Should senior employees get bigger AI budgets?

A public cybersecurity company gives junior staff an AI budget of $800 a month. Senior staff get between $1,600 and $4,000, and data scientists get the most of anyone. A global travel tech company starts everyone at $200 and raises the limit, sometimes to tens of thousands of dollars, depending on seniority and role. Uber introduced a cap of $1,500 per employee per month after its AI coding budget for the year ran out in four months.

These examples come from SemiAnalysis, which spoke to more than 50 companies about token spend this summer. The amounts vary a lot from one company to the next, but most of them share one rule: the more senior you are, the bigger your AI budget.

Why budgets follow the org chart

Seniority is the line that already exists. It sits in the HR system, finance trusts it, and it mirrors how companies set salaries and software access. When AI spend started growing faster than anyone had planned, it was the quickest way to put a limit in place.

It also rests on a fair assumption. Senior people usually own the work with the most at stake, like the key account, the board paper or the architecture decision. If AI helps with that work, it should be worth more.

Some companies use role instead, which is why data scientists often get the biggest allowance. One airline in the SemiAnalysis conversations ties budgets to the revenue of each project. Each of these is a reasonable guess about where AI creates value, made before anyone has measured it.

What the usage data tells us

In August, OpenAI published a working paper, written with researchers from Columbia and Wharton, on how 1,764 organizations use ChatGPT Enterprise. Once the authors added controls, revenue per employee showed no meaningful link with the output tokens each employee generated.

The same paper found that early-career workers send many more messages than their senior colleagues. So the heaviest users sit lower in the org chart, while the biggest budgets sit higher up. Usage volume and job title each describe something real, but neither tells you what the spend produced.

Where seniority helps and where it doesn’t

The research on who gains most from AI points in two directions. Studies of customer support teams found the largest gains among newer, less experienced agents. More recent studies of coding agents suggest experienced engineers get more out of them on complex work, because they know what to ask for and when to reject the answer.

Both findings can hold in the same company. Seniority mixes two separate things: how valuable a piece of work is, and how well a person uses AI to do it. A senior manager might do high-value work and use AI poorly. A junior analyst might use AI very well on work nobody has priced yet. A budget set by job title can’t tell these two people apart.

A worked example

Picture a 40-person sales team spending $6,300 a month on AI. Most of it goes on three types of work. The figures are illustrative.

A seniority budget would protect the proposals, and it should. Each successful proposal returns more than three times its cost in time saved.

The other two rows are where the real decisions sit. Call recaps take the largest share of spend because everyone writes them, every day. They are cheap per result, so the question is whether they need the most expensive model at all.

Pipeline analysis gives the team its best return: each successful analysis is worth more than seven times its cost, and it comes from the most junior people on the team. Half of those analyses get abandoned, though. A shared template or skill that lifts that rate would do more for the team than a bigger allowance for anyone.

A budget set by job title would miss all three of these decisions.

Three questions for every task type

1. Did the task succeed? A proposal that gets sent or a report the team uses is worth funding. A task abandoned after four attempts cost money and produced nothing.

2. What does a successful task cost? Divide the AI spend on a task type by the number of tasks that succeeded. This is the number to compare across teams, in place of spend per head.

3. What is a successful task worth? Start with the time saved, multiplied by the hourly cost of the person doing the task. Where you can trace it further, into a deal won or a ticket closed, use that.

Time saved is capacity. It becomes a cash saving only when costs actually fall. That makes it a good basis for comparing task types, and a weak basis for promising finance a budget cut.

Senior people often come out ahead on these questions, because their tasks tend to be worth more. When they do, their bigger budget has evidence behind it, and it can change when the evidence does.

Setting it up

Give everyone the same default allowance, large enough to try AI on new work without asking for permission. A new use of AI can’t prove its value until someone tries it.

Add headroom for the task types that reliably produce results people keep, whoever does them. Prefer soft limits that alert a manager over hard caps that stop someone halfway through a task.

Review the split every quarter against cost and value per successful task, and report it by team and task type. Individual rankings tend to reward volume. Amazon and Meta both saw internal token leaderboards gamed this year, with staff running AI agents on busywork to climb the rankings.

A better answer for finance

Sooner or later, a CFO asks why one team spends three times as much on AI as another. Pointing to seniority is a quick answer, but finance has no way to check it. Cost and value per successful task take longer to explain, and finance can test every part of them: which work the money went to, how much of it produced something people used, and what that was worth.

Evidence for every AI budget decision

Nebuly gives AI leaders the data to answer those three questions. It reads the conversations employees have with their AI tools, splits them into the tasks people were trying to get done, and flags which conversations ended successfully. ROI counts time saved only on successful work, multiplied by the hourly cost you set. AI spend sits next to it, broken down by task, model, team or tenure, so you can see whether longer-serving staff actually get more from their budget. Model savings then shows where a cheaper model would do the same job, like the call recaps above. Nebuly doesn’t set or enforce budgets. It gives the people who set them the evidence to do it well.

Want to see this in practice? Book a demo with us.