Recently, one of our clients, a regulated financial services firm in Europe, enabled Claude alongside Copilot, giving its teams two AI assistants to accelerate productivity. The leadership team was thrilled but I must admit I was taken aback by a regulated firm handing its teams that much power. Three months later, finance disabled access because of exorbitant costs, about $10,000 a month that wasn't budgeted or tied to a tangible outcome.
I told the team, and almost every HR leader since, that high costs are not bad per se. If usage is high enough to make a dent in the budget, people are building great products. The AI bill for your knowledge workers means your people are using AI. The job now is putting it to work on hiring, on building best practice around hiring, and on the people processes behind it. Done well, a team of 100 starts to look like a team of 150.
The Costs Become Manageable With Time
AI is expensive, and sometimes it has to be. What catches people out is that the whole spend sits on one budget line when it behaves like five different things. Seats are the line everyone can see, at about $30 a month each, cheap and predictable. The knowledge worker who has learned to use the tools properly is a different animal, running past $100 a month by my own estimate from the people we have trained, and an engineer with the right access can burn $1,500 a month in tokens without anyone noticing. Then there are software renewals, where every vendor wants to charge you for the AI it switched on, and agents, which nobody has budgeted for at all. Everyone watches the seats. The money is going to the engineers and the agents, where nobody is looking yet.
The other thing I see is how far apart companies are. In Europe, most of the HR rooms I walk into are still talking about ethics and governance, and the question of cost has not come up. In Las Vegas last month, half the room told me they were still working out who gets a license. The United States went into build mode a year ago and has now swung the other way: stop building, this is way too expensive. Only a handful have reached the stage HR should own, which is asking what the company gets back and who decided.
Nobody Owns the Bill, So Make It an Allowance
According to Harness, 52 percent of enterprises say nobody clearly owns AI costs. Futurum puts 46.9 percent already over plan. No employee controls the bill, but every employee understands an allowance. Give each role one, and treat it as a privilege that grows with use.
In practice, every job gets two numbers: a salary and a monthly AI allowance. The allowance rises with seniority and with how much the person does with it. When spend reaches 80 percent of the allowance, the manager gets a note. At 100 percent, manager and employee sit down and decide whether to raise it. Nobody gets switched off, since the first person to hit the cap is usually your heaviest user. Cutting off your best user tends to be the most expensive cap of all.
This is where HR earns its seat. Of the five lines on the bill, three can carry an HR name: seats, knowledge workers, and the agents that run HR's own workflows. AI is bought like software but behaves like labor, so it belongs with the people who already manage labor. As someone said at a table in Las Vegas, it is harder to open a requisition today than it was in 1985, yet anyone can now spend on AI capacity without asking.
The Work Got Faster, the Waiting Stayed the Same
A job description now takes minutes, but opening the role still takes weeks. In an illustrative hiring workflow, opening one role takes five handoffs over 15 days, with 24 hours of hands-on work inside them. Put AI on every step and the same five handoffs take 11 days and nine hours of work. The drafts got faster. The queue did not move. Redesign the workflow, with fewer handoffs and a packet that gives each approver what they asked for, and the same role opens in two days.
The hours AI saves go somewhere. Workday found that 37 percent of the time AI saves goes to rework. BetterUp and Stanford found that four in ten people have received AI work that only looks finished. In 1955, C. Northcote Parkinson wrote that work expands to fill the time available for its completion, and AI is testing his law at scale. Saved time with no decision about what it is for fills up with comfort work, which contributes to nothing and looks impressive.
Where This Leaves HR Outsourcing
If your HR services sit with a partner, the same three questions move into the contract. What are we paying for? Ask whether the provider bills by seat or by task, and ask for the cost of a job description, a shortlist, an exit-interview summary, and a policy answer. What is the impact? A provider that adds AI to every step of your process and leaves the handoffs alone has sold you the 11-day version. Who decided? Put one name on each line, on your side as much as theirs, and agree a simple rule for where the human sits: if a task can be undone, let the automation run. If it cannot, a person signs. Automated rejections in hiring fall on the wrong side of that line, and most HR leaders I spoke to this year agreed.
So the next time finance asks why the AI bill keeps growing, somebody is probably building. The question for HR is what that time is now for.
Dominique Rose Van-Winther is co-founder and CEO of Final Upgrade AI and its Chief AI Evangelist. She spoke on the price of intelligence at HR Tech in Las Vegas in October 2026.
