How a FinOps or finance team gets AI and LLM spend under control: attribute every dollar to a team across OpenAI, Anthropic, Azure OpenAI, Bedrock, and Vertex, set budgets that alert before they blow, catch anomalies before the invoice, and export defensible numbers for chargeback.
A FinOps team gets AI spend under control the same way it got cloud under control — inform, optimize, operate — but the inputs are different: cost is metered per token by software loops rather than per provisioned hour, it can spike in hours rather than weeks, and it arrives on several provider invoices that share no common denominator. The practical sequence is: get one normalized cross-provider cost plane, attribute every event to a team at ingestion, set budgets with deduplicated alerts, add anomaly detection against a trailing baseline for the failure modes budgets miss, and only then negotiate commitments. Quordo is the plane that does the first four.
Who this is for: FinOps practitioners, finance business partners, and the platform leads who share accountability for the AI bill.
Key takeaways
AI spend needs the FinOps loop — inform, optimize, operate — applied to a cost that is metered per token, generated by software, and split across provider invoices that share no common denominator.
Attribution has to happen at ingestion, with provider, model, and cost denormalized onto each usage event, or historical spend shifts under you when a key is rotated or a price changes.
Budgets catch the slow overrun; anomaly detection against a trailing baseline catches the fast one — a retry loop or a model swap can multiply unit cost while staying under the monthly limit.
Chargeback is only defensible when the mapping from usage to cost center is recorded and auditable, which is why showback usually comes first.
The questions this role asks
How do I track AI spend by team when the invoice is one number?
You need a mapping layer between raw provider usage and the team that incurred it, applied when the usage is recorded rather than reconstructed later. The two workable mappings are per-connection (each team gets its own provider key or project, and the key implies the team) and per-request tagging (the caller stamps a team on every request). Per-connection is the one most organizations can actually deploy, because it needs no change to application code. Whichever you pick, store the cost on the event at ingestion so a later price change or key rotation cannot rewrite last quarter.
Start with showback. Showback publishes what each team spent without moving money, which surfaces the attribution errors — misassigned keys, shared service accounts, untagged workloads — while they are still cheap to fix. Chargeback moves the cost onto the team's budget and only works once the mapping is trusted and auditable, because the first disputed invoice will be about the mapping, not the amount. Most organizations run showback for a quarter or two before flipping.
Two different controls, because they catch different failures. Budgets with deduplicated threshold alerts catch the slow overrun — a team trending 30% above plan — and they alert once when a threshold is first crossed, then again only on escalation, so the channel stays worth reading. Anomaly detection catches the fast one: a day's spend far above its trailing baseline, which is what a retry loop, a misconfigured agent, or a quiet model swap looks like. A runaway loop can burn a month's budget in a day while never tripping a monthly limit until it is already too late.
Can I forecast AI spend when usage is this volatile?
Yes, but not with a single trend line. Forecast per team and per workload against a trailing baseline, separate the steady component (production inference at a fairly stable request rate) from the bursty component (experimentation, evals, agent runs), and treat model-price changes as a discrete input rather than something the trend absorbs. Forecasting on the blended org-wide total is what produces forecasts that are wrong in both directions at once.
Should we build AI cost tracking in-house or buy it?
The honest test is whether cross-provider normalization is a differentiator for you. Ingesting one provider's usage API is a week; keeping five of them normalized as their schemas, price tables, and billing granularity drift is the ongoing cost, and it is the part that gets abandoned. Build if you have one provider and an existing data platform to land it in. Buy if you have several, or if the numbers need to be defensible to finance rather than indicative to engineering.
Five different products get marketed as AI cost management, and they answer different questions: native provider dashboards (what did we spend with this vendor), cloud cost platforms (what did our cloud account bill), LLM gateways (enforce routing and limits on traffic routed through them), LLM observability platforms (why is this chain expensive), and cross-vendor cost planes (what did this team spend across every provider, against budget). The clarifying test is to ask what percentage of the organization's AI spend a tool would see on day one, before any migration or instrumentation work.
Connect OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, and Google Vertex. Quordo normalizes their usage into one model with provider, model, team, and dollar-cost on every event, so a single dashboard slices spend the same way regardless of who billed it.
Per-team attribution recorded at ingestion
Assign each provider connection to a team; Quordo stamps that team onto every usage event and stores the cost denormalized at the time of use. Rotate a key or watch a price change and last quarter's numbers stay exactly as they were reported.
Budgets with deduplicated threshold alerts
Set a monthly limit org-wide or per team, with on-track, near-limit, and over states. Alerts fire once when a threshold is first crossed and again only on escalation. Budgets can be set to hard-stop, which pauses the scope's provider connections when the limit is breached.
Anomaly detection against a trailing baseline
A scheduled sweep compares each day's spend against its trailing average per team, model, and provider, and raises the spikes that budgets miss — by email, ops webhook, or Slack.
CSV export and an audit log of every mutation
Export attributed spend for chargeback, and answer 'who changed this allocation rule, and when' from an append-only audit log rather than from memory.
Reading path
What is AI FinOps? — Start here for the discipline and how it differs from cloud FinOps.
AI FinOps is the practice of bringing financial accountability to AI and LLM spend: making cross-provider cost visible, attributing it to the teams and workloads that drive it, and operating budgets, alerts, and anomaly detection against it. It applies the established FinOps loop — inform, optimize, operate — to a category of spend that is metered per token, generated by software in real time, and split across providers whose dashboards share no denominator.
Which AI providers can Quordo attribute cost for?
All five it connects: OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, and Google Vertex. Per-team cost attribution is live for each, and every usage event carries provider, model, team, and dollar-cost so the same slice works across all of them.
Do I need to change application code to attribute AI costs?
No, if you attribute per connection. Give each team its own provider key or project and assign that connection to the team in Quordo — the key implies the team, and no calling code changes. Per-request tagging gives finer granularity within a team but does require the caller to stamp each request.
How is this different from my provider's usage dashboard?
A provider dashboard reports one vendor, in that vendor's denominator, for the whole account. It cannot show you what a team spent across five vendors, cannot alert on a cross-provider budget, and cannot tell you what one piece of multi-agent work cost. A cross-vendor plane exists to answer exactly those three questions.
Can I export AI spend for chargeback?
Yes. Attributed spend exports to CSV with the team, provider, model, and period on each row, and every allocation-rule and budget change is recorded in an append-only audit log — which is what makes a chargeback number defensible when a cost-center owner disputes it.