NEW · v2.4PR-level AI attribution shipped Nov 14

Measure the ROI of AI in your
engineering org
.

Devology attributes every line of code to AI or human, joins that with what each AI tool actually costs, and tells you what your $/PR merged, $/story point, and bug ratelook like — across Cursor, Copilot, Claude Code, Codex, and Windsurf.

SOC 2 Type IISelf-hosted availableRead-only by default
platform/api · PR #4,182 · 2h agoPROVENANCE
retry-with-backoff.ts+87 / −14 · 4 files
// Co-Authored-By: Claude <noreply@anthropic.com>import { sleep } from './utils';export async function retryWithBackoff<T>(  fn: () => Promise<T>,  opts = { max: 5, base: 200, cap: 8000 }): Promise<T> {  let attempt = 0;  while (true) {    try { return await fn(); }    catch (err) {      if (++attempt >= opts.max) throw err;      const delay = Math.min(opts.cap, opts.base * 2**attempt);      await sleep(delay + Math.random()*100);    }  }}
AI · 11 linesHuman · 3 lines
claude-3.5-sonnet · $0.34
Measured across 2,400+ engineering orgs
LinearVERCELPlaidRampMERCURYRetoolsupabasePostHogLinearVERCELPlaidRampMERCURYRetoolsupabasePostHog
THE GAP

Your finance team sees the AI bill.
Nobody sees what it shipped.

You're paying for Cursor, Copilot, Claude Code, Codex, and Windsurf seats and tokens. But none of that data is joined to your repo. So when the CFO asks "is this working?" the honest answer is "we don't know — it feels faster."

Devology closes the loop. We attribute every commit, line, and PR to a human or an AI assistant — then divide the actual token bill by what shipped.

AI spend, last 12mo
+412%
Average across surveyed orgs
Orgs with attribution
6%
Can answer $/PR merged
% of code AI-touched
58%
Median, our customer base
Bug rate, AI vs human
1.14×
Indistinguishable, with review
ONE DASHBOARD · NEUTRAL · MEASURABLE

Provenance, attribution, ROI — joined.

Devology ingests AI tool usage logs, joins them with your Git history, and produces a single picture of what AI is actually doing in your codebase. No judgment, no slop-shaming — just the numbers.

devology.ai/platform-team/overview
live data

AI in engineering · Q4 2025

Oct 1 – Dec 31
$/PR merged
$184−44%
$/story pt
$41.20−18%
% AI-authored
62%+8pp
Bug rate · AI vs hu
1.04×parity
Code provenance · weekly
% of merged lines, by author type
AI-authoredAI-assistedHuman only
100%75%50%25%62% AI · this weekW36W40W44W48
Spend by source
AI tokens, last 30 days · $52,481
Cursor$24,180 · 46%
Claude Code$14,902 · 28%
GitHub Copilot$8,140 · 16%
Codex / OpenAI$3,621 · 7%
Windsurf$1,638 · 3%
Tokens · 30d
412M
PRs merged · 30d
284
HOW IT WORKS

Three signals. One ground truth.

We don't ask your engineers to label anything. Devology reads provenance directly from three independent signals and reconciles them into a single confidence-scored verdict per line.

01Tool API

Direct usage logs from each AI tool

We integrate read-only with Cursor, Copilot, Claude Code, Codex, and Windsurf. Every prompt, completion, and accepted suggestion comes with author, timestamp, model, and exact token count.

eventcompletion.accepted
toolcursor@0.42
modelclaude-3.5-sonnet
tokensin 2,418 · out 184
fileapi/retry.ts:18-29
View source schema
02Git metadata

Co-author trailers & commit fingerprints

Every commit is parsed for Co-Authored-By trailers (Claude, Copilot, Aider, …) plus stylometric fingerprints — keystroke cadence, comment density, function arity — that flag AI authorship even when trailers are stripped.

commit 7a4b2c9
author jdimov <j@…>
+ Co-Authored-By: Claude
cadence 3.2k LOC · 1.8s/line · burst
verdict AI-assisted · 0.94 conf
How fingerprinting works
03PR analysis

PR-level diff classification at open time

When a PR opens, Devology classifies each hunk: AI-authored, AI-assisted (edited by human), or human-only — and ships an inline label so reviewers know what they're reading.

retry-with-backoff.ts11/14 lines AI
payments.ts2/47 lines AI
CHANGELOG.mdhuman only
See the GitHub check
SOURCES

Every AI coding tool your team uses.

Direct integrations with the five biggest AI coding sources, plus a generic webhook for in-house assistants and CI bots.

Read-only by default · no credentials leave your VPC
Backfill 24 months of historical token + Git data on connect
Self-hosted edition for regulated environments
CursorNative
Completions · Composer · Tab · Agent runs
Token + accept rate per file
GitHub CopilotNative
Inline · Chat · Workspace · Edits
Org + seat-level usage logs
Claude CodeNative
Terminal sessions · Tool calls · MCP
Per-prompt token attribution
Codex / OpenAINative
CLI · API · Codex tasks
Org usage + dashboard ingest
WindsurfNative
Cascade flows · Edits · Inline
Per-session token + diff
Custom / WebhookOpen
In-house assistants · CI bots
POST /v1/events · SDKs
Git sourceGitHubGitLabBitbucketSelf-hosted GitLinear · Jira · Shortcut for story-point joins →
THE ROI MODEL

What you actually want to know.

Token spend and PR throughput are easy to count separately. Joining them is the hard part. Devology does it per author, per team, per project, every day.

$/PRAI tokens spent on a PR ÷ PRs merged in window. Compared against humans-only baseline.
$/ptTokens ÷ story points shipped (Linear / Jira / Shortcut joined).
bug Δ30-day bug rate on AI-authored hunks vs human-authored hunks, in the same file.
rev/$Optional · attributed feature revenue per AI dollar, if you wire up product analytics.
ROI calculator · per quarterindicative
Quarterly AI bill
$52,800
Engineering output (est)
858 PRs
$/PR merged
$62
Visibility before Devology
$0
Assumes 50 productive workdays / quarter, baseline 2.2 merged PRs per AI-active engineer per week. Configurable per team in-app.
FOR REVIEWERS

A check on every PR.
Visible to the people who matter.

Devology posts a GitHub / GitLab check on each PR with the provenance breakdown, the token cost, and the model used. Reviewers see what they're reading. Managers see what was actually shipped.

All checks have passed12 successful checks
LDevology/ provenance
62% AI · 38% human · $0.34 · claude-3.5-sonnetDetails
By file
retry-with-backoff.ts+11 AI
payments.ts+2 AI
CHANGELOG.mdhuman
By source
Claude Code$0.28
Cursor$0.06
Copilot$0.00
Quality signals
Tests added+4
Coverage+1.2pp
Similar PRs · bug rate0.9×

We were spending almost $90k a quarter on Copilot, Cursor, and Claude Code seats. The board kept asking what we were getting back. Devology gave us the first honest number — $71 per merged PR, 1.06× bug rate vs human. That ended the debate.

Priya Anand
VP Engineering · Plaid
−$2.1M
annualised, vs benchmark

"Finance finally stopped treating AI tools like a marketing line item. We can show $/PR and $/story point on the same page as Slack and Notion."

Marco Reyes · Director of Eng Ops, Ramp

"The PR check is the killer feature. Reviewers know what's AI before they start reading. Our review time on AI-heavy PRs dropped 31%."

Ada Chen · Staff Platform Eng, Mercury
PRICING

Priced like an observability tool, not a percentage of your AI bill.

Per seat. Predictable. No clawback if your team uses AI well.

Team
$0/ eng / mo

Up to 10 engineers · single repo

Start free
  • Provenance check on every PR
  • $/PR merged · $/story point
  • 2 AI sources connected
  • 30-day data retention
  • ·SSO / SCIM
MOST POPULAR
Business
$18/ eng / mo · billed annually

Up to 500 engineers · unlimited repos

Book a demo
  • Everything in Team
  • All 5 AI sources + custom webhook
  • Bug-rate attribution (AI vs human)
  • Linear / Jira story-point joins
  • SSO / SCIM / audit log
  • 24-month retention · CSV export
Enterprise
Custom

Self-hosted or VPC · unlimited everything

Talk to sales
  • Everything in Business
  • Self-hosted edition
  • Custom provenance rules
  • Dedicated CSM + quarterly review
  • SOC 2 Type II report · DPA · MSA
  • 99.95% uptime SLA

Stop guessing whether AI is paying off.
Start measuring it on Monday.

No CC required10-minute setupRead-only access