Customer service leaders have spent the past several years pursuing an ambitious vision for AI: faster resolutions, more productive agents and end-to-end automation of customer issues.

The investment behind that vision is substantial.

Gartner’s 2026 C-Suite AI Survey of more than 1,400, which included responses from 95 customer service leaders across various industries and regions, found customer service and support organizations allocated an average of 12.7% of their functional budgets to AI initiatives, with median spending approaching $6 million in 2025.

Yet the returns remain decidedly mixed. Across the customer service AI use cases evaluated by Gartner, about one-quarter produced positive returns, about one-quarter produced negative returns, and 11% broke even

Most concerning of all, however, is that 42% of leaders did not know how much they spent, their return, or both. Even so, 78% of service leaders planned to increase AI investment in 2026 — a disconnect that should give executives pause.

Customer service organizations are expanding their AI commitments before many have established a clear return on their existing investments. At the same time, they may be building financial plans on an even riskier assumption: that AI will inevitably become cheaper. That assumption is worth challenging.

Today’s prices may not last

The current price of many AI services does not necessarily reflect the full cost of providing them. AI vendors have subsidized their offerings to encourage adoption and capture market share, separating what customers pay from what the technology costs to operate.

Those economics are unlikely to persist indefinitely. Providers will eventually need to recover more of their costs, whether through direct price increases, revised rate structures or new revenue streams. Even when vendors announce lower model prices, changes elsewhere in the commercial model can increase a customer’s total bill.

Recent pricing changes illustrate the risk. In June, GitHub Copilot moved from a largely flat per-user model to one combining access and consumption. The change increased the average Gartner client’s monthly Copilot bill by three to 10 times. The following month, OpenAI shifted ChatGPT Enterprise from a per-user fee with pooled credits to a usage-based model, creating greater allocation flexibility but also exposing customers to potentially significant overage charges.

The lesson for service leaders is straightforward: Improvements in model efficiency do not guarantee reductions in enterprise AI spending.

That distinction matters because AI business cases are often based on current prices. If those prices rise while adoption and usage are also increasing, AI spending could outpace anticipated productivity gains or labor savings. Organizations may then be forced to scale back, restart or abandon initiatives they can no longer afford.

Make generative AI the exception

Preparing for this possibility begins with a more disciplined automation strategy.

Generative and agentic AI should not automatically be the first choice for every workflow. Before automating an issue, leaders should ask whether they can eliminate its root cause. If automation is still necessary, they should determine whether conventional AI, robotic process automation, business process automation or another deterministic technology can deliver the required result more economically.

Generative AI should be reserved for work in which its additional flexibility or reasoning capability creates enough value to justify the additional cost and risk.

This approach leads naturally to hybrid workflows. A single customer journey may combine traditional automation for predictable steps, a rules-based script for routine interpretation and generative AI for tasks requiring complex reasoning.

Service organizations should apply the same discipline when selecting models. Rather than defaulting to the highest-performing option, they can use gateways and complexity thresholds to route straightforward tasks to faster, less expensive models. More capable models can then be reserved for the interactions that genuinely require them.

Identifying what a customer wants to accomplish, for example, does not require a frontier model. A complicated request that depends on several data validations might. The goal is not to minimize AI use indiscriminately. It is to use the least expensive capability that can reliably produce an acceptable result.

A hybrid automation strategy introduces architectural complexity, but it can also reduce the risk that every automated interaction carries premium-level costs.

Negotiate for predictability, not just price

A stronger technology architecture must be matched by a stronger commercial strategy.

Consumption-based pricing gives organizations flexibility, but it also makes spending rise with adoption. That can become particularly difficult in customer service, where interaction volumes fluctuate and a long or complex exchange may consume far more resources than a routine one.

Alternative pricing models can offer greater predictability. Activity-based pricing ties charges to units of work, such as tickets or interactions handled. Outcome-based pricing connects payment to an agreed business result. Hybrid arrangements can combine committed access with variable usage. Each model carries trade-offs, and outcome-based arrangements can create disputes if success is not precisely defined. Nonetheless, less-variable structures may provide better forecasting and cost control as subsidized pricing recedes.

Leaders should also look beyond the headline pricing model.

The final bill a leader receives is determined by several factors:

• The pricing model and how it interacts with other commercial commitments, like minimum spend.
• The formula used by the vendor to determine how charges accrue.
• How metrics in that formula are defined.

Two apparently similar proposals can produce very different costs once scaled across actual workloads.

Contracts should therefore include protections against uncontrolled increases. Service leaders, working with procurement and finance, should negotiate usage limits, overage forgiveness and restrictions on a vendor’s ability to change token or credit multipliers during the contract term. If a vendor retains that right, organizations should seek substantial advance notice and the ability to exit the agreement.

Treat AI cost as an operating discipline

Once AI is deployed, active cost management becomes essential and benefit attribution becomes non-negotiable.

Organizations can cache prompts for frequently requested, static information rather than generating the same response repeatedly. They can monitor token consumption to identify inefficient workflows and analyze usage patterns to find tasks that should be moved to cheaper, deterministic automation. They can also establish thresholds that escalate an interaction to a human when continuing with AI would cost too much.

Finally, service leaders should conduct portfolio-level sensitivity analyses with finance. They need to understand how changes in adoption, token consumption and vendor pricing could affect total cost and ROI. They should also know how large an increase their budgets could absorb before a use case becomes uneconomical.

The next phase of AI adoption in customer service will not be defined solely by who deploys the most advanced technology. It will be defined by who exercises the greatest economic discipline.

AI can still transform service delivery. But sustainable transformation requires leaders to challenge the assumption that better AI will always mean cheaper AI. The organizations best positioned to succeed will automate selectively, route work intelligently, negotiate for predictability and manage AI consumption as rigorously as any other major operating expense.

For customer service leaders, the question is no longer simply what AI can do. It is whether it will be worth it when the bill comes due.