The Rising AI Cost Facing Tax Departments

As AI usage grows, so does the cost. Discover a path forward by standardizing AI solutions and governance and measuring the cost and value of outputs.

For many organizations, early AI adoption was intentionally broad: Employees received access to generative AI and were encouraged to experiment and incorporate it into everyday work. Fast forward and today’s tax professionals now use AI for research, drafting, document analysis, data work, and complex problem-solving. As a result, the question has shifted from “How do we get people to use AI?” to “How do we scale that use economically?”

AI changes technology economics. Unlike traditional software, whose costs often are driven largely by licenses, AI adds a variable cost that rises with use. Tokens – the small units of text a model reads and generates – multiply as usage and task complexity increase, sometimes producing unexpectedly high costs.

The problem: Democratized AI can be difficult to scale economically

During early adoption, decentralized experimentation is a feature, not a problem. It helped employees learn where AI works well and where human judgment remains essential. But a model optimized for learning is not necessarily optimized for scale. Twenty employees facing similar tasks might build 20 slightly different solutions, each consuming tokens, requiring maintenance, and producing its own answer, when the organization ultimately needs one well-designed, governed solution.

AI agents can magnify this issue by reviewing information, retrieving documents, using tools, and refining work on a user’s behalf. While that capability can create value, it can consume far more tokens than a simple question-and-answer interaction, leading to increased costs for the organization. Rising AI costs do not necessarily signal failure. They can, instead, reflect rapid adoption and more use cases than anticipated. The challenge for management, therefore, is to direct consumption toward the highest-value work.

Restricting access can solve the wrong problem

One intuitive response to rising AI costs is to reduce licenses or broadly limit usage. Some financial control is appropriate, but removing access can create a different problem. Broad license reductions can interrupt investments in training, fluency, and experimentation while concentrating capability in a small group of specialists. An organization that restricts access might lower its AI costs, but it also is reducing the return on its AI investments. The better objective is not necessarily less AI. It is more intentional AI.

The solution: Standardize the work, not the access

As tax departments mature, standardization can become a powerful cost control. Organizations should identify repeated use cases and consolidate the strongest approaches into shared, governed agents or workflows.

Instead of having several tax professionals build tools to analyze the same type of notice or guidance, an organization could create one approved solution with standardized instructions, review procedures, and monitoring. It can then optimize consumption once, centralize maintenance, and reuse the solution across teams.

Governance should occur before a token is consumed. Predictable processes could be better handled with workflow tools, analytics platforms, scripts, or other established technologies. An economical design automates predictable work first and applies AI where language understanding, reasoning, or unstructured information adds value. AI then becomes a productivity multiplier for automation rather than its replacement.

Good AI governance should address economics alongside security, data, risk, and compliance. For each AI-enabled process, tax departments should be asking:

• What problem does it solve?
• How much time does it save?
• Does every step require AI?
• Is another team solving the same problem?
• Can solutions be consolidated?
• What is the AI cost per completed outcome?

The broad goal should not be to minimize tokens. A high-consumption workflow that eliminates substantial repetitive work can outperform a low-consumption workflow that adds little value. The better question is whether AI consumption produces a measurable business return.

From cost per token to value per outcome

The first chapter of enterprise AI usage democratized access, built fluency, and surfaced useful applications. The next chapter should industrialize that progress by turning successful experiments into governed, reusable solutions, applying AI only where it adds value, and making AI costs and outcomes visible. The organizations that manage this transition successfully might not have the lowest AI bills, but they will be the ones that understand which tokens are worth buying because the return justifies the cost.