AI spend is moving from hyperscaler capex to enterprise opex, and the buyers are now asking for oversight of where and how AI is used. AICost.ai is a deterministic decision layer: 120+ engines that model any AI workload before it exists, a policy engine that compiles the approved numbers into gateway and IAM configuration, and a signed evidence loop that proves the workload stayed inside its envelope. Same product for the enterprise agent fleet and for the solo developer's subscription.
An agent loop has a cost view (turns per task, tokens per turn, runaway probability) and a governance view (maximum turns, cost cap, fallback model, kill switch, review threshold). They are the same variables. Whoever models them once owns both the FinOps conversation and the compliance conversation, and can turn the model into the rule that runs in production.
Retrospective by design. They price tokens and allocate compute. They cannot cost a workload that does not exist, and they carry no controls, obligations or evidence chain.
Forms and attestation storage. No models, no agents, no telemetry. The policy is a document and the enforcement is a reminder.
In-path routers apply a rule someone else wrote. AICost runs before the routing layer, out of band, and exports the decision as the gateway's own config.
Every AI workload, at every scale, has a cost shape that a single model prices wrong and a governance obligation that arrives with the first regulated user. The layer that prices and governs them is horizontal.
Hundreds of workloads across engineering, support, legal, finance, HR and marketing, each with a different cost shape and data class.
Providers accountable for their clients' AI, needing discovery, policy and proof per tenant at a flat monthly price.
Seat vs metered, agent loops, context growth, the $29 developer who became $750. Calculators, playgrounds and MCP keys.
Local models on laptops and phones that no proxy sees; the placement decision between device, edge and cloud priced per call.
Per-episode economics with no token in them, demonstration corpora, teleoperation, buy vs RaaS.
// One rate catalogue underneath all five: 150 models, 17 providers, 488 Bedrock rates across 15 regions, refreshed daily with 23 automated checks.
| Layer | What it is | Status |
|---|---|---|
| Decision engines | 120+ deterministic workload models across agentic, RAG, voice and media, training and MLOps, evaluation, platform, seats, routing, cloud placement, governance, diligence and physical AI. Typed contracts, golden test suites, provenance on every answer | Live, human calculators and machine-callable |
| MCP server | mcp.aicost.ai. Multi-tenant, OAuth for humans and bearer keys for pipelines. Chains assemble because each engine names its successors | Live since July 2026 |
| Pricing single source of truth | Vendor pages read directly, never aggregators. 23 daily checks for duplicates, reference disagreement, retirement, scheduled changes, catalogue size, provenance and routing coherence | Live, refreshed daily |
| Policy engine and CostWall | Eleven intake questions, six ordered gates per route, six templates, elimination trace, policy_patch export to LiteLLM, Portkey, Cloudflare, Kong, OpenRouter, Bedrock IAM, Azure RBAC, Vertex org policy | Live; downgrade and hard-stop proven on live traffic 14 July 2026 |
| Governance assessment | 20 modules, 200 questions, 247 CSA AICM controls mapped to ISO 42001, NIST AI RMF, EU AI Act and 12 US state statutes. Connected to Entra, Defender, Intune, licensing, Bedrock, Azure and GCP exports | Live; 15 modules with working engines, the rest questionnaire-backed |
| CostProof ledger | Hash-chained, HMAC-signed journal; seven weekly watchers; monthly drift report with a customer-runnable verify command; FOCUS 1.x, OSCAL and STAR exports | Live |
| Adjacent assets | ToolsInfo: 115,968 tools across 5,700+ categories. Vendor evidence graph: 39,531 governance facts on 4,148 vendors, quoted and sourced. CostOptimization.ai for cloud bill diagnosis. MSPCost for the provider market | Live |
| Commercial | Stripe billing live. Service ladder from free Quick-Check to fixed-fee assessment to flat monthly Continuous Governance. MSP resale per tenant | Live; enterprise pilots in progress |
The same engines that govern an enterprise workload tell a buyer whether a target's or portfolio company's AI unit economics hold. Vendor-exact rates, the full cost stack rather than the token line, and the governance exposure the acquirer inherits.
Re-prices a company's claimed cost per unit against vendor-exact rates and the eight layers most decks omit: retrieval, refresh, retries, guardrails, evaluation, human review, observability, maintenance.
Agent products fail on variance, not on the mean. Runaway tail, turn ceilings and the reserve a business plan needs before the number is believable.
Margin on an AI-powered product line at current volumes and at 10x, including the tier transitions on the way.
What a single provider repricing does to the P&L, with the announced-but-cancelled changes tracked rather than remembered.
Shadow AI on the card, vendors training on submitted data, workloads with no owner, EU AI Act classification. Findings priced and owned.
One dollar figure for the controls a regulated portfolio company must run, computed at current volumes, not estimated once.
Subu Vdaygiri built the CloudBlue data lake at Ingram Micro that reconciled bills for hundreds of MSPs, launched Ingram's Microsoft NCE, AWS and GCP programs for its global marketplaces, and spent the prior decade on enterprise cloud at Siemens. Ten AWS and Azure certifications, Wharton CTO and Kellogg CPO credentials, NVIDIA Inception member. Co-founded with Hanvish Vdaygiri; Ravi Erraguntla leads the Austin practice.
For a confidential conversation about the raise, or a diligence layer for AI unit economics across your portfolio.
✉ [email protected]