Autonomous AI FinOps

The token bill is one line of your AI cost. [i]Unravel finds the rest — then cuts both.[/i]

Tokens land on the model invoice. The retrieval, jobs, indexes and storage around them land on the data platform invoice. Most AI cost tools see one invoice or the other. Unravel joins both at the workload level — attributed to an owner, measured against the outcome, optimized only where eval pass rate and p95 latency hold, and held there by policy.

1.7×

end-to-end cost vs. token bill · retrieval-heavy support agent

92–98%

cost cut on one customer’s largest AI pipeline

<4 wk

typical time to first applied savings

The challenge

AI spend is compounding — [i]and ownership hasn’t kept up.[/i]

Per-token prices keep falling, but agents, reasoning models and longer contexts consume more tokens per task — so total spend climbs anyway. It lands on a dozen invoices that close on different cycles. Finance sees a total — not who spent it, on what, or what it returned.

Fragmented

The spend is scattered across a dozen bills.

Data platforms, foundation-model APIs, coding assistants, cloud AI services, gateways, MCP servers. Each with its own invoice and its own unit — input, output, cached and reasoning tokens, DBUs, credits, seats, provisioned throughput.

What Unravel does

Normalizes each connected provider into one AI workload data model, FOCUS-aligned where billing data allows — so AI cost and data cost land in the same view, under the same owner.

Unattributed

A meaningful share of AI spend belongs to nobody.

Untagged endpoints turn chargeback into a negotiation. The team that shipped the agent isn’t the team that sees the bill, and nothing on the invoice connects the two.

What Unravel does

Infers ownership from workspace, identity, job lineage and call path where tags are missing — shows the confidence of each assignment — and surfaces the untagged endpoints left to fix.

Unoptimized

Waste compounds quietly, with no guardrail to catch it.

Frontier models on trivial tasks. High reasoning effort on simple extraction. Idle endpoints. Uncached repeats. Verbose system prompts on every request. An agent stuck in a retry loop. A runaway job that burns a month of budget in five hours.

What Unravel does

Right-sizes the model, trims the prompt, caches the repeats, evicts the idle capacity — low-risk changes automatically, the rest with approval, each validated on eval pass rate, p95 latency and error rate before it ships.

——— The end-to-end claim

AI cost and data cost [i]aren't two problems.[/i]

An agent doesn't just call a model. It retrieves from a table, runs a job, refreshes an index, writes traces back. The token cost is the visible part. For retrieval- and batch-heavy workloads, the compute around it is often as large or larger — and it’s on a different invoice, owned by a different team, in a different tool.

Unravel reports both against the same agent, the same job, the same owner. That’s the precondition for a unit cost you can defend — cost per resolved ticket, per document processed, per merged PR — the figure that decides whether a workload gets more budget or gets retired.

1.7×

retrieval-heavy agent

2.0×

batch inference pipeline

1.9×

coding agent fleet

End-to-end cost as a multiple of the token bill, for the three example workloads above. Your multiple depends on architecture — that’s what the teardown measures.

Why we can act, not just report

Behind every charge is a query, a job or a prompt. [i]That’s where the fix goes.[/i]

Unravel ties each charge to the query, job, trace and prompt that produced it — and, where you connect a repository, to the line of code. That’s what separates a recommendation from a fix: the change is specific enough to review, test and apply.

Prompt optimization, not prompt advice

Unravel sees the prompt as it’s sent — in traces, or in the code that assembles it — so it can show a 4,200-token system prompt going out on every request, the 800-token version that held its eval pass rate within 0.3 points on a 1,000-query set, and the static prefix that provider prompt caching would bill at a fraction of the input rate.

Model right-sizing per call site

Opus where Sonnet would do. A frontier model on FAQ classification. High reasoning effort on a lookup. Unravel spots the mismatch at the call site, measures the change in eval pass rate and p95 latency on a sample of your traffic, and routes simple requests down a tier while complex ones stay put.

The job around the call

The paging pattern, the unnecessary re-scan, the cluster sized for the worst batch, the real-time calls that could run on a provider’s discounted batch tier. This is the part most AI cost tools don’t model — and on batch inference it’s often the larger share.

Fine-grained chargeback by consumer

Warehouse and endpoint cost split by the app that actually drove it — ServiceNow, Power BI, Genie, an internal agent, a named developer. Not one line item called "AI."

Coverage

A dozen invoices, a dozen units of measure. [i]One data model underneath them.[/i]

Unravel ingests the signals that explain AI spend, populates one data model built for AI workloads, and delivers attribution, optimization and enforcement from that same model — so the chargeback finance sends, the saving engineering ships and the budget platform enforces all start from the same figure.

1. Ingest

Bills and FOCUS exports, system tables, OpenTelemetry GenAI traces, gateway and MCP logs, job and query plans — and source code, if you connect it. Sources below.

2. Populate

One model of AI workloads: agent → model → endpoint → infrastructure → storage → user → team → outcome. Token types, DBUs, credits, seats and throughput units normalized to one cost basis, kept current by pipelines and agents, with a confidence score on every inferred owner.

3. Deliver

Attribution, optimization and enforcement all read the same model — so the number in the chargeback is the number the optimization saved and the number the budget enforces. Detail in the next section.

Stage 1 · Ingest — what we connect to

Data & analytics platforms

Databricks · Snowflake · BigQuery

Model serving, AI functions, Cortex, BigQuery AI functions, and the jobs and warehouses underneath them.

Foundation-model APIs

Anthropic · OpenAI · Google · the long tail

Direct API spend per model, endpoint and destination — split by input, output, cached and reasoning tokens, including fallback traffic.

Coding assistants

Claude Code · Cursor · Codex · Copilot

Per team, developer and session, with accepted suggestions and merged PRs alongside the spend.

Cloud AI platforms

Bedrock · Vertex · Azure AI Foundry

Managed inference, provisioned throughput, and the GPU capacity sitting idle behind it.

How it lands

Attribute. Optimize. [i]Enforce.[/i]

Three layers on one data model. Build them in order — attribution makes optimization credible, optimization makes enforcement defensible.

Every AI dollar
has an owner.

Attribution first, because nothing downstream survives a number finance doesn’t believe. Every call rolls up to a team, a use case, a model and an endpoint — even where tagging is imperfect, with the confidence of each inferred assignment shown.

Team, use case, model, endpoint

One hierarchy across every provider. Business unit, application, workspace, environment, user, SKU — pivot however finance asks.

Platform vs. usage split

Infrastructure and platform cost separated from token consumption, with shared cost allocated rather than dumped in an “other” bucket.

Tag the untagged

Ownerless endpoints surfaced by name and value, so the gap in your chargeback is a work item instead of a mystery.

Unit economics that mean something

What a resolved ticket, a processed document or a merged PR costs to produce, trended next to volume and quality — so “AI is getting expensive” becomes “cost per resolved ticket fell 11% while volume doubled.”

Service-to-destination routing

Apps call a stable endpoint, not a model. Unravel tracks which destination actually served each request — including the 2.3×-cost fallback nobody knew was firing.

unravel · ai / finops · last 30 days

AI spend

$284.2K

+12.4% MoM · 9.8% of cloud spend

Savings opportunity

$41.3K

11 recommendations · 5 high-priority

Blended cost / request

$0.295

690,000 requests · 43% cache hit

Unattributed

$18.7K

6.6% of spend · 3 endpoints

Spend by business unit · against budget

Customer Success 111% of budget $61.2K

AI Research 79% of budget $47.3K

Data Platform 86% of budget $38.9K

Product 83% of budget $29.1K

(untagged) no owner $18.7K

Platform $12.1K · usage $49.1K · exportable by BU, team and pipeline

Chargeback ready

From recommendations
to results.

Most AI cost tools stop at a list. Unravel applies the changes you approve. AutoApply for low-risk changes, human-in-the-loop for anything touching production behavior. You set the threshold per workload.

Model right-sizing with a quality number

Not “consider a cheaper model.” A measured quality delta on your traffic, a latency delta, and a routing split that sends only the simple queries down a tier.

Prompt and context reduction

Redundant system-prompt instructions, unfiltered payloads, oversized outputs — trimmed at the call site, tested before apply.

Caching and duplicate elimination

Semantic caching on repeat traffic, embedding reuse instead of recompute, duplicate requests collapsed.

Idle and cold-start waste

Scale-to-zero on under-used endpoints, consolidation of low-traffic ones, GPU pools at 4% utilization evicted. Launch spend tracked separately from operational spend.

The job, not just the call

Query rewrites, partition pruning, cluster right-sizing on the pipeline the inference runs inside — the same engine Unravel already runs on data platform spend.

Pricing-tier arbitrage

Batch tiers for work that isn’t latency-sensitive, prompt caching for stable prefixes, provisioned throughput only where utilization justifies the commitment — the same tokens, billed at a lower rate.

unravel · ai / optimize

11 actions · $41.3K/mo recoverable · 4 low-effort

5 high-priority

⇄

support-copilot · model routing

Frontier-tier model on FAQ classification. Route confidence >0.85 to economy tier: 70/30 split.

$8.2K

Needs Review

✎

prod-ml-workspace · prompt optimization

System prompt 4,200 → 800 tokens. Eval pass rate 96.4% → 96.1% on a 1,000-query set.

$3.4K

AutoApply

◈

rag-service · semantic caching

31% of retrieval traffic is a near-duplicate of a request served in the last hour.

$5.1K

AutoApply

⏻

ft-h100-pool · idle GPU capacity

4% utilization for 3 consecutive days. Scale to zero, retain the image.

$4.9K

AutoApply

Before apply · measured, not estimated

Baseline (current)

frontier tier

$42.8K/mo

96% quality

420ms latency

→

Recommended

economy tier · 70% of traffic

$30.0K/mo−30%

93% quality−3%

180ms latency−57%

A 3-point drop in eval pass rate (96% → 93%) is inside the 5-point threshold you set for this query class — and the complex 30% never leaves the frontier model.

Optimization without enforcement
is a savings estimate.

The same overrun comes back next quarter unless something stops it. Budgets, quotas, routing rules and model policy enforced where the call happens — not reported after the invoice clears.

Budgets and quotas per team

Thresholds that escalate before the money is gone. Per-user and shared budget models, tracked against actual spend hourly.

Per-task ceilings for agents

Max steps, max tokens and max spend per task, so an agent caught in a loop stops at a dollar figure you chose — not when the budget runs out.

Anomaly detection from telemetry, not the invoice

A runaway job that burns 8,158 credits in a five-hour window gets flagged within minutes where streaming telemetry is available — not on the month-end statement.

Model and asset governance

A first-seen model appearing in production traffic is an event, not a surprise line item. Approval gates on new models, endpoints and provisioned capacity.

Enforcement at the gateway

Throttling, routing and policy applied through the AI gateway you already run — or through Unravel’s, deployed in your VPC. Governed and ungoverned endpoints, clearly separated.

Change control with a full audit trail

Every action carries a state — pending review, needs input, approved for auto-apply, applied — with an owner, a target date and a record of who signed off.

unravel · ai / governance · active

12 active policies · all enforcing

1 team over budget

◇

Model approval

No first-seen model serves production traffic without review. gpt-5.2-vision caught at $2.1K in its first 48 hours.

Active

◷

Budget escalation

Alert at 80% of monthly team budget, escalate to VP if no action in 48h. Customer Success at 111% — escalated.

Active

⇄

Routing policy

Fallback to the 2.3×-cost destination only on rate limit, never on latency. support-copilot fallback rate 14.2% → 3.1%.

Active

⚠

Anomaly · high severity

ft-h100-pool idle at 4% for 3 days. 8,158 credits burned by a single job in a 5-hour window, flagged in minutes.

Escalated

▤

Ungoverned endpoints

3 endpoints serving production traffic outside policy — surfaced with owner, spend and traffic split.

Active

Approval workflow · 1 auto-applied · 4 pending review · 5 needs input

0 rejected

Why existing tools fall short

Everyone is shipping an AI cost view. [i]Here's where each one stops.[/i]

Every category owns a slice of the problem. None of them, on its own, closes the loop from the token to the job to the enforced policy — and several are inputs Unravel reads from.

——— The enforcement layer

An overrun you only report [i]is an overrun you’ll see again.[/i]

Unravel decides the policy from the full cost-and-quality picture, then enforces it wherever your traffic actually flows. Integrate with the gateway you already run, or deploy Unravel’s gateway inside your VPC so prompts and responses stay inside your perimeter.

Will right-sizing models quietly degrade my product?

No — that's why the quality delta is measured before anything is applied. Every right-sizing recommendation carries a quality score on your own traffic, a latency number, and a routing split that keeps complex requests on the frontier model. If eval pass rate drops past the threshold you set, or p95 latency rises past its budget, the change doesn’t ship.

Do I have to tag everything before this works?

No. Attribution rolls up from telemetry, code and job metadata, not tags alone — so you get showback on day one, and chargeback as attribution confidence rises. Untagged endpoints are surfaced by name and spend so the remaining gap is a short work item, not a prerequisite.

Does it cover AI spend outside my data platform?

Yes — direct foundation-model APIs, coding assistants, cloud AI platforms, gateways and MCP servers, alongside in-platform AI functions and model serving. That's the point of a shared data model: one normalized view, not one integration per bill.

How do I know an AI workload is worth what it costs?

Attach an outcome to it — a resolved ticket, a processed document, a merged PR. Unravel trends cost per outcome next to task-success rate and volume, so a workload whose cost rose 40% while it resolved twice the tickets shows up as a better deal, and one whose cost rose on flat output gets flagged.

Trust by design

Cheaper is only an optimization [i]if the output still holds up.[/i]

Every change is validated against your real traffic before it's applied — model swaps against sampled output quality, prompt reductions against scored responses, job rewrites against production history. If validation shows eval pass rate, p95 latency or error rate moving past your threshold, Arvix doesn’t apply it. That’s why customers run AutoApply on 33% of actions.

Zero production incidents attributed to Arvix-applied optimization

——— Customer story

A global cybersecurity company cut its dominant AI workload [i]92–98%.[/i]

A support-transcript analysis pipeline running LLM calls inside a Databricks job. The AI spend showed up in one place, the compute in another, and no one could say what the pipeline actually cost — let alone whether it was running efficiently.

Unravel attributed both halves to the same job, then rewrote the work around the call: batch ranges instead of full re-scans, transcripts pre-filtered before the prompt, and an economy-tier model for a structured-extraction task that never needed a frontier one. Model right-sizing, batching and pre-filtering alone — before caching or budget controls entered the picture — with field-level extraction accuracy held at baseline.

"We knew what the model calls cost. We had no idea what the pipeline around them cost — and that turned out to be the larger number."

Director of Data Engineering · Global cybersecurity company

92–98%

cost cut on the dominant workload

8,158

credit spike caught in minutes, not at month-end

99.6%

of answers unchanged after the 93% cost cut

Recommended resources

How to Stop Burning Money (or at Least Slow the Burn)

Articles

Cost Opt & FinOps

IDC Analyst Brief: The Role of Data Observability and Optimization in Enabling AI-Driven Innovation

Research

AI & Automation

Logistics giant optimizes cloud data costs up front at speed & scale

Customers

CI/CD

Equifax Optimizes GCP Costs at Scale

Customers

Cost Opt & FinOps

The token bill is one line of your AI cost. [i]Unravel finds the rest — then cuts both.[/i]

AI spend is compounding — [i]and ownership hasn’t kept up.[/i]

The spend is scattered across a dozen bills.

A meaningful share of AI spend belongs to nobody.

Waste compounds quietly, with no guardrail to catch it.

AI cost and data cost [i]aren't two problems.[/i]

Behind every charge is a query, a job or a prompt. [i]That’s where the fix goes.[/i]

Prompt optimization, not prompt advice

Model right-sizing per call site

The job around the call

Fine-grained chargeback by consumer

A dozen invoices, a dozen units of measure. [i]One data model underneath them.[/i]

1. Ingest

2. Populate

3. Deliver

Stage 1 · Ingest — what we connect to

Databricks · Snowflake · BigQuery

Anthropic · OpenAI · Google · the long tail

Claude Code · Cursor · Codex · Copilot

Bedrock · Vertex · Azure AI Foundry