The measurement layer for AI spend

You spent millions on AI. Which tools are actually earning it?

Every vendor's dashboard says their tool is essential. Surveys tell you what people think you want to hear. Pyx measures what actually got done — the same people, before and after each tool — so you know which AI to keep, shrink, or cut.

Book a demo hello@pyx.legal
Illustrative · one operating company, one quarter

The board keeps asking. Now there's an answer that traces to a person, a tool, and a matter.

Recoverable this quarter
$847,000
across 47 findings — 22% behind peer benchmark
47 FINDINGS · 78% HIGH OR MEDIUM CONFIDENCE · 184 LAPTOPS · $12.4M ANNUAL AI + SAAS SPEND
Kill
$527K
3 tools · net-negative or unused
Kira · Casetext · LegacyDocAI
Shrink
$337K
2 tools · too many seats
CoCounsel · Gong AI
Negotiate
2 tools
counter-proposal packets drafted · renewals within 90 days
SFDC Einstein · Ironclad

Every dollar traces to a person, a tool, and a matter.

The proof

Same person. Same work. Before and after each tool.

No vendor dashboard, no survey, no screen monitor. We compare the same people doing the same kind of work — before a tool went in, and after — and measure what actually got done.

✓ Keep · net positive · $0 to recover

Cursor · engineering

218 min168 min23% faster
48 engineers. Same people, same type of work — 218 matters before Cursor, 264 after. Code quality, measured after merge, unchanged.
✗ Shrink · net negative · $187K/yr recoverable

CoCounsel · legal ops

142 min238 min1.6 hrs slower
$187K/yr in extra people-time, on top of the $294K license. Keep the 12 identified power users; cut 33 marginal seats.

Vendor dashboards are biased. Surveys are unreliable. Screen-time trackers are invasive and measure the wrong thing. Engineering-metrics platforms only see engineers. Nothing else measures the work itself, before and after, for everyone.

Per person · the input to everything else

Every person. Every tool. What got done — not time at the keyboard.

184
People measured · 88.5% opted in
3.2×
Top 10% vs median output
0.58
Bottom 10% · dragged by one tool
−124 → +186
Net minutes per matter

Managers see their team. Team leads see their tools. The CFO sees the aggregate. Every number rolls up from the same per-person measurement — so the decision to roll a tool out, hold it back, or pull it from one team is made on what actually got done.

Kill · shrink · negotiate

The CFO gets a list. Procurement gets the receipts.

ToolActionImpact / yrWhy
Kira SystemsKill$396KPeople work at half the speed of peers using it. Almost nobody actually uses it.
CasetextKill$84KNet-negative, and deprecated by the vendor anyway.
LegacyDocAIKill$47KZero uses in 90 days. Dormant seats.
CoCounselShrink$250KAdds cleanup time per use. Only 12 identified power users are net positive; cut the rest.
Gong AIShrink$87KOn a third of summaries, the same person invokes the AI and "reviews" it within minutes. That's not a real review.
SFDC EinsteinNegotiateTarget −30%Renewal inside 34 days. Counter-proposal packet ready.
Ironclad PlaybookNegotiateTarget −20%Renewal inside 62 days. Usage supports a smaller commitment.

Every recommendation carries its defense trail — the person, the tool, the work. Export the list, send it to the CFO, hand procurement the packet.

Watch it work

Signed records. Checked four ways. Tamper-proof log.

Each time someone uses an AI tool, Pyx writes one signed record — sensed, resolved to a matter, recorded, signed, checked, and appended to a log that can't be edited after the fact.

18,204
Records this quarter
100%
Verified
18,204
Log entries
184
Signing keys
Live pipeline
In production this runs on each person's own laptop, not on this page — same code path. This preview fires plausible records with live timestamps.
AI-tool usesCaptured
Work-product eventsCaptured
Time at deskBlocked in code
Keystroke loggingBlocked in code
Screen contentsBlocked in code
Per-minute session trackingBlocked in code
The moat

Four things at once. Nobody else has all four.

01 · Buyer-side

Runs on the buyer's own laptops.

No AI vendor can block it, throttle it, or shade the numbers in their favor — because it doesn't depend on any vendor's data.

02 · Patents

Patent-pending.

Seven provisional applications filed with the USPTO on how the work is captured, signed, and measured.

03 · Consent

Employees will actually accept it.

No time-at-desk, no keystrokes, no screen capture — those are blocked in code. It measures the work, not the person.

04 · Deterministic

No AI in the analytics.

The tools being measured guess. The measurement does not. The same evidence produces the same number every time.

Built to sit underneath the platforms firms already run — not another dashboard on top of self-reported time.

See it on your own AI stack.

Book a demo