Can AI answer questions about your IT inventory?

An assistant can answer questions about your IT inventory, but only the questions your records already contain the answer to. Ask which contracts end in the next 60 days and it reads them off the register in a second. Ask what you paid for Salesforce when nobody ever recorded the price, and the only honest answer is that the data is not there. The difference between those two behaviours is the whole design, and it is worth understanding before you buy anything that says AI on the box.

What can it actually answer?

Questions whose answer is a fact already in your register, phrased in whatever words come to mind rather than the filter names the software expects.

The pattern: an assistant turns a question into a lookup. It does not turn a question into knowledge you never captured.

What makes one trustworthy?

One behaviour above all: what it does when it does not know. A model asked about a contract price it has never seen will, unless told otherwise, produce a plausible number. That number looks exactly like the true ones, and it is the reason a lot of people distrust AI in an audit context.

The fix is not cleverness, it is instruction and scope. Give the model only a summary of the workspace, tell it to answer from that summary alone, and tell it to say so when the answer is not there. Then test it: ask something you know is missing. In InventorIA that answer reads like this, which is the behaviour to insist on from any vendor:

The provided workspace summary does not contain information about the Salesforce renewal price. This detail would be located on the Contracts screen.

An assistant that says that is more useful than one that is right most of the time, because you can act on its answers without checking each one.

What should it never be sent?

Every question sent to a model provider leaves your building. That is a fact to design around, not a reason to avoid the feature.

• No personal data. Names, email addresses and serial numbers are never needed to answer a question about renewal dates or seat counts. Send counts and vendor names, not people.
• No other customer's data. The summary must be built from the caller's own workspace, scoped at the query, not filtered afterwards.
• Not the whole database. A bounded summary is cheaper, faster and leaks less. Ours carries counts plus the contracts and licences ending within 60 days, capped at fifteen each.

And use a paid API tier. Free tiers at several providers reserve the right to use prompts to improve their products, which is a poor place for your customers' vendor names to end up.

How do you stop it becoming an open invoice?

A text box that calls a paid API on every keystroke of curiosity needs a ceiling, and the ceiling has to live somewhere a restart cannot reset. Three limits, checked before the call rather than after:

1. Per user, per minute. Stops one person, or one script, looping.
2. Per workspace, per day. Stops one customer draining the month.
3. Across everything, per month. The hard stop that protects the card.

Two details matter more than the numbers. Keep the counter in the database, not in memory, or a restart hands out a fresh budget. And charge on total tokens, not just the visible answer: reasoning models bill the thinking they never show you, and counting only the visible part under-counted a real call of ours by six times.

What does it cost in practice?

Less than most people expect. A grounded question over a bounded summary is a small prompt and a short answer. Measured on our own deployment with Gemini Flash, seven real questions cost 135 micros of a euro, about a hundredth of a cent. The monthly ceiling of 10 euros is therefore not a budget, it is a tripwire: it exists to catch a bug or an abuse, and normal use never approaches it.

That changes how you should think about pricing such a feature. If a vendor charges a large AI surcharge per seat, ask what the underlying calls actually cost.

Is it worth having at all?

It is worth having when it removes a lookup you would otherwise skip. Nobody opens a contracts table to check renewals on a Tuesday afternoon; plenty of people will type a question if the box is there. The value is not that the assistant knows more than you, it is that it lowers the cost of asking to nearly zero, and questions nobody asks are how a renewal slips past its notice deadline.

It is not worth having if your register is empty. An assistant over three spreadsheets and a memory is a confident voice with nothing behind it. Record the data first: see what to log for a renewal reminder and how to measure licence utilization.

Frequently asked questions

Can AI manage my IT assets automatically?

No, and be wary of anything that claims it. An assistant reads what you recorded and answers questions about it. Deciding what to renew, cancel or reassign stays a human decision with money attached.

Will my inventory data be used to train a model?

It depends on the tier. Several providers reserve that right on their free tiers and not on paid ones. Ask your vendor which tier they use, and what fields they send. The answer should exclude names, emails and serial numbers.

What happens if the AI does not know the answer?

A well-built one says the data is not in the record and names the screen that would hold it. If it produces a number instead, test it with three questions you know the answers to before trusting any of them.

How much does an inventory assistant cost to run?

Far less than most pricing implies. A grounded question over a bounded summary costs a fraction of a cent on a current fast model. Measured on our deployment, seven questions came to about a hundredth of a cent in total.

Can I try one without signing up?

In InventorIA, yes. The public demo workspace includes the assistant, so you can ask it about a populated register without creating an account.