As we move from experimentation to pilots to artificial intelligence (AI) in production, fear of messing up (FOMU) is still a big issue. Although leaders often focus on the risk of AI doing something deleterious that it shouldn't, an equally big concern is whether AI delivers a positive return on investment. There are two sides to the ROI story: picking the AI projects that deliver the most value and ensuring those projects deliver quality outcomes at the lowest possible cost.
Every vendor seems to have a laundry list of new agents and AI features to try, with new ones being released at a rapid pace. You can use Valoir's free CX AI Value Coach (based on our ROI methodology and data) to compare agents and identify the potential projects likely to deliver the greatest returns and least risk.
On the benefit side, with vendor pricing models continuing to evolve, fluctuating token costs, and changing model quality, it can be hard to determine how much you're paying and how much you should be. If you buy AI through a contact-center-as-a-service (CCaaS) or customer relationship management (CRM) software vendor, you'll want to know how they're metering so you know what you can control. Some charge per-agent fees, some charge by resolution, and some charge by action or intent. Insist on spending caps and alerts when you're reaching certain thresholds, an agreed-upon definition of resolution (if it's metered that way), and the ability to move committed budget between seats and different AI workloads.
You can do several practical operational things to optimize your AI token economics (whether you buy them through another vendor or not). They include the following:
- Don't use a large language model (LLM) call when a deterministic workflow will do. In the early days of agentic AI, many agents reasoned through things that could have been solved without reasoning, and calling LLMs will always be more expensive and less efficient than simple workflows.
- Set limits and escalate early. An AI agent that loops through 15 turns before handing off to a human is still burning tokens without value (even if you’re only paying on resolution) and frustrating customers. Set maximum turns and LLM calls and trigger escalation on frustration signals. Always give the customer the option to escalate easily.
- Clean up your knowledge base. Bloated, duplicated, or outdated articles give the model more content and context through which to reason, slow response times, and lower answer quality. Shorter, current articles cost less to retrieve and are more likely to deliver accurate responses.
- Fix the workflows that drive contact volume. Fixing billing errors, operational bottlenecks, and confusing policies that drive repeat customer contacts lower spend no matter which pricing model you're on for AI.
You or your vendor can and should also do several things to optimize AI ROI from a token cost perspective. They include the following:
- Optimize routing to the most appropriate model. Vendors should help you understand and match the right AI workload to the most effective and cost-effective model, not just the expensive frontier models. Small models cost less and are just as good or better than large ones at certain tasks; open-weight models are obviously a great option from a cost perspective.
- Avoid rerunning the same reasoning. Take advantage of prompt caching to cache the parts of the prompt that don't change and leverage notes from existing reasoning turns from similar requests with a positive resolution to reduce the reasoning burden and cost.
- Track cost per resolution, not cost per token. You should have dashboards that show you not just token spend and cost but how many resolutions are being delivered and at what cost.
Token spend is becoming the new handle time, a cost metric that, if managed in isolation, pushes teams to cut the wrong things. Taking the practical steps to ensure AI can operate most efficiently long before AI touches a process and taking steps to meter model usage and cost based on resolutions and value will help you optimize token spend and, ultimately, AI ROI.
Rebecca Wettemanmn is founder and CEO of Valoir.