AI budgets do not leak because a tool costs too much. They leak because of paid seats nobody touches, duplicate tools across teams and subscriptions with no owner. Each line looks small, and it all adds up at the end of the month. The answer is measurement, not prohibition.
Where your AI budget leaks
The leak almost always comes from the same places. None of them is dramatic on its own, but together they become the fastest growing line in the IT budget. We see it in every company that started buying AI tools before it had a process.
- Seats purchased and never used. They bought it for the whole team, and in practice three people log in.
- Duplicate tools. Marketing, development and support bought three separate tools that do the same thing.
- Shadow subscriptions. A personal credit card, an expense reimbursement, and nobody in the IT department knows about it.
- Trials that rolled over. A free trial that turned into an annual subscription without anyone approving it.
- Expired credits. Paid up front for usage, never used it, the date passed.
- Employees who left. The seat stays active and the billing keeps running.
What all these lines have in common is that nobody owns them. When nobody is responsible for a line item, nobody cancels it either. This is not a story about wasteful employees, it is a story about a missing process, so the fix is a process too, not a lecture.
A dormant seat is worse than an expensive one
This is the logic missing from most cost discussions. A manager sees a high price on a tool and tries to negotiate it down, while paying for 20 seats that nobody touches.
An expensive tool that 10 people use every day pays for itself. A dormant seat returns nothing, and it distorts the picture. It inflates spend, drags down average utilization and makes it harder for you to justify expansion when you genuinely need more seats.
There is a quieter cost as well. When leadership sees a large sum with no results, the reflex is to freeze purchases of new tools. So the teams that would have benefited from the right tool go without, because of someone else's empty seats.
Cost per active employee, not price per license. Two tools at the same price can be one cheap and one very expensive, depending on how many people actually use them.
Duplicate tools, personal cards and shadow subscriptions
In companies that grew fast, this is almost inevitable. Every team solved its own problem, and every solution came in through a different door. The result: there is no single list of tools, and nobody sees the full total.
- Build one list: every AI tool, who owns it, what it costs and which payment method it is billed to.
- Go through six months of expense reports. That is where the shadow subscriptions show up.
- Ask accounting for every recurring charge, especially the ones billed in dollars.
- Consolidate tools that do the same thing and pick one per need.
- Move every subscription to a single company card or a managed vendor account.
Once you have the list, you can start managing. Before that, every conversation about savings is an educated guess. It is also worth defining who approves a new tool, otherwise the list will be stale within two months and you will be back where you started.
Per seat or per usage
Before approving a new tool, understand how it bills. There are a few common models and each suits a different situation. Keep in mind that the price on the vendor's site is almost always in dollars and before VAT, and in Israel 18% is added on top.
| Pricing model | When it pays off | What to watch |
|---|---|---|
| Fixed per-seat | Daily, steady use by the same people | Every unused seat is a total loss |
| Per usage or tokens | Uneven use, projects and peak periods | Easy to miss a sudden spike in the bill |
| Base subscription plus usage | A small team that runs a lot | Needs a monthly cap and an alert |
| Prepaid credits | A discount in exchange for commitment | Credit not used by the deadline is lost |
In practice, tools the team opens every morning are best bought per seat, and tools used for projects are best run on usage pricing with a monthly cap and an alert. If a vendor offers a big discount on an annual commitment, first check how many seats were actually active over the last three months. Committing to a number you never verified is exactly how dormant seats are born.
How to measure real usage per employee
A team lead's gut feeling is not a measurement. You need data, and almost every vendor exposes it in the admin console or through the API. The good news: it is an hour of work the first time, and after that it runs on its own.
- Pull the list of paid seats from every vendor, including who is assigned to each seat.
- Pull login and usage data from the API, not from the team lead's report.
- Flag every seat untouched for 30 days as dormant.
- Split the active users into heavy, medium and light usage.
- Calculate the cost per active employee for each tool. That is the only number that goes to the management discussion.
The way to stop guessing is to see everything in one place. In theIT management dashboard that we build for every client, there is a module that shows, for each tool, paid seats against active users, plus the split between heavy and light use. For illustration only, here is what one row looks like: 40 paid seats, 28 users who signed in over the past month, 12 dormant, 70% utilization. When you see that in a single row, deciding what to cancel takes two minutes. The numbers here are illustrative examples, not client data.
A 30-minute monthly routine
One-time savings do not last. What works is a short meeting on a fixed date, with the same table and the same four questions.
- Open the subscription table and compare spend against last month.
- Review the dormant seats and decide whether to cancel or reassign to another employee.
- Check whether new tools came in, and whether they duplicate a tool you already have.
- Approve next month's budget based on actual usage, not on what was requested.
- Flag renewals and credits that expire within the next 30 days.
Half an hour a month sounds like very little, and that is exactly the point. A short meeting that always happens is worth more than a big audit that happens once a year, because canceling a dormant seat only saves money if it is done before the next charge, not after.
What to ask before approving a new tool
Cost is only half the decision. An AI tool receives business data, and sometimes customer data too, so there are a few questions to ask before approving it, not after.
- Where the data is stored, and who on the vendor's team can access it.
- Whether what you enter is used to train models, and how to turn that off.
- What deletion actually means on their side, and how long logs are retained.
- Whether it connects to Entra ID or Google Workspace, so access can be cut off the day an employee leaves.
- Who in the company approves a new tool, and which payment method it is billed to.
With our clients managed IT services we go through these questions before a tool comes in, and we connect it to your organizational identity. That way both the budget and the data stay under control, and nobody finds out after the fact that a client contract was pasted into a tool with no idea where it stores things.
FAQ
What counts as a dormant seat?
A paid seat that nobody has signed into during the period you define, usually 30 days. Some vendors show this directly in the admin console, with others you need to pull the data from the API. Once a seat is flagged as dormant, the decision is to cancel it or reassign it to an employee who will actually use it.
How long does it take to see savings?
The first month is mostly mapping, and that is where shadow subscriptions and seats belonging to former employees surface. Cancellations can take effect as early as the next billing cycle. From the second month on, it becomes a short routine instead of a project.
Should we just ban employees from using AI tools?
A blanket ban pushes usage onto personal credit cards and accounts you cannot see, which is worse from a security standpoint. It is better to approve a small number of tools, define what may be entered into them and measure usage. That keeps both the budget and the data under control.
How do we know a tool is safe for company data?
Ask where the data is stored, whether it is used to train models and how to turn that off, and what actually happens when you delete it. It should also connect to your organizational identity, so offboarding an employee cuts their access immediately. We go through these questions with the client before any new tool is approved.
Who should own the spend?
Every tool needs one owner, usually the manager of the team that uses it, with someone above them who sees the full picture. Without an owner, nobody cancels a dormant seat. In a managed service, we hold the full picture and bring it to you once a month for a decision.
Related to: IT management dashboard · Managed IT services · What managed IT services include · Assessment call