Zero-instrumentation per-customer AI margin meter that plugs vertical SaaS billing into real token/GPU cost truth.

1. DoiT just proved zero-instrumentation, kernel-level AI cost attribution works at enterprise scale, derisking the core metering technique for a smaller entrant to adapt to the SaaS-tenant margin use case.
2. DoiT's own survey shows 85% of enterprise leaders cannot calculate AI ROI without major bottlenecks, meaning the pain is broad enough to extend past DoiT's internal-chargeback focus into the external customer-margin question.
3. An established FinOps incumbent chose to acquire an attribution startup rather than build the capability internally, confirming enterprise buyers will pay for this category now even though the winning product surface is still being defined.
4. Reporting ties the attribution failure directly to shared GPU pools and agentic workloads, which are exactly the conditions vertical SaaS companies face once they embed third-party AI agents behind a single multi-tenant backend.

Catalyst. DoiT's acquisition of Attribute and its zero-instrumentation kernel-level metering, launched the same week enterprises admitted only 15% can calculate AI ROI, proves the underlying attribution technology and buyer urgency now exist, while DoiT's internal-chargeback focus leaves the external per-customer margin question open for a wedge-focused entrant.

A lightweight proxy or sidecar sits at the model-call layer (LLM gateway, API gateway, or network tap) and fingerprints each request back to the initiating end-customer tenant or session without requiring code changes from the client team. It aggregates token counts, GPU-second estimates, and vector-store query costs across every inference provider the company uses, then streams a per-tenant, real-time margin ledger into the company's billing system (Stripe Billing, Orb, Metronome) and BI stack. Pre-built margin-guardrail rules let a product or finance owner alert, throttle, or nudge an account to a higher tier before a customer segment becomes structurally unprofitable, closing the loop between cost data and pricing action.

What's different. Unlike DoiT/Attribute, which sells into central IT and FinOps teams to chargeback AI spend across internal departments, this product sells to vertical SaaS companies that need per-external-customer margin visibility feeding directly into their own billing and pricing systems, a narrower wedge that horizontal FinOps platforms are not built to serve. Because the product plugs into usage-based billing engines rather than an internal chargeback dashboard, it captures value at the exact moment of a pricing or renewal decision, not just a monthly cost-report cycle.

Jobs to be done

flowchart LR
  Customer[End customer request] --> Proxy[Metering proxy]
  Proxy --> Providers[Multi-provider LLM and GPU calls]
  Providers --> Ledger[Per-tenant margin ledger]
  Ledger --> Billing[Usage-based billing engine]
  Ledger --> Alerts[Margin guardrail alerts]
  Billing --> Finance[Finance and product decision]
  Alerts --> Finance

Executive takeaways

• DoiT’s Attribute launch and Portkey’s later acquisition show that AI cost attribution and gateway control are already strategic budgets, but the open wedge is external-customer margin rather than internal chargeback.
• The enabling stack is real today: gateways and observability tools already capture user and session metadata plus costs, while billing platforms can rate real-time usage; the missing layer is finance-grade tenant profitability and billing action.
• Buyer urgency is credible: Menlo says enterprise generative AI spend reached $37B in 2025, FinOps Foundation says 98% of practitioners now manage AI spend, and CloudZero still finds widespread AI ROI and attribution blind spots.
• Competition is fragmented rather than settled. Horizontal FinOps platforms, AI gateways, and billing vendors each own part of the workflow, so a focused entrant can win if it proves invoice-grade accuracy, zero-code deployment, and direct billing-system outcomes.

Market definition

The relevant market is software that attributes multi-provider AI COGS to external SaaS tenants and connects that cost truth to billing, pricing, and margin-control workflows.

Customer and buyer

The daily operator is usually an AI product or platform leader plus a finance or RevOps owner trying to understand margin by tenant, feature, or plan. The economic buyer is typically a CFO or VP Finance partnering with the Head of AI or Product once AI COGS begin to affect renewals, packaging, or board narratives.

Buying triggers

• A quarterly review, renewal, or board conversation reveals that the team can see total AI spend but still cannot explain margin by customer or product tier quickly enough to act. [9][2]
• The company is launching or revising usage-based AI packaging and needs accurate per-tenant cost truth before shipping a self-serve or mid-market plan. [43][55][67]
• AI traffic expands across multiple model providers plus retrieval or vector layers, turning one clean invoice into many hidden cost surfaces that product and finance can no longer reconcile manually. [23][104][107][120]

Willingness to pay

Willingness to pay is credible because AI spend is already large and still compounding, FinOps teams now explicitly manage AI spend, and leading AI vendors already buy external real-time billing infrastructure to operationalize usage transparency. The startup is not asking buyers to fund a theoretical category; it is asking them to close a known control gap between AI COGS and revenue decisions. [3][6][56][67][41]

Category dynamics

Growth signal 3.2x YoY enterprise generative AI spend growth from 2024 to 2025

Validation signals

• DoiT bought Attribute and launched zero-instrumentation tokenomics instead of waiting to build a similar capability internally, validating buyer urgency for AI spend attribution.
• FinOps Foundation reports that 98% of practitioners now manage AI spend, confirming that AI economics has become a standard budget and operations problem.
• OpenAI and Replicate both rely on Metronome for real-time usage, credits, and billing workflows, proving that sophisticated AI companies already buy external metering infrastructure.
• Palo Alto Networks’ intent to acquire Portkey shows that AI gateways are strategic control points rather than disposable developer tools.

Regulatory & technical constraints

• Provider pricing is heterogeneous across input, output, cached input, regional uplift, and partner-cloud variants, so a margin ledger must continuously normalize changing unit types rather than assuming one token price.
• Vector and retrieval layers add separate storage, read, write, and index-compression cost surfaces that can materially affect per-tenant COGS beyond raw model tokens.
• Shared infrastructure allocation needs to account for workload, idle, and overhead costs rather than summing only directly observed API calls or token counts.
• Prompt, response, and tenant metadata logs may trigger retention, minimization, security, and audit obligations, especially in regulated or cross-border deployments.

Three adjacent stacks are converging on the problem: internal FinOps and cloud unit-economics platforms, AI gateways and observability vendors, and usage-based billing platforms. The gap is the finance-grade tenant margin ledger that sits between request telemetry and billing action for external customers rather than internal cost centers.

Why incumbents do not win by default

• Cloud FinOps platforms. CloudZero, Finout, and adjacent FinOps tooling translate infrastructure cost into unit economics, but their center of gravity is internal optimization and reporting rather than inline margin action for external SaaS tenants.
• AI gateways and observability vendors. Helicone, Portkey, and Langfuse can already capture model usage, metadata, and traces at the request layer, but they stop short of becoming the finance-owned profitability and billing decision system.
• Billing platforms. Orb, Metronome, and Stripe excel at rating, credits, invoices, and plan changes, but they assume someone else has already normalized the underlying AI cost and tenant attribution data.
• Model and cloud providers. OpenAI, Anthropic, Google, and Azure expose pricing and some retention controls, but native dashboards are provider-specific and do not answer cross-provider customer-margin questions for SaaS vendors.
• In-house scripts and spreadsheets. Manual reconciliation remains common because it is the fastest local fix, but it breaks as soon as providers, token types, or vector and retrieval costs proliferate across the stack.

AI Margin Metering is a finance-grade tenant COGS ledger for Series B-D vertical SaaS companies that recently launched embedded AI copilots and now need to defend gross margin before renewals or pricing changes. The researched opportunity is not generic AI observability: gateways, FinOps tools, and billing systems already exist, but none of them owns the per-external-customer margin decision inside a SaaS vendor's revenue workflow. The initial wedge is a reconciliation-first proxy or gateway layer that attributes token, GPU, and vector costs to each tenant without SDK changes, then pushes that ledger into Stripe Billing, Orb, or Metronome before a pricing or renewal event. Research supports a roughly $0.4B TAM, a $120.0M near-term SAM, and a $3.6M year-3 SOM, but those estimates assume the company can convert a narrow trigger-based sales motion into repeatable channel-assisted distribution. The first customer should buy because finance has already flagged an underwater AI account or the company is about to ship usage-based AI packaging and cannot defend unit economics with spreadsheets. Product sequencing matters: the MVP should prove invoice reconciliation and one billing integration before automating throttles or broader pricing controls, because finance distrust and compliance objections are the main adoption risks. The company wins if it becomes the fastest way for a CFO and Head of AI Product to go from total AI spend visibility to a trusted answer on which tenant, feature, or plan is profitable and what pricing action to take. The largest disconfirming risk is not raw market demand; it is whether buyers fund a standalone margin-control layer instead of stretching existing FinOps, gateway, or billing vendors, and whether zero-SDK tenant attribution holds up across async agent workflows.

Problem

• Vertical SaaS teams embedding AI across multiple providers cannot see per-tenant AI COGS until provider invoices and internal logs are manually reconciled, so unprofitable accounts survive until a quarterly close or renewal.
• When pricing teams roll out AI usage tiers, they lack invoice-grade cost truth across tokens, GPU, and vector retrieval, so pricing, throttling, and packaging decisions rely on estimates.

Solution

• Insert a proxy or gateway-side ledger that fingerprints requests to the initiating tenant, normalizes token, GPU, and vector costs across providers, and reconciles them against actual provider bills.
• Push per-tenant COGS into the customer's billing and BI stack, starting in read-only reconciliation mode and then enabling alerts, tier recommendations, and guardrails once finance trusts the data.

Why we win

• The product sits in the gap between gateways or observability and billing: closer to the request layer than CloudZero or Finout and closer to finance action than Helicone, Portkey, or Langfuse.
• Each deployment builds reusable reconciliation logic across provider pricing units and billing integrations, creating switching friction and a dataset of which pricing actions actually improve margin.

Milestones

flowchart LR
  Wedge[Renewal or pricing trigger] --> MVP[Reconciliation first tenant ledger]
  MVP --> Proof[Invoice grade trust plus billing sync]
  Proof --> Expansion[Margin guardrails across more workflows]

Founding team

Experiment roadmap

Risk assessment

What must be true

• At least 8 of the first 20 ICP interviews must reveal a live renewal, board, or pricing trigger tied to AI margin uncertainty.
• The first 3 pilots must reconcile per-tenant AI COGS to provider invoices within 5% variance on the initial workflow.
• At least 2 of the first 4 paid pilots must convert to annual production once one billing integration is live.
• At least one of Stripe Billing, Orb, or Metronome must cover the majority of early time-to-value integrations in the beachhead.
• Prospects must keep preferring a dedicated margin-control layer over stretching existing FinOps, gateway, or billing tooling.

Open diligence questions

• What evidence shows zero-SDK tenant identity holds through async jobs and multi-step agents?
• Which budget line signs first: CFO, AI product, RevOps, or platform engineering?
• How often do pricing or renewal projects stall because per-tenant AI COGS is unknown today?
• What invoice variance will finance accept before refusing to use the ledger for pricing or renewals?
• How quickly can DoiT, CloudZero, Portkey, or a billing vendor close this gap with adjacent features?

Model sanity

• Revenue engine. Base revenue comes from growing from 3 paying accounts at Y1 exit to 24 by Q4Y3 while exit annualized ARPU rises to about $124K through usage fees and second workflows.
• Must go right. The company must keep pilot-to-production conversion near one quarter and show at least one repeatable partner referral path, because sales cycle is the largest revenue and runway sensitivity.
• Model breaks if. If buyers delay standalone budget ownership and gross margin stalls near 66%, the downside case pushes cash below zero before the seed proof point arrives.
• Next-round proof. The seed story is 6-10 production logos by late Y2, at least 3 expanded workflows, and proof that one billing or gateway channel can source repeatable qualified pipeline.

Scenarios

Sensitivity

flowchart LR
  TargetAccounts[Trigger-based target accounts] --> PaidPilots[Paid pilots]
  BillingPartners[Billing and gateway partners] --> PaidPilots
  PaidPilots --> ProductionLogos[Annual production logos]
  ProductionLogos --> Expansion[Second workflow expansion]
  Expansion --> Revenue[Revenue]
  Revenue --> GrossProfit[Gross profit]
  GrossProfit --> Cash[Cash and runway]

Flags: CustomersEop includes paid pilots plus production contracts, so recurring-only production logos lag the headline account count through Y1. · Base-case cash bottoms near $0.25M in Q3Y3, so a two-quarter slip in pilot conversion likely forces slower hiring or an earlier seed. · Exit ARPU rises above the researched $60K initial ACV because the model assumes usage-based fees and second-workflow expansion; single-workflow accounts would undershoot revenue. · Y2 burn remains heavy while direct sales and partner channels are still being proven, so standalone budget ownership is the main commercial risk. · Cash is modeled as EBITDA; implementation prepayments, deferred revenue timing, or security-review capex could move actual cash collections materially.

• Incumbent downmarket move. DoiT/Attribute or a horizontal FinOps platform could extend its internal chargeback product to cover per-external-customer margin metering. Mitigation: Win on billing-system-native integration and vertical-specific margin guardrails that a horizontal chargeback tool is not built to ship first.
• Provider API fragmentation. Each new LLM provider, agent framework, or GPU orchestration layer changes how usage data is exposed, risking constant adapter maintenance. Mitigation: Build a provider-agnostic metering core with a plugin architecture and prioritize adapters for the 3-4 providers the beachhead segment actually uses.
• Attribution accuracy disputes. If per-tenant cost estimates are wrong, customers will distrust the margin ledger and refuse to base pricing or renewal decisions on it. Mitigation: Publish a reconciliation report against actual provider invoices during onboarding and let customers audit the metering logic before it drives billing.