Stop measuring AI cost. Start measuring AI value.
Image: Depositphotos
The AI dashboard shows you the exact cost of a chatbot interaction, but not the business value of that interaction. That gap is about to become more than a budget problem. It is becoming a competitiveness problem, and most businesses have no pipeline in place to close it.
The FinOps Foundation’s 2026 State of FinOps survey captured the issue in one practitioner’s quote: “Is your AI providing value? No one can answer that question yet.” The survey spanned 1,192 respondents and more than $83 billion in annual cloud spending. Forrester predicts enterprises will defer 25% of planned 2026 AI spend to 2027 as CFOs demand measurable returns that many organizations can't produce. The most important metric, cost per valuable outcome, is absent from standard FinOps dashboards today.
An April 9 article on CIO.com highlighted the most obvious symptom. Users have been tricking customer service chatbots into running off-topic code generation tasks. Tokens burn. Costs climb. None of it shows up in any anomaly report, because every session looks like a customer conversation. In that article, Greyhound Research’s Chief Analyst, Sanchit Vir Gogia, captured the prevailing sentiment: “Dashboards show what happened, but not whether it should have happened.”
The observation goes further than just chatbot freeloading. As a contributor to the Coalition for Secure AI, I’ve seen similar behavior in every multi-agent AI deployment I’ve advised on. Businesses using a large number of agents are very strict about tracking tokens, latencies, and errors. However, most can’t tell you what percentage of those sessions led to a sale, a ticket being resolved, or any other actionable outcome downstream. There is usage telemetry. There is no value telemetry.
CIOs who experienced the mid-2000s will remember this. Back in 2005, enterprises monitored page views but ignored conversions. It took a long time to plug in tracking pixels, funnel analytics, and multi-touch attribution models into the standard e-commerce stacks. Today, all serious retailers operate with attribution. Enterprise AI is in the 2005 moment.
Figure 1. Three data streams that already exist but were never connected.
Currently, three data streams have no integrations between them.
The first data stream consists of token-level session data, which shows input/output, reasoning tokens, model version, session duration, etc. While most enterprises have some version of this, few have it at the per-session level.
The second data stream includes metadata about intention. Without this layer, in the cost data, all sessions would look the same. This data stream further differentiates every session by type of intent (e.g., “product inquiry,” “support issue,” “account management,” “other”) and by outcome. For example, two sessions can burn the same tokens and look identical on a cost dashboard. Tag them by intent and they separate: one was a password reset that should be cheap, the other a product inquiry that ended in a purchase. Same cost, opposite value. Intent is what lets you tell a bargain from a waste.
The third stream includes business outcomes data from CRM systems. This stream captures purchasing data, ticket resolution, conversion data, and any downstream value signals. This data exists. It has never been connected to chatbot session data because no integration was ever designed for it.
Once all three streams of data have been integrated, the CFO gets what they want: cost per valuable outcome.
Three phases, each delivering independent value
The pipeline allows for incremental rollouts.
Each phase delivers value on its own, so the pipeline can roll out incrementally.
Phase one entails monitoring session-level tokens, including intent classification. This means recording tokens spent during each session, and categorizing them into four or five groups. One of these groups is “other,” which serves as the parasitic traffic proxy. The cost conversation shifts when a CIO sees 8% of sessions in the “other” group consuming 22% of total tokens. There is no need for CRM integration in phase one.
Phase two is outcome tagging. This includes integrating chatbot interactions with individual customers and their associated CRM events within a particular timeframe. For customer service bots, label “resolved,” “escalated,” and “abandoned.” For sales bots, tags are either “purchase within 24 hours” or “no purchase.” This will create metrics such as cost-per-resolution and cost-per-conversion. These will be the metrics the CFO will use to defend the costs associated with AI.
Phase three consists of portfolio-level attribution modeling. AI attribution is where digital marketing attribution was in 2010. The importance is clear, but there is no consistent approach. Recent work, such as the tokenomics paper at MSR 2026, has begun formalizing where tokens are spent inside multi-agent systems. The first teams to pair that research with business outcomes will enjoy a lasting competitive edge.
The organizational block
From a technical perspective, the pipeline is indeed buildable. The larger issue, however, is organizational. The data spans three teams that infrequently collaborate on common infrastructure. Session data is owned by AI engineering. The CRM is owned by marketing and customer success. Cost data is owned by finance. None of them own the join.
Developing the pipeline requires a unified schema, a unified timeline, and a unified definition of value. Alignment of that sort does not happen on its own. It will need a top down directive from the CIO or CFO, and that is before the next AI budget review, not after.
The AI budget discussions for next year will give the most value to teams that can demonstrate cost-per-resolution, cost-per-conversion, and cost-per-retained-customer metrics. The teams that will be left behind showing only token burn will be defending the spend from the previous quarter. The point was never to spend less on AI. It is to prove where AI pays back, and put more behind it.
Nik Kale
Nik Kale is a Principal Engineer at Cisco Systems and a CoSAI Contributor. He is also an ISSIP Ambassador, working at the intersection of enterprise platforms, AI systems, and service innovation.