobody is cut while the agent proves itself, and jobs that open up on the team stay open until the record is in. The people you need most shouldn't have to guess.

3. Count for a full cycle of the work

   A quarter for most back-office work, longer if your year has a peak season, counted the way your accountant would.

4. Freeze the team's open jobs and offer its people your openings first

   Fill any job the team can't cover without, starting with whoever handled the exceptions.

5. Turn hours into whole jobs in your busiest week

   Hours spread thinly across ten people are a lighter week for each of them, not a role you can remove. Then check that the people left can still cover the exceptions, holidays and sick days.

6. Let departures do what they can

   Set the whole jobs the record found against how many people left that team in the last two years. Cut only the difference that leavers and transfers won't close within a year.

Run it that way, and the conversation about the team's size starts from what the agent did instead of what its proposal said.