Tagging in theory, tagging in reality, and what AI changes

I have written before that tagging is an input rather than a precondition. That distinction sounds like a technicality until you notice what it decides: whether your business gets an answer about its technology spend this month, or after a programme of work that has to finish first. This piece is about why that question is becoming increasingly consequential for you and your business every quarter, not less.

Start with tagging at its best, stated fairly. A convention is agreed, a key and a value, applied to each resource at the moment it is created, inherited consistently by everything downstream, and finance ends up with something close to a coded ledger of technology transactions. Cost rolls up by team, by environment, by product. On paper the model is sound, and where it has been maintained with real discipline it does useful work.

The reality is the one every finance team recognises. The coding is done outside finance, at provisioning time, in free text, by someone whose priority that morning was shipping. There is no validation, no approval, no chart of accounts to code against, and no equivalent of a rejected journal, so an untagged resource is not an error to be corrected, it is simply a cost that arrives without an owner. Then the estate moves underneath it. Teams reorganise, products get renamed, an acquisition arrives carrying a convention nobody still at the business wrote, and the labels quietly stop describing the thing they were applied to.

None of that is new, and it is worth being honest that an entire category of technology cost management tooling was built with those problems already understood. Its answer was to treat the tagged estate as the entry condition. Implementation begins with a tagging remediation programme, the first deliverable is a taxonomy, governance is stood up to enforce it, and the actual answer about what technology spend produced arrives once the data has been made worthy of the tool. The assumption underneath that sequence is that the estate can be brought to a standard and then held there. Everything else follows from it.

Tagging assumes a human was in the room when the cost was created, a provisioning-time answer to what is now a runtime question.

Which is the assumption AI-accelerated development is taking apart, and it goes first at the volume. Infrastructure is defined in code, provisioned by pipelines, scaled by policy, and increasingly authored by agents rather than typed by people. Nobody needs convincing that the speed of software development is rising sharply, and that AI is why, so the resources behind that work are now created at a rate no tagging convention was designed to keep up with. A standard that depends on a person remembering it at the moment of creation degrades in direct proportion to how few creation events involve a person at all. The remediation programme is sized against an estate that is growing and changing faster than the programme can close.

The second failure is the structural one, and it is the reason discipline cannot fix this even in principle. Tagging is a provisioning-time concept: it labels an object at the moment that object comes into existence. AI cost is not shaped like that. An inference call is not a resource you can tag, it is an event, triggered by a feature, at a volume nobody provisioned and in a quantity that varies with how customers behave that week. There is frequently no durable object to attach a label to. So for the fastest-growing line on the technology bill, the entire mechanism does not apply, and no amount of governance will make it apply.

Put those together and the calculation runs the wrong way on both axes. The total grows, and the proportion of it that any labelling scheme could ever have captured shrinks, so the absolute size of what finance cannot explain rises twice over. A model whose entry condition is a well-tagged estate is not merely inconvenient in that environment. It is chasing a precondition that is receding faster than it can be met, which is another way of saying the model is structurally obsolete rather than temporarily behind.

To be clear, this is not an argument against tagging. A well tagged estate is genuinely easier to reason about, it improves the resolution of everything built on top of it, and the discipline is worth having. The argument is against tagging as a gate. The thing being waited for is moving away from the person waiting.

This is the difference in how OPTIMAZE is built, and it is worth stating plainly because it is the part that changes what a finance team can do this quarter. Tagging is a beneficial but not essential input. It is not a prerequisite, not a phase one, not a data-readiness gate, not a programme to be funded before the platform will tell you anything. OPTIMAZE derives attribution from how resources, accounts, services and any other available metadata actually relate to one another, so a defensible view of what technology spend produced exists on an untagged estate, on a partially tagged estate, and on an estate carrying three incompatible conventions inherited from three acquisitions. Attribution is maintained as rules over relationships rather than a static chart of accounts, which is why a reorganisation or a rename is absorbed rather than requiring the estate to be retagged and the previous year restated. Tags, where they exist, are read and used. Where they do not, the answer still arrives.

A business that can only understand the spend its engineers remembered to label will understand less of its technology cost every quarter from here. A business whose attribution is derived rather than declared will understand more of it. That second position, and the confidence it gives finance to stand behind a technology number without a data programme in front of it, is what OPTIMAZE calls Technology Capital Performance.

If this sounds like something you or your business needs, get in contact with our sales team.