Measuring AI ROI requires knowing, for every AI request, who made it, what it cost, and what it produced. That takes identity-aware instrumentation at the request level, applied at the boundary where requests are dispatched to models, because billing-level data arrives too late and too aggregated to connect spend to outcomes.

Three things make AI ROI measurable, and most enterprises today have none of them:

1. Per-request identity. Which user, team, or agent made this call.

2. Per-request cost. Tokens consumed, model tier, and inference time, computed when the request runs.

3. Per-request outcome signal. What task the request served and whether it succeeded.

This guide explains why the ROI question is currently unanswerable in most organizations, how to build the three measurement capabilities, and how attribution turns into an optimization loop that lowers cost each cycle. It extends the cost-control mechanics covered in our guide to AI token optimization and control and pairs with our piece on governing AI at the boundary, both companions to the NexusOne AI Data Control Plane launch post.

Why AI ROI Is Unanswerable for Most Enterprises Today

The numbers on this are blunt. MIT's State of AI in Business 2025 research found roughly 95% of enterprise generative AI pilots deliver no measurable P&L impact [1]. McKinsey's global survey found 88% of organizations now use AI regularly, yet only 39% attribute any EBIT impact to it, and most of those put the impact under 5% [2]. An IBM study of 2,000 CEOs found only about a quarter of AI initiatives delivered their expected ROI, and just 29% of executives said they can measure ROI with confidence [3]. PwC's 2026 CEO survey found 56% of CEOs report neither increased revenue nor decreased costs from AI in the past year [4].

Read those figures carefully and a measurement problem hides inside the value problem. An initiative that cannot be measured reports as an initiative that failed, and most enterprises cannot measure. Gartner's CFO survey found 54% of organizations cannot accurately state their total AI spend because costs fragment across budgets [5], and an RGP survey of 200 US finance chiefs found only 14% report a clear, measurable impact from AI investments [6]. Meanwhile the organizations that do measure behave differently: Wharton's 2025 adoption study found 72% of enterprise leaders now formally track productivity and profitability metrics tied to AI usage, and those organizations report far higher ROI confidence [7]. BCG finds only about 4% of companies are creating substantial AI value, and they are distinguished by disciplined measurement and scaling practices rather than better models [8].

The mechanical root cause is humbler than any of those reports make it sound: the shared API key. When one provider key serves 50 engineers across five teams, the invoice is accurate in total and empty in detail, one number covering every team, project, and agent [9]. Attribution data that was never captured at request time cannot be reconstructed at billing time, and each provider's console only sees its own slice. Finance wants cost per outcome; the data to compute it does not exist [5][9]. Virtual keys, one governed identity per user, application, and agent, are the structural fix, and they have become a product category in their own right [10][11].

The Three Things You Need to Measure AI ROI

What is AI usage attribution? AI usage attribution is the practice of tagging every AI request with the identity that made it, the cost it incurred, and the task it served, so that spend and results can be rolled up accurately by team, agent, model, or time window. Attribution is the layer that turns a consolidated AI bill into an answerable ROI question.

Each of the three capabilities has a concrete implementation, and all three have to be captured at the same point, on the request, or they never join up again.

1. Per-request identity. Every call carries a principal: the user, the team, and, when an agent made the call, the agent's own identity plus the user it acted for. In practice this means deriving identity from your SSO provider and issuing distinct credentials per user, application, and agent rather than sharing keys [9][10]. Identity is the join key for everything else; without it, cost and outcome data aggregate into mush.

2. Per-request cost. Token counts in and out, the model that served the request, and inference time, recorded when the request completes and priced against the model's rate table. The telemetry standard for this already exists: OpenTelemetry's generative AI semantic conventions define gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, and the surrounding span structure for tracing each call [12][13], and observability platforms compute per-request cost from token counts across hundreds of models, broken down by tags for team, application, and feature [14][15]. Multi-step agent work traces as one record: an agent that made six model calls to answer one question shows the full chain with per-step cost [14].

3. Per-request outcome signal. The hardest of the three, and the one that separates cost reporting from ROI measurement. At minimum, tag each request with its task category (summarization, SQL generation, document extraction, reasoning) so cost rolls up by the kind of work performed. Better, capture task completion: agent evaluation practice has converged on metrics such as task success rate, tool-call accuracy, and step count [16][17], and academic surveys of agent benchmarks treat task success rate as the primary outcome measure [18]. One enterprise customer put the principle well in the context of Anthropic's admin analytics: token usage alone tells you little, while which skills get run again and again is the real signal of value [19]. Repeat usage of a workflow is a revealed-preference outcome signal, and it is available on every request at zero instrumentation cost.

Instrument at the Request Boundary, Not the Billing Layer

Billing-layer attribution fails for reasons that are structural rather than fixable with effort. Cloud cost tags cannot cover usage-based line items that do not accept tags, and untagged spend is the norm rather than the exception; in one documented case more than half of a $700,000 cloud bill sat unallocated before a tagging overhaul [20]. FinOps practice treats 90% cost-attribution coverage as a target worth celebrating [21], which is a fine bar for infrastructure and a hopeless one for per-team AI ROI. Billing data also arrives on billing's schedule, daily at best, and never carries request-level identity [20]. The FOCUS billing-data specification is steadily improving cross-vendor consistency, and its 2026 revisions add AI usage detail [22], but a better-formatted bill is still a bill: an after-the-fact aggregate.

Boundary instrumentation inverts every one of those properties. Because each request passes the boundary at execution time, coverage is 100% by construction, including self-hosted models that never appear on any cloud bill. Cost is computed from actual token counts rather than allocated from invoices, identity is attached before dispatch rather than inferred afterward, and the data is available immediately rather than at month-end [12][14][23]. The FinOps community has effectively conceded the point by making AI cost management its fastest-growing discipline, with 98% of organizations now managing AI spend, up from 63% a year earlier [24], and its own guidance distinguishes showback, giving teams visibility into their consumption, from chargeback, formally landing cost in team budgets [25][26]. Boundary data supports both; billing data supports neither at the granularity ROI requires.

Scale makes the gap wider every quarter. Token throughput on one major routing platform grew roughly 4 to 5x year over year, with average request length more than tripling since 2023 [27], and Goldman Sachs Research projects a roughly 24x multiplication of token consumption between 2026 and 2030 [28]. Attribution debt compounds with volume: every month of unattributed usage is a month of ROI evidence that can never be recovered.

How Attribution Works in the NexusOne Control Plane

The NexusOne AI Data Control Plane implements the three capabilities as properties of the boundary itself. Every AI request from every user and agent passes through the control plane, where it is matched to an identity, its role, and its permissions before anything else happens, so the principal is attached before dispatch rather than reconstructed after [29]. Cost is computed and recorded at the boundary from what the request actually consumed, split between cloud token spend and local execution on your own models, which keeps self-hosted inference inside the same ledger as API calls [29]. Routing decisions are logged with their reasoning, what was sent where, why, and what it cost, and because the control plane classifies each request to route it, every request carries a task category as a byproduct [29]. Underneath the dashboards sits the full record of each request: what was asked, what came back, whether it was cached, and which checks were applied. Usage reports roll up by team, agent, model, or time window, and you can build whatever leadership view your organization needs on top of the same records [29]. To see attribution running against your own model mix and teams, bring your current bill and your hardest ROI question to an expert consultation.

From Usage Data to ROI: A Practical Example

Take a 200-person engineering organization running four AI assistant deployments: a coding assistant for all engineers, a documentation assistant, a text-to-SQL assistant for the data team, and an incident-summary agent for the platform team. Here is the 30-day version of doing this properly.

Days 1 to 3: set up attribution. Each deployment gets its own identity, each engineer authenticates through SSO, and the incident agent gets an agent identity tied to the platform team. Task categories map onto the four deployments, with the coding assistant further split by request type. Budgets are set per team with alerts rather than hard caps for the first month, since the goal is baseline data [9][10].

Days 4 to 30: let it run and resist tuning. The boundary records roughly everything: which teams consume most, which models serve them, cache hit rates, and cost per task category [14][15].

At day 30 you produce the first per-team summary, and it will contain surprises. A plausible one, consistent with published usage research: the coding assistant dominates spend but also dominates repeat usage, the revealed-preference signal that the workflow earns its keep [19][30]. The SQL assistant costs little and completes a high share of tasks. The documentation assistant shows heavy spend and low repeat usage, an early sign it is not landing, worth investigating before renewal rather than after. "ROI positive" now means something specific at the team level: the data team can compare the SQL assistant's monthly cost against analyst hours reclaimed, using time-savings baselines from controlled studies such as GitHub's finding that developers completed a benchmark task 55% faster with an AI assistant [31], applied skeptically to your own numbers rather than adopted wholesale. Organization-level ROI is then a sum of team-level cases plus shared costs, not a single number conjured from the total bill. DORA's 2025 research is a useful caution here: AI adoption amplifies existing team strengths and weaknesses rather than fixing them [30], which is precisely why per-team measurement beats org-level averages.

Connecting ROI to Routing: The AI Token Optimization and Control Loop

Attribution data does not just report cost; it tells you where cost is wrong. Usage records reveal which workloads consume the most tokens for the least outcome signal, and those workloads are routing candidates: work running on a frontier model that a small model or a cheaper path could serve. The published headroom is large. Routing frameworks report cost reductions from 25% at near-full quality retention to 85% on specific benchmarks [32][33], and NVIDIA researchers argue small language models are the economically rational default for most agentic invocations [34]. Deterministic work is the extreme case: a request the control plane can answer with a query engine or an existing ML pipeline costs no tokens at all [29].

The control plane closes the loop because measurement and enforcement live at the same point. Measure a month of attributed usage, update routing policy for the worst cost-per-outcome workloads, and measure again; each cycle's savings are themselves attributed, so you can show leadership what the optimization recovered, by team and by workload [29][32]. This is the practical payoff of AI token optimization and control done at the boundary: the same records that answer the ROI question power the changes that improve it.

What to Include in an AI Usage Report for Leadership

A leadership usage report should answer four questions, in this order: what did AI cost, what did it produce, is it governed, and what changes next. A structure that works:

• Total AI spend by team, model, and time period, with deltas against the prior period and against budget, and the cloud-versus-local split if you run your own models [15][29].

• Cost per task category: summarization, SQL generation, document extraction, reasoning, and the rest, because category-level cost is what makes routing decisions legible to non-engineers [14][29].

• Outcome signals where trackable: task success rates for agents, repeat-usage rates for assistants, and team-level value cases where a baseline exists [16][19][31].

• Governance summary: policy violations, boundary enforcements, and audit events in the period, which belongs in the same report because spend and risk share the same records at the boundary [29].

• Optimization actions: which workloads move to cheaper routes next period, with the projected savings attributed to the change [32][33].

Keep it to a page, keep the definitions stable so periods are comparable, and, once trust is established, let the report drive showback and eventually chargeback [25][26].

FAQ: AI ROI and Usage Attribution

What Tools Exist to Track and Attribute AI Token Costs by Team or User in a Data Platform?

Three tool classes cover it. LLM observability platforms such as Langfuse and Datadog LLM Observability record per-request token usage and computed cost against tags for team and application, following OpenTelemetry's generative AI conventions [12][14][15]. Gateways issue virtual keys per user, team, and agent, attaching identity and budgets to every call [10][11]. Control planes such as NexusOne's attribute every request, including self-hosted model calls and cached responses, to an SSO-derived identity at the request boundary and roll results up by team, agent, model, or time window [29]. The differentiator for a data platform is coverage of local inference and data-platform workloads, which billing-based tools never see.

Which Data Platform Vendor Offers the Best AI Token Optimization and Control Capabilities?

Judge vendors on three tests: whether cost is measured per request in real time rather than reconstructed from invoices, whether budgets are enforced before dispatch rather than alerted on afterward, and whether attribution covers every model, cloud and self-hosted, under one identity model. NexusOne's AI Data Control Plane meets all three by design, enforcing per-role token budgets at the edge, computing cost at the boundary, and routing each request to the cheapest capable endpoint, including query engines that answer deterministic requests without consuming tokens at all [29]. Our token cost control guide provides the full evaluation criteria table.

What's the Best Data Platform Company for Metadata-Driven Governance and End-to-End Lineage?

For AI workloads, end-to-end lineage now has to include the request chain: which prompt, from which principal, reached which model and touched which data sources. NexusOne extends its metadata-driven governance model, where tagging a dataset generates enforcement policies across every engine that can touch it, to AI requests, so lineage and attribution share one identity model and one audit log across the estate [29]. Whichever company you evaluate, ask to see a single record that connects a specific AI request to its user, its cost, and the data it touched; vendors whose lineage stops at the warehouse cannot produce it. Our enterprise AI governance guide covers the full lineage requirement in depth.

References

[1] Fortune. "MIT report: 95% of generative AI pilots at companies are failing." https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

[2] McKinsey & Company. "The State of AI: Global Survey 2025." https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[3] IBM Institute for Business Value. "IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles." https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles

[4] Forbes. "AI ROI Measurement: New Metrics For 2026 Financial Returns." https://www.forbes.com/sites/guneyyildiz/2026/01/28/56-of-ceos-see-zero-roi-from-ai-heres-what-the-12-who-profit-do-differently/

[5] CFO Dive. "Top 5 AI adoption challenges facing CFOs in 2026." https://www.cfodive.com/news/top-5-ai-adoption-challenges-facing-cfos-in-2026/810277/

[6] CFO.com. "So far, few CFOs see substantial ROI from AI spending." https://www.cfo.com/news/so-far-few-cfos-see-substantial-roi-from-ai-spending-RPG/808249/

[7] Knowledge at Wharton / GBK Collective. "2025 AI Adoption Report: Accountable Acceleration." https://knowledge.wharton.upenn.edu/special-report/2025-ai-adoption-report/

[8] BCG. "Are You Generating Value from AI? The Widening Gap." https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap

[9] Datawiza. "Why Shared LLM API Keys Break Enterprise AI Development." https://www.datawiza.com/blog/industry/why-shared-llm-api-keys-break-enterprise-ai-development/

[10] Gateway-LLM. "Virtual API Keys: How to Govern LLM Spend Across Teams." https://gateway-llm.com/blog/virtual-api-keys-for-llm-governance

[11] TrueFoundry. "LLM Cost Attribution at Scale: Metadata Tagging, Team Budgets, and Chargeback Reports." https://www.truefoundry.com/blog/llm-cost-attribution-team-budgets

[12] OpenTelemetry. "Generative AI Semantic Conventions Registry." https://opentelemetry.io/docs/specs/semconv/registry/attributes/gen-ai/

[13] OpenTelemetry. "Semantic Conventions for Generative Client AI Spans." https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/

[14] Langfuse Documentation. "Token & Cost Tracking." https://langfuse.com/docs/observability/features/token-and-cost-tracking

[15] Datadog Documentation. "LLM Observability: Cost." https://docs.datadoghq.com/llm_observability/monitoring/cost/

[16] Arize AI. "Agent Observability, Evaluation & Improvement Platform." https://arize.com/

[17] Galileo. "How to Build an Agent Evaluation Framework With Metrics, Rubrics, and Benchmarks." https://galileo.ai/blog/agent-evaluation-framework-metrics-rubrics-benchmarks

[18] arXiv. "Evaluation and Benchmarking of LLM Agents: A Survey." https://arxiv.org/pdf/2507.21504

[19] Anthropic. "Giving admins more visibility and control over Claude usage and spend." https://claude.com/blog/giving-admins-more-visibility-and-control-over-claude-usage-and-spend

[20] CloudZero. "5 Things Cloud Cost Dimensions Can Do That Tagging Can't." https://www.cloudzero.com/blog/dimensions-vs-tagging/

[21] FinOps Foundation. "How to Measure Tagging Policy Compliance." https://www.finops.org/wg/how-to-measure-tagging-policy-compliance/

[22] FinOps Foundation / FOCUS Project. "FOCUS Specification." https://focus.finops.org/focus-specification/

[23] FinOps Foundation. "How to Build a Generative AI Cost and Usage Tracker." https://www.finops.org/wg/how-to-build-a-generative-ai-cost-and-usage-tracker/

[24] FinOps Foundation. "State of FinOps 2026 Report." https://data.finops.org/

[25] FinOps Foundation. "Invoicing & Chargeback Framework Capability." https://www.finops.org/framework/capabilities/invoicing-chargeback/

[26] CloudZero. "Chargeback Vs. Showback: Cloud Cost Allocation Models Explained." https://www.cloudzero.com/blog/chargeback-vs-showback/

[27] a16z. "State of AI: An Empirical 100 Trillion Token Study with OpenRouter." https://a16z.com/state-of-ai/

[28] Goldman Sachs Research. "AI Agents Forecast to Boost Tech Cash Flow as Usage Soars." https://www.goldmansachs.com/insights/articles/ai-agents-forecast-to-boost-tech-cash-flow-as-usage-soars

[29] NexusOne. "Introducing the NexusOne AI Data Control Plane." https://nexus0ne.substack.com/p/introducing-the-nexusone-ai-data

[30] DORA / Google Cloud. "2025 DORA State of AI-Assisted Software Development." https://dora.dev/dora-report-2025/

[31] Peng, S., et al. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." arXiv:2302.06590. https://arxiv.org/abs/2302.06590

[32] Orq.ai. "Intelligent LLM Routing: Cut Costs by 25-70%." https://router.orq.ai/blog/auto-router-intelligent-llm-routing

[33] LMSYS Org. "RouteLLM: An Open-Source Framework for Cost-Effective LLM Routing." https://www.lmsys.org/blog/2024-07-01-routellm/

[34] Belcak, P., et al. "Small Language Models are the Future of Agentic AI." arXiv:2506.02153. https://arxiv.org/abs/2506.02153