What the finance department knows that IT doesn't

The bill doesn't lie

A subscription gets taken out because someone needs it. That is the quality that makes procurement and licensing data valuable when building an AI inventory. An employee who uses a language or image model without IT knowing usually still pays for it: with a credit card, through an expense claim, via a team card, via a department budget. That transaction is recorded somewhere, even if no one has flagged it as "AI use."

That makes financial data a different entry point than the IT list. Where why the IT list doesn't add up at an organization with multiple departments explains why formal approval and actual use diverge, this trail shows what does get recorded despite that gap: money spent on something.

Where you get it from

The sources don't sit with one department. Relevant ones include:

Credit card statements and expense claims. Recurring charges to known AI vendors, often under names that aren't immediately recognizable as "AI."
Invoices held by accounts payable. Contracts that run through procurement, including one-off implementation costs that point to a subscription.
SaaS management tools, if the organization uses one: these often show which applications are actively linked to company accounts, regardless of whether they were requested through IT.
Single sign-on and identity logs. Who logs into which service with a company account, even without a central contract in place.
Departmental procurement registers. Especially at organizations where departments manage their own budget, subscriptions run through here that were never reported to IT.

None of these sources gives a complete picture on its own. Together they expose a pattern that the IT list doesn't show.

What you record per item found

The goal is not a list of amounts, but a list of applications with their origin. For each subscription or contract found, it's worthwhile to record:

the name of the vendor and product, as it appears on the invoice
the department or cost center the expense comes from
whether it's an individual subscription, a team license, or an organization-wide contract
the presumed period of use, derived from the first and last invoice date
whether the application already appears on the existing IT list, or whether this is a new find

These fields are a starting point. A more complete overview of what's needed per application to reach a classification is described in what you record per application at an organization with multiple systems. Procurement data mainly provides the starting point: a name, a department, an amount. The rest of the classification follows later.

The limit of this method

Procurement and licensing data mainly finds paid use. Free versions of AI tools, trial periods, and personal accounts not billed through the company remain invisible this way. Services embedded in an existing contract — an AI feature added to software that was already in use, without being billed separately — also don't surface in an invoice overview.

That's one of the reasons why financial data is one source alongside others, not the only one. Other material, such as network traffic or browser extensions, finds precisely the use that costs no money. Which technical signals are useful for that, and what they do and don't show, is described in which IT signals are useful for detecting AI use. Both trails together give a better picture than either alone.

Asking questions without presenting the bill

Anyone who comes across an unfamiliar subscription in the records can ask the department that took it out directly. That works better if the question isn't posed as an audit. A department head who suspects an expense is being questioned is less likely to give complete information than someone who thinks the question is only meant to build a full picture. How to phrase that question so that someone answers instead of shutting down is worked out in how to ask employees without it feeling like a reckoning, at an organization with existing AI applications.

One source in a larger process

Procurement data is one of the building blocks in building an organization-wide inventory, not the whole process. How this source relates to employee surveys, technical signals, and the existing IT list, and in what order it's practical to tackle this, is described in how to build an AI inventory at an organization with multiple departments.

From knowing what's running to knowing what it's worth

An inventory based on procurement data answers the question of which AI applications exist and who pays for them. That doesn't yet say what those applications actually deliver, or what share of the work they were bought for is genuinely being done by AI. That question falls outside the scope of a governance inventory and is answered by the work scan from FTE TO AI, which calculates per task what share of the work can be taken over by AI. Once you know which applications are running in the organization, that scan lets you determine what work they are actually shifting.