Insights
AI Budget Planning Finance Can Defend
A $180,000 monthly AI bill is not a budget. It is an invoice waiting for an explanation. Without ownership data, finance cannot tell whether the cost came from a customer support agent, an engineering copilot, a marketing workflow, or a forgotten proof of concept still running in production.
AI budget planning starts by turning that invoice into accountable business spend. The goal is not to put a hard cap on experimentation or force every token into a spreadsheet. It is to give finance, IT, and business leaders a shared view of who is spending, what they are getting, and what should change when usage grows faster than expected.
Why AI budgets break before finance sees the problem
Most organizations begin with a simple assumption: AI spend can sit in a central software, cloud, or innovation budget. That works for a pilot. It breaks when several teams use multiple model providers, cloud AI services, and gateways at once.
The first issue is variable consumption. A subscription is generally predictable. AI costs can change with model selection, prompt and response length, traffic volume, retries, batch jobs, and agent behavior. A support team may double its usage after rolling out an AI-assisted resolution workflow, while an engineering agent can generate unexpected cost through repeated tool calls.
The second issue is missing attribution. Vendor invoices often identify an account, project, or API key, but not the business owner that finance needs for planning. If a shared key serves five applications, or an AI gateway aggregates requests from dozens of agents, the invoice alone cannot support a credible allocation.
The third issue is accounting timing. Platform teams can see usage data in near real time, but finance needs monthly accruals, department-level actuals, variance explanations, and journal entries that can be reconciled to source invoices. Those are related processes, but they are not the same process.
A budget built from last month's aggregate invoice will therefore be directionally useful at best. It will not tell a controller whether a department exceeded plan, a CFO whether a product has acceptable unit economics, or a platform leader whether a model change reduced cost without creating a business problem elsewhere.
AI budget planning needs a cost model first
Before setting budget targets, define the objects that will own AI cost. For most organizations, that means at least four dimensions: legal entity or cost center, business team, application or project, and agent or workflow. A customer service agent and an internal knowledge assistant may use the same model provider, but they should not disappear into the same cost pool.
The right level of detail depends on the operating model. A company with a handful of centrally managed AI applications may budget at the application level and allocate to departments quarterly. A larger organization with many product teams may need a direct mapping from every production agent to a cost center and technical owner.
Direct attribution should be the default. If an agent has a known owner and usage is metered, assign its cost directly. Shared costs should be allocated with a documented rule, rather than left in an unowned corporate bucket.
For example, a central AI gateway might incur $24,000 in shared platform and observability charges during a month. If the gateway supports three business units, allocation by tracked request volume may be reasonable: Customer Operations generated 50% of requests, Product generated 30%, and Corporate Services generated 20%. The resulting allocation is $12,000, $7,200, and $4,800.
That rule is not automatically correct. Request volume can be misleading when one group uses a high-cost reasoning model and another uses a lower-cost model for short classifications. In that case, allocate by measured spend, tokens adjusted for model pricing, or a blended method. The important point is that the method reflects the cost driver and is applied consistently.
Separate controllable costs from shared enablement
A useful budget distinguishes between costs a business owner can influence and costs required to operate the shared capability. Model inference, embedding generation, and agent tool calls are often controllable at the application level. Gateway operations, security review, monitoring, and central platform engineering may be shared enablement costs.
Blending the two creates unproductive conversations. A department should be accountable for the demand it creates, but it should not be held responsible for a centrally chosen platform architecture it cannot change. Show both values: direct AI consumption and allocated shared cost. That preserves accountability without pretending every dollar is equally controllable.
Build the budget from operational drivers
A credible AI forecast is usually a driver-based model, not a percentage increase on the prior invoice. Start with the workflows expected to run in the planning period, then estimate the variables that create cost.
For a customer support agent, the drivers may be monthly eligible tickets, adoption rate, average model calls per ticket, average input and output volume, model mix, and escalation rate. For an internal coding agent, the drivers may be active developers, requests per developer, average context size, and the share of requests routed to premium models.
Consider a support workflow expected to process 200,000 eligible tickets per month. If 40% are routed through the agent, each interaction averages 1.8 model calls, and the blended cost per resolved interaction is $0.11, direct inference spend is approximately $15,840 per month. Add expected tool-call, retrieval, and shared-platform allocations separately. If adoption is forecast to rise to 60% in the second quarter, the budget should reflect that planned operating change rather than classify it as a surprise variance.
Use scenarios where uncertainty is material. A base case can assume planned adoption and current model mix. A growth case can reflect higher transaction volume or broader rollout. A control case can model actions such as routing simpler tasks to a lower-cost model, reducing unnecessary context, or applying usage limits to nonproduction environments.
This does not require false precision. Estimates should carry assumptions, ranges, and named owners. A forecast that says $48,000 to $62,000 based on adoption and model-routing assumptions is more useful than a single $55,000 figure with no explanation.
Turn the plan into monthly financial control
Budgeting fails when it lives only in an annual planning workbook. AI usage needs an operating cadence because the underlying consumption can move quickly.
At minimum, finance and technology should review monthly actuals against budget by cost center, application, and major agent. The review should distinguish price variance, volume variance, and mix variance. Price variance occurs when provider rates or discounts change. Volume variance occurs when users, requests, or workflows exceed plan. Mix variance occurs when more work shifts to a higher-cost model, longer context window, or more agent steps.
Those distinctions lead to different actions. Volume above plan may be good news if it reflects productive adoption. Mix variance may point to a routing decision that needs review. An unexplained spike tied to retries or an abandoned test environment is a control issue, not an adoption success.
Set thresholds that match the scale and risk of each workload. A production customer-facing agent may require a weekly review once it reaches 85% of its monthly budget. A sandbox may have a lower spend ceiling and a shorter approval path. Hard caps are appropriate for some environments, but they can interrupt revenue-generating workflows if used indiscriminately. Alerts, approval gates, and fallback routing often provide a better balance.
Make the accounting trail explicit
Finance also needs a repeatable path from usage data to the general ledger. Suppose the company receives a $100,000 consolidated AI invoice. Metered usage and approved allocation rules identify $42,000 for Product, $31,000 for Customer Operations, $17,000 for Engineering, and $10,000 for shared corporate functions.
The monthly entry may debit the relevant departmental AI expense accounts for those amounts and credit accounts payable or accrued expenses for $100,000, depending on invoice timing. If the invoice arrives after month-end, an accrual can recognize the estimated consumption in the correct period, followed by a true-up when the vendor invoice is finalized.
The exact account structure depends on the chart of accounts and materiality policy. What should not vary is the evidence: source usage, allocation logic, owner mappings, approval history, and reconciliation to vendor charges. That evidence matters during close and becomes essential when a business leader asks why their cost center is carrying a charge.
Treat unit economics as the decision layer
An AI budget becomes strategic when it is connected to a business unit. Cost per automated ticket, cost per generated sales proposal, cost per active developer, or cost per claim processed can show whether spending is rising for a productive reason.
These measures need context. The cheapest model is not always the best economic choice if it lowers resolution quality and creates rework. Likewise, a higher-cost workflow may be justified when it shortens cycle time, improves conversion, or reduces outside service costs. Finance should ask for the operational outcome alongside the usage metric, not demand that every agent meet the same cost-per-request target.
The practical test is simple: can the business owner explain the cost driver, the budget owner, and the expected outcome in the same meeting? If not, the organization does not yet have a budget for AI. It has a growing bill. Start with ownership and evidence, then let the budget become a tool for scaling the work that earns its place.
Not sure which of your AI costs are being booked? Run the free Agent Spend Assessment.
Onaro Meridian is FinOps for agentic AI: the system of record that attributes, controls and books what AI agents spend.

Brian Diamond
Brian Diamond is a fractional Chief AI Officer and founder of Onaro. He has spent 30 years running infrastructure operations and founded LANStatus, a Connecticut managed services provider and Microsoft partner, in 2001. He holds a Chief AI Officer certification and writes the CAIO Brief on AI leadership for finance and operations.