AI budgeting requires forecasting demand, not just approving invoices
Budgeting for AI is more than setting a spending limit. It also means estimating how demand may change, assigning responsibility for costs, and assessing whether usage supports business goals. A model invoice is only one part of the financial picture.
Deloitte’s State of Generative AI in the Enterprise survey, conducted with The Wall Street Journal, reports on organizations’ adoption and use of generative AI. Its findings offer a view of surveyed organizations, not a universal forecast of what any one company will spend. Use the report as context for planning, not as a substitute for an organization’s own usage and cost data. Source Source
Why AI spending crosses budget owners
A model charge may begin in one team while related expenses appear elsewhere. A product group may fund a feature; infrastructure teams may carry cloud and data costs; finance may track vendor commitments; and operations may absorb the work of reviewing output quality. When each group sees only its own costs, local efficiency can be mistaken for enterprise-wide efficiency.
As a practical matter, organizations can make AI costs easier to manage by connecting each use case to its consuming team, underlying services, accountable business owner, and intended outcome. This is an operating recommendation, not a finding attributed to the survey.
Separate committed capacity and other relatively fixed costs from variable consumption. They behave differently as usage changes, and a blended budget line can hide that distinction. Show owners what could drive changes in variable costs, such as more requests, longer inputs, additional evaluations, or broader adoption.
Build forecasts around behavior, not hype
A forecast is a set of assumptions, not a promise. For each use case, document expected users, frequency, average task size, and service path. Record what is known and what remains uncertain. A team testing an internal assistant may have little usage history; a mature document workflow may have enough records to support a more informed estimate. Treat those forecasts differently.
As a practical planning approach, model a base case using measured consumption, a higher case that allows for adoption or larger workloads, and a lower case that reflects delayed rollout or substitution. Review assumptions regularly and compare forecasts with actual usage. If the gap widens, investigate whether it reflects user growth, repeated calls, larger context, changed pricing, or a changed workflow. The cause should guide the response.
Make consumption legible to finance
Total spend alone is a blunt measure. Where the data permits, track cost by application, business owner, model or provider, and workload. Pair cost with a quality or throughput measure, such as completed cases, processed documents, or resolved requests. Choose a measure that reflects the work: a low-cost answer requiring substantial human correction may not be efficient.
Usage-based charges can vary with input and output volume, while application design can affect how often models are called. Finance and product teams can work together to understand these cost drivers before setting allowances. When usage is still changing, a range with stated assumptions is more useful than a falsely precise single figure. These are practical recommendations for planning, not claims about results established by the cited survey.
Put guardrails close to the work
A central policy can define approved providers, reporting expectations, and escalation thresholds. Teams close to the workflow should help shape the controls: product owners understand which steps are essential, data teams understand what information is being sent, and finance can establish a review cadence. This is a shared operating responsibility, not a reason to transfer every decision to finance.
Treat exceptions as signals to investigate. A sudden increase may reflect successful adoption, an unbounded workflow, duplicated requests, or a change in service or usage. Examine the cause before imposing a blanket cap. If a use case is valuable, the goal is to make it sustainable rather than suppress it indiscriminately.
Environmental impacts should be assessed separately with appropriate operational data and a clearly stated measurement boundary. Financial cost estimates alone do not establish energy use or emissions.
A useful definition for the next budget cycle
For an organization using AI, budgeting means aligning expected consumption with accountable owners, business outcomes, and explicit uncertainty. It means funding work based on its value, monitoring the variable costs that may follow adoption, and revising plans as evidence changes.
As a practical starting point, name an owner for each material AI use case, document its cost drivers, and compare actual consumption with a forecast on a dependable schedule. Give teams enough visibility to act before overspending becomes a surprise. Bigger budgets may be justified, but they should follow evidence rather than be automatic.
For guidance on managing changing model charges, see Managing AI Token Spend: A Budgeting Guide for Volatile LLM Costs.
ai budgeting requires forecasting demand, not just approving