As a former CFO, I see AI as a financial lever. It can reduce cycle times, expand capacity, improve decisions and lower risk. But those outcomes do not come from deploying a model. They come from sound investment choices, operating discipline and accountability for results.
The question for finance leaders is not whether AI belongs in the business; the question is how to fund it when technology, pricing and usage patterns change faster than a standard budget cycle. AI does not fit a fixed annual plan, model capability can shift within months and provider pricing can change. Adoption can accelerate after one workflow proves its value, or stall when data, controls or process design are not ready.
That uncertainty does not mean AI spend cannot be managed. It means finance needs a different approach to planning, cost control and value measurement.
Start With Cost by Design
Many organizations use the most capable model for every AI task. That choice is often expensive and unnecessary. A task such as classification, routing, document extraction or content summarization may not require a frontier model. A lower-cost model may deliver the required level of accuracy at a fraction of the cost.
The right model should be selected based on the task, the required quality threshold, response time and risk level. This is where model routing matters. A system can send simple tasks to lower-cost models and reserve higher-cost models for work that requires reasoning, judgment or complex analysis. In high-volume workflows, this approach can reduce inference spend by 60 percent or more when the starting point is overuse of premium models.
Other cost controls also matter:
- Caching: Reusing responses for repeated questions and standard requests can reduce duplicate compute
- Prompt design: Clear prompts reduce unnecessary tokens and improve response consistency
- Batch processing: Reporting, document review and other non-time-sensitive work can often run in batches at lower cost
- Usage limits: Spending caps, role-based access and workflow controls help prevent uncontrolled consumption
- Quality monitoring: A lower-cost model has no value if it creates more rework, errors or review effort
Model cost is visible, but it is not the whole cost base. In many early AI programs, the largest investments sit elsewhere: preparation, integration, security, governance, testing, process redesign and employee training. Finance should measure the total cost of ownership, not just token spend.
Forecast Through Stages
AI investment should be forecast in stages because the cost profile changes as adoption grows.
Phase 1: Build the Foundation
The first phase includes platform selection, data access, governance, controls, integration and pilots. Inference costs are often low at this point. The main expense is building the capability needed to use AI in business processes. This phase should have clear limits. Each pilot needs a defined use case, a business owner, a baseline metric and a decision point.
Finance should ask:
- What problem are we solving?
- What process measure will change?
- What controls are required?
- What will it cost to operate at scale?
- What result would justify further investment?
Phase 2: Scale Proven Use Cases
Once use cases move into production, usage rises. Inference spend can grow fast, especially in customer service, sales operations, finance operations, HR and knowledge management. At the same time, unit economics can improve. Model routing, caching, prompt improvement and process redesign can lower cost per task.
Volume can also support better vendor terms. Forecasting should use scenarios rather than a single annual number. Finance should model a conservative case, a base case and a growth case. Each should include assumptions for users, transaction volume, model mix, human review, vendor pricing and expected business impact. The key metric is not total AI spend, it is cost per business outcome.
Examples include:
- Cost per invoice processed
- Cost per report produced
- Cost per sales lead qualified
- Cost per hour of analyst capacity released
Phase 3: Manage AI as an Operating Capability
At maturity, AI is no longer a series of pilots. It is part of how the company operates. The discussion shifts from model cost to business outcomes. What does it cost to reduce a process from five days to one? What is the value of fewer errors? How much capacity can be redeployed? What risk can be prevented or identified sooner? This is where finance should allocate more funding to workflows with proven results and stop funding work that does not produce value. The purpose of governance is not to prevent experimentation. It is to ensure the company learns at a controlled cost and invests more where evidence supports the decision.
Measure Outcomes, Not Activity
Tokens, prompts and model calls are operational measures. They help manage consumption, but they do not measure value. A 50-cent inference that saves two hours of analyst time may create strong returns. A no-cost tool that produces poor output, adds review work or creates compliance risk has no economic value. For each use case, finance should require three things:
- A pre-AI baseline: What does the process cost today? How long does it take? What is the error rate? How much capacity does it consume?
- A target outcome: What improvement should the AI-enabled process deliver?
- A named owner: Who is accountable for adoption, process change and realized value?
Finance should not assign every improvement to AI. A process may improve because of training, workflow changes, staffing or policy changes. Value cases should separate projected value from realized value and document the assumptions behind both.
Preserve Flexibility, Then Earn Commitment
In the early stages, organizations should avoid locking into large commitments before demand and workloads are understood. Consumption pricing, short contract periods and portable architecture can preserve options. Once a workload is stable and demand is predictable, committed spending or volume agreements may improve unit economics. Companies should earn the right to commit through evidence, not enthusiasm.
The companies that succeed with AI will not be defined by how much they spend. They will be defined by how well they connect spending to business results. AI will show up on the P&L. Its returns should as well. The CFO’s role is to ensure that each investment has a purpose, a baseline, an owner and a path to measurable value.
