At The AI Enterprise Conference 2026, leaders from New Balance, Mastercard, Omnicom, Earnest and Crum & Forster explained how they measure AI value as token-based costs climb. Their advice: tie AI to business metrics, track cost per outcome, right-size models and build shared platforms.
Key takeaways
- AI serves the business strategy. Rege said teams should track existing metrics like time to insight and decision quality, not impact per prompt.
- Measure cost per outcome. One prompt can trigger large compute bills, so Rege's teams count cloud, AI and support costs against each outcome.
- Not all tokens are equal. New Balance sorts usage into efficient, inefficient, unproductive and necessary tokens.
- Build platforms, not one-off agents. Shared governance, semantic and orchestration layers keep agents consistent across business units.
- Trust decides adoption. Continuous evaluation, human review for sensitive outputs and change management keep people using AI.
As usage-based AI pricing replaces flat subscriptions, enterprise leaders are under pressure to show what their AI spending returns. At The AI Enterprise Conference 2026, held September 1 in New York City, a panel moderated by Joshua Ainsley, Senior Director and Head of Data Science at New Balance, took on that question directly.
He was joined by Pushkar Rege, Vice President and Co-Founder of the Business Intelligence Program at Mastercard; Rachna Narem, EVP at Omnicom; Justin Burke, VP of Information Systems at Earnest, a packaging distributor; and Anuj Garg, AVP and Enterprise Architect at Crum & Forster. Ainsley framed the stakes early: venture capital money "isn't going to last forever," and AI vendors will eventually need to be profitable.
How should enterprises measure the value of AI?
The panel agreed that counting AI use cases is no longer enough. Narem said Omnicom builds AI into measurement frameworks for clients from the start, so teams can tell whether a tool is improving results, doing nothing, or hurting them through added token cost.
Rege argued against treating AI as its own goal. "There shouldn't be an AI strategy just for the sake of AI strategy. It should be a business strategy and AI is sort of enabling it," he said. Instead of measuring impact per prompt, teams should track existing product metrics, such as time to insight or decision quality, and see whether AI improves them.
Garg described the key metric as time to a quality decision, such as faster insights for underwriters and claims adjusters. At Earnest, Burke said the goal is removing administrative work from unforgiving SAP workflows. The company is rolling out an accounts payable agent to handle two to three times as many invoices without adding headcount as it grows, including into Mexico.
How do companies control rising AI costs?
Narem said Omnicom used AI for nearly everything at first, until costs multiplied and the CFO pushed back. Teams then started asking which work actually needs a sophisticated solution and which tasks are one-off jobs better done by hand.
Rege said AI changes the cost model of software. In older products, cost tracked roughly with time spent in the platform. Now a single prompt can trigger enormous back-end compute, and "a wrong prompt could have you wake up the next morning with a really large bill," he said. He has his teams focus on cost per outcome, counting cloud compute, AI fees and support, and then work backward. That leads to choices like right-sizing models. "You don't need premium models for everything that you're doing," he said.
At New Balance, Ainsley said, his team sorts token usage into categories: efficient tokens that produce a good outcome from the right prompt, inefficient tokens that take several iterations, unproductive tokens with no good result, and necessary tokens for work like documentation that doesn't drive ROI but keeps the business running.
Garg said Crum & Forster tracks the cost per transaction of every agentic workflow as a business KPI tied to the company's expense ratio. Design matters too, he said, including giving a model only the context it needs.
Why build AI platforms instead of individual agents?
Scaling across regions and business units exposes the same problems again and again, Rege said. If each business unit builds its own solution, customers get a disjointed experience, like talking to a different assistant in every product.
He described a platform with a shared governance layer and semantic layer, centralized guardrails, and one orchestration layer that ties business-unit agents together. Narem said Omnicom is building agents that coordinate other agents across regions for global clients, while trying to avoid "a lot of AI agents flying around."
What does generative AI change about AI governance?
Ainsley said much of AI is still the same math, but the output is now text, and he often can't judge whether a legal or insurance document is good. Rege pointed to non-deterministic outputs, bias and hallucinations as new dimensions of governance.
Burke described a sales rep who built a tool outside the company network with nightly sales reports. He welcomes the creativity, he said, but told employees, in effect, "here's the six-lane highway that I want you to play on."
Garg called for continuous evaluation, since new models arrive every quarter and responses can drift from a trusted version unnoticed. Rege expects a mix of live LLM-as-a-judge checks and after-the-fact analysis of telemetry. Ainsley said New Balance weighs the risk of an error and the cost of human review, keeping people in the loop for sensitive content like legal documents.
How do leaders choose which AI projects to fund?
Ainsley said his team runs simulations when a problem is well understood. When it isn't, they size the problem and estimate what share AI could address, then track which assumptions held. Narem noted that the qualitative side of ROI, such as why a team still needs people after automation, is harder to pin down.
Burke said Earnest asks whether a project advances the company's mission and differentiation, rather than "automating just for the sake of automating." Garg said Crum & Forster favors capabilities that several business units can share.
In a closing lightning round, Narem urged leaders to "measure the value more than the technology," Rege said governance should make "the right way to get something done" the easiest way, and Burke stressed change management. Ainsley summed it up as trust. "If people don't trust it and don't see the value, they're not going to use it," he said.
Hear more on AI ROI at The AI Enterprise Conference 2027
The AI Enterprise Conference returns to Pier Sixty in New York City on Tuesday, August 31, 2027, with more Fortune 500 leaders sharing how they take AI from pilots to production. Data Science Connect's free webinars continue the conversation online between conferences.
Frequently asked questions
How do you measure the ROI of enterprise AI?
The panelists said to tie AI to existing business metrics such as time to a quality decision, run simulations or size the problem before investing, and track which assumptions hold after launch.
What is cost per outcome in AI?
Rege described it as the full cost of achieving a business outcome, including cloud compute, AI fees and support, which teams then optimize by choices such as using smaller models.
How can companies keep AI token costs under control?
Panelists said to reserve sophisticated solutions for work that scales, right-size models, give models only the context they need, and track cost per transaction for each agentic workflow.
How do enterprises monitor generative AI quality in production?
Panelists described continuous evaluation as models change, LLM-as-a-judge checks, telemetry analysis, and human review where errors carry high risk, such as legal documents.
Watch: Scaling AI That Works: Proving Value, Managing Cost, and Driving Impact on YouTube
Event: The AI Enterprise Conference · More coverage: The AI Enterprise Conference insights
