Step 1: Decide what you are allocating to
Before touching any numbers, choose the dimensions you will charge cost against. For most enterprises the useful ones are team or cost centre, use case, and model or provider. Team answers who is spending. Use case answers what the spending is for, which is the dimension finance and the board actually care about. Model tells you where a cheaper option would change the bill. Pick these three to start. Adding more later is straightforward once the plumbing exists, but starting with too many makes the model impossible to populate.
Step 2: Inventory every source of AI spend
List every place AI cost originates: each model provider, every copilot subscription, the point tools individual teams bought, and the usage sitting on personal accounts and expensed cards. Shadow AI belongs in this inventory even though it is inconvenient, because a chargeback model that ignores a third of real spend will be wrong from day one. The output of this step is a single list of spend sources, each tagged with an owner. Expect this to take longer than you think and to surface tools leadership did not know were in use.
Step 3: Tag usage so it can be attributed
Allocation only works if each unit of usage carries enough context to route it. Where you control the integration, pass a team and use-case identifier with every request so the cost lands in the right bucket automatically. Where you do not, fall back to the account or key the usage runs under and map that key to a team once. The goal is that the great majority of spend attributes itself without a human sorting invoices at month end. The minority you cannot tag goes into an unallocated bucket you work to shrink over time.
Step 4: Set the allocation rules
Write down how each cost is split. Direct usage that carries a team and use-case tag is allocated straight to that bucket. Shared platform cost, the audit trail, the gateway, the monitoring, is split by a fair driver such as share of total calls or seats. Unallocated spend is either held centrally or apportioned across teams by usage share, but decide which and be consistent. These rules are the heart of the model. Keep them plain enough that any team lead can understand their own bill, because a chargeback nobody trusts gets ignored.
Step 5: Produce the allocated view
Combine the tagged usage, the inventory and the rules into one view that shows cost by team, by use case and by model for a given period. This is the artefact you take into budget conversations. It lets a CIO see which use cases cost the most, which are pilots still spending with no value attached, and which teams are driving growth in the bill. Build it so it refreshes on a regular cadence rather than as a one-off, because the value of allocation is in the trend, not a single snapshot.
Step 6: Attach value and act on it
A cost allocation model is only half the picture until you put a value line next to each use case. For every material use case, record what it is worth in hours saved, revenue influenced or risk reduced, and set that beside its allocated cost. Now the model drives decisions: pause or renegotiate the use cases that cost real money with no return, and fund the ones that earn it. This is where allocation stops being an accounting exercise and becomes the discipline that decides which AI work deserves to scale into production.
Step 7: Make it a governed, repeating process
Give the model an owner, usually within the platform or FinOps function, and a cadence, usually monthly, aligned to the budget cycle. Set budgets and alerts on the use cases running at real volume so overspend is caught in hours rather than discovered at quarter end. Route usage through shared services with a single audit trail so the allocation is built on accounted-for calls rather than reconstructed from provider bills. Done this way, cost allocation becomes part of how the organisation governs AI, and the same visibility that controls spend is what proves, to a regulator or the board, exactly what your AI is doing.