AI token consumption is becoming a significant line item on monthly budgets. Tokens are the basic unit AI models process—every prompt, response and agent action consumes tokens, each carrying a small cost. As AI usage grows, so does the token bill, and “tokenomics” is the practice of keeping that cost visible, predictable and measurable.
Globally, enterprises spent an estimated $2.5 billion on AI tokens last year, a figure projected to reach $3.6 billion within 24 months with current usage trends. While token prices have been falling as technology matures and AI labs compete for market share, this has not translated to smaller bills. New Accenture research reveals a paradox – when prices drop, organizations use AI more – leading to higher bills.
When asked what they’d do if token prices fell another 25 percent, only 5 percent of executives in our survey said they’d bank the savings, while the rest claimed they’d expand usage. Token volume growth is currently outpacing price declines — surveyed companies expect token consumption to grow 78 percent and price to decline by 19 percent over the next two years. As organizations move from pilots to embedding AI into more complex workflows, this trend is likely to accelerate.
There’s a second challenge equally central to tokenomics: AI’s value is still hard to quantify. Fewer than one dollar in five spend on tokens can currently be traced to a measurable financial outcome such as influenced revenue or increased productivity. The rest is investment that organizations believe is worthwhile but haven’t yet proven. Moreover, token costs account for just 25 percent of the total AI spend – while the bulk of the investments remain in infrastructure, maintenance costs, cloud hosting, talent training and operational changes necessary to use the technology at scale.
Anticipating the Challenge
As AI adoption increases across government agencies, this consideration matters in a particular way: taxpayer money is funding the tokens. Our survey of public service workers found that 70 percent of agencies now describe themselves as advanced in implementing AI technologies, and 77 percent of workers say the everyday tools they use have changed measurably in just the past two years—a clear signal that AI is moving from pilot to routine use.
That growth in use is a genuine opportunity: nearly two-thirds of public service workers expect AI to meaningfully reduce their workload by automating repetitive tasks, freeing time for more complex, judgment-driven cases. Moreover, 63 percent are optimistic that AI will help improve customer service experiences.
How agencies pay for AI differs from the private sector—they often access them via vendors for pilot projects, use government-wide licenses tracked per department, and secure significant discounts. However, the tokens still have a cost, and monitoring how value is tracking against that expense will become increasingly important.
Between January and July of 2026, the share of public service organizations reporting widespread, sustained business value from AI declined from 29 percent to 19 percent, while the share reporting moderate value in specific areas rose from 21 percent to 31 percent. Usage and implementation are increasing but confidence in seeing organization-wide return is not keeping pace.
How to See, Control and Account for AI Spend
This is a planning opportunity, not an inevitable problem. Agencies that build the right foundations now can scale AI with confidence rather than retrofitting governance later.
- Make AI work visible. Know which workload is running, on which model, at what cost. Define in advance what it should improve and how it will be measured against clear KPIs such as process time reduction or resident satisfaction improvement.
- Build-in financial accountability. Give each workload an owner responsible for its cost and outcomes. Tying spend to a person or team, rather than absorbing it anonymously, makes it far easier to explain where taxpayer-funded token spend is going.
- Route tasks to the right model. Not every interaction needs the most advanced model available. Routine tasks can be served by lower-cost models, while complex or sensitive cases require stronger models; always supervised by skilled staff that remain in the lead.
- Match deployment to your needs. Open-source models often carry a lower cost per token, but that saving shows up in the infrastructure, hosting and specialist skills needed to run them safely. Closed-source models cost more, but up-front investments are handled by the provider. Balance capability and cost depending on an organization’s needs.
- Build AI skills across the workforce. Give employees practical guidance and train them to use AI tools confidently and responsibly. Rather than mandating one training path, let employees choose how they build AI skills based on what works for them.
Government use of AI is set to keep growing, and that’s a positive development for service delivery. Agencies that get ahead of the cost-and-value equation now will be best placed to make the case for continued investment with confidence.
The author, Prasanna Ellant, is Global Public Sector Digital Core (Civilian) Lead at Accenture.