AI usage intelligenceFor CEOs, CTOs, and whoever signs the AI bill

Do you know
how your company
uses AI?

Metermark shows you what your AI spend actually buys, across every tool your people use. Who gets the most from it, what kind of work it does, and where the rest of the company can catch up. No transcripts stored, ever.

One view acrossClaude CodeCursorGitHub CopilotChatGPT EnterpriseClaude.aiCodex CLIGemini CLIAWS BedrockGoogle VertexAzure OpenAI

Adoption is no longer the story. Value is.

More than 90% of developers use AI every day. Vendor dashboards only see their own tool. Engineering-intelligence tools only see engineers. Nobody can tell a CEO what all that usage is actually doing.

$37B

Enterprise gen-AI spend in 2025, up 3.2× year over year. The money is committed. The accountability question is arriving on schedule.

Menlo Ventures

95%

of gen-AI pilots show no measurable P&L return, despite $30–40B invested.

MIT, State of AI in Business 2025

<33%

of decision-makers can identify a specific financial outcome from their AI investments.

Gartner

~40%

of employee AI time savings are lost to rework. Only 14% see consistent net-positive outcomes.

Workday, January 2026

Your #2 AI spender is probably a product manager.

And no engineering dashboard can see them. In most companies the heaviest AI users sit in product, design, marketing, and operations, where none of today’s tooling looks.

The gap is cross-vendor, semantic, and cross-functional.

Security tools read everything and sell fear. Vendor dashboards are free and single-vendor. Engineering tools stop at the engineering org. The question every executive is about to be asked, what did we get for it?, needs a neutral layer that sees every tool, understands what the work was, and covers every kind of worker.

Know what your AI spend buys.

Metermark joins the usage every AI vendor already reports into one cross-vendor account of spend, work, and outcomes. Three parts, none of them in your critical path.

Collect01

Every tool, one ledger

Lightweight collectors tap the telemetry your AI tools already emit — Claude Code, Cursor, Copilot, ChatGPT Enterprise, Bedrock and more — and normalize it into one record per session. Who, when, which tool, which model, what it cost, what shipped.

Classify02

Understood inside your boundary

Prompts are read where they already live, on your machines or in your cloud, labeled by the kind of work they represent, and discarded. Only the labels leave. About 2 KB per session; a 500-person company produces megabytes a day, not terabytes.

Answer03

Never in the request path

No gateway, no proxy, nothing between your people and their tools. No latency story, no uptime story, nothing to break on a Tuesday. Metermark reads what already exists and answers in plain English.

The record

One record per session. Fixed spine, extensible leaves.

The spine — identity, time, cost, outcomes — never changes, so records stay comparable across vendors and later power benchmarking. The leaves are yours: your project mapping, your definition of work-related, your labels, your custom fields. Corrections made in the dashboard feed back into the classifier, so accuracy is yours too.

who
user (SSO), team
when
start, duration, active time, turns
where
harness · surface · auth mode
what
models, usage in/out/cached, cost, tool mix
context
repo or project, languages, branch
derived
kind of work · topics · work-relatednesscomputed at the edge, content discarded
outcome
commits, PRs, lines added and removed
custom
namespaced, typed, size-capped fields

The questions you can finally answer.

Classifying what people ask AI to do — never storing what they said — turns a usage bill into a map of how your company actually works.

Spend by kind of work

Last 30 days · every tool · one company

Sample
  • Building features34%
  • Fixing bugs19%
  • Research14%
  • Writing & docs11%
  • Product planning9%
  • Data analysis7%
  • Learning4%
  • Other2%
Semantic coverage: 91% of sessionsSpend and activity only: 9%
Power users

Who gets the most out of AI, and what do they do differently?

Find the people whose AI sessions actually ship. See how they work — how they brief, how they iterate, which tools they reach for — and turn it into a playbook for everyone else.

Upleveling

Which teams are AI-native, and which need a hand?

Spot the teams still using AI like a search box. Give them a concrete path to how your best people work, and watch the gap close quarter over quarter.

Spend attribution

What did the money buy, by team, project, and kind of work?

Features, fixes, research, documents, planning. Across Claude, Cursor, Copilot, and ChatGPT. One ledger, one number, and the ability to defend it in a renewal review.

Everyone

What is the rest of the company doing with AI?

Product, design, marketing, ops, finance. The people engineering tools can't see are often your heaviest users, and your biggest upside.

Coverage

What can we see, and what can't we?

Every gap stated plainly. “91% of your AI usage has semantic coverage” is a real number. “Total visibility” isn't, and we won't claim it.

Ask

Which team spends the most on research?

Ask in plain English. Metermark answers with the numbers and shows the query it ran, so finance and engineering argue about the same figure.

Every tool. Honestly labeled.

Coverage is a first-class feature. Each surface is qualified by what its vendor actually exposes — per auth mode, not per logo — so the number on your dashboard is one you can repeat to your board.

SurfaceHow it’s readDepth
Claude CodeClient telemetry, enforceable through managed settings. The only per-user path on Bedrock and Vertex.Kind of work + spend
Claude.ai & CoworkCompliance API on Enterprise plans.Kind of work + spend
ChatGPT EnterpriseCompliance logs platform, ingested continuously.Kind of work + spend
Codex CLITelemetry via config, fleet-manageable.Kind of work + spend
Gemini CLI & appTelemetry via settings; Google Vault export.Kind of work + spend
AWS BedrockInvocation logging into your own S3 or CloudWatch.Kind of work + spend
Google VertexRequest-response logging into BigQuery.Kind of work + spend
CursorAdmin API: per-event usage, cost, model, user. Content-blind by design.Spend + activity
GitHub CopilotMetrics API: per-user daily activity. No content path on any admin surface.Spend + activity
Azure OpenAIDiagnostic settings into Log Analytics.Spend + activity
Windsurf, Devin, AmpEnterprise analytics APIs; shared threads where admin-visible.Spend + activity
Claude DesignAdmin toggle with active-user counts only. A gap we disclose rather than paper over.Presence only
Work classified at the edge, spend priced from real usage Spend and activity from admin APIs, no content available Vendor exposes nothing beyond seat counts today

Verified against live vendor documentation, September 2026. Vendors change; the map is versioned and re-verified.

No transcripts stored. Ever.

Built for enablement, not oversight. People who feel watched leave, and regulators agree: the EU AI Act treats workplace AI monitoring as high-risk. The architecture is the promise.

0

Transcripts stored. Prompts retained. Responses kept. Metermark never receives the content, so it can never be asked for it, subpoenaed for it, or breached for it.

Classified at the edge, content discarded.

Prompts are read inside your boundary, labeled, and dropped. What leaves is a handful of words per session, and the label is barred from quoting content, names, secrets, or links.

Aggregate by default.

Team-level views first. Person-level detail is role-gated, audit-logged, and switched off below a minimum group size. Shipped as admin policy; offered to legal as a feature.

Your data stays in your store.

Want to reclassify later with new labels? Keep redacted content 30 to 90 days in your own bucket and the job runs there. Compute goes to the data. The sensitive asset never lands on our books.

Nothing in the request path.

No gateway, no proxy, no agent sitting between a person and their tool. We can't read what we never receive, and there is nothing of ours to go down.

Priced per covered user. Paid for by what it finds.

At a 300-person company with 150 AI users, Growth runs about $27k a year, roughly 15% of the AI spend it explains. The goal is for Metermark to pay for itself out of the misattributed spend it uncovers.

Free
$0

up to 20 covered users

  • One team, one collector
  • 30 days of history
  • Spend across the tools it sees
  • The coverage map

Adoption counters are free everywhere. The understanding starts at Growth.

Start free
Growth
$15

per covered user / month · $12k minimum

  • Every collector, every vendor
  • Kind-of-work classification at the edge
  • Cross-vendor rollup and spend attribution
  • Power-user and team upleveling views
  • Questions in plain English

Guided self-serve. An IT admin finishes rollout in a day or two from the runbooks.

Get an assessment
Enterprise
$25–30

per covered user / month

  • Compliance-API sources: ChatGPT Enterprise, Claude Enterprise
  • Classifier deployed inside your VPC
  • Bring-your-own-bucket retention
  • SSO, SCIM, audit log, API export
  • Hands-on rollout

For cloud-log and compliance-grade sources, deployed with you.

Talk to us

Get a free AI usage assessment.

Point Metermark at one team for a week. You get a report of what your AI spend bought, who your power users are, and where your coverage gaps live. No transcripts stored. No contract.

One team, one week, one report. We reply from a human, not a sequence.