Prove the ROI of every AI dollar.
Larridin reads what AI costs from your invoices, and what it produced from your workflows and your engineering systems. Every department, one ledger, trended week by week.
AI ROI
What AI cost
“What did we spend on AI?”
$547K +6% vs prior 12 weeks
Seats, tokens and cloud models, priced from the invoices
The return
“What does an hour of AI work cost us?”
$31.65 per AI hour
spend ÷ human-equivalent hours of AI work
“Is it paying back?”
2.5× AI ROI
$1.35M net value in 12 weeks · hours at $110 loaded cost
What AI produced
“What did we get for it?”
1,440 hrs / wk back · 36 people-equivalents
Hours returned, measured from the work itself
“Which departments return the most?”
Compare all 7| Department | AI spend | Hours back / wk | Cost / AI hr | Adoption | Automatable | 12-wk trend |
|---|---|---|---|---|---|---|
| Engineering 310 employees | $237,120 | 520 | $38.00 | 84% | 14% | Increasing over 12 weeks |
| Customer Support 420 employees | $93,840 | 340 | $23.00 | 71% | 36% | Increasing over 12 weeks |
| Sales 380 employees | $86,400 | 240 | $30.00 | 58% | 24% | Increasing over 12 weeks |
| Finance 110 employees | $27,360 | 60 | $38.00 | 43% | 46% | Nearly flat over 12 weeks |
| Company total 2,000 employees | $546,960 | 1,440 | $31.65 | 58% | 23% |
“Where is the next return?”
Invoice processing and reconciliation
Finance · $164K a year · first of 195 ranked candidate workflows. 23% of observed effort is automatable, worth $3.6M a year in potential capacity.
See the ranked candidatesSpend is invoiced seats, token and API usage and cloud model calls, mapped to the department that used them. Hours back converts observed AI-assisted work and merged engineering output into human-hour equivalents. AI ROI values those hours at loaded cost, minus spend, over spend. Team level by default. Estimated capacity value, not realized cash savings.
Trusted by AI-forward enterprises
The budget review
The invoices are the easy half of AI ROI.
78% of IT leaders reported AI charges they never budgeted for (Zylo 2026 SaaS Management Index). The spend half of AI ROI arrives every month with a total at the bottom: seats, tokens, cloud invoices. The impact half arrives as anecdotes. Developers like Cursor. Sales uses ChatGPT for prospecting. Neither survives a budget review.
Larridin measures the impact half the way finance already measures the spend half: from records, not recollection. Observed workflows for every department. Merged code, CI runs and agent sessions for engineering. Both carry the spend that produced them, so the return on the slide agrees with the invoices in the drawer.
How AI ROI is measured
Measured from the work, priced from the invoice.
Three streams of evidence feed one ledger. Two measure what AI produced, one measures what it cost, and the ledger divides one by the other for every department, every week.
What AI cost
Observed workflows, every department
Engineering systems
Larridin
Prices
every dollar from the invoice, mapped to a team
Observes
workflows; AI-assisted vs non-AI runs of the same work
Scores
merged code as Engineering Output
Divides
spend by the hours AI gave back
The AI ROI ledger
Spend is priced from the invoice. Workflows are observed as metadata only, from admin-configured apps, at team level. Engineering is read from merged code, CI and agent sessions.
01 · What AI cost
Every place an AI dollar leaves the company, in one spend model.
Larridin connects to seat licences (ChatGPT Enterprise, Microsoft 365 Copilot, Cursor), provider invoices (Anthropic, OpenAI), cloud model calls (Bedrock, Vertex AI, Azure OpenAI), API gateways (LiteLLM, Portkey, Cloudflare AI Gateway), coding-agent sessions and browser plugins. Each source measures cost differently. Larridin normalizes them and maps every dollar to an owner, a team and a cost centre, with human and agent spend kept apart.
Priced from the invoice , so the total reconciles with finance. A developer on a personal Claude or ChatGPT plan still shows up, priced at list as estimated value and labelled as such. Unattributed spend, agents with no owner and seats idle for 30 days are flagged before the renewal, not after it.
Explore Spend IntelligenceSpend model
Lead qualification in the CRM
Mined steps of one AI-assisted run
Same workflow, with and without AI
AI-assisted runs
11.2 min per run
32 runs in 30 days
Non-AI runs
23.4 min per run
23 runs in 30 days
Verdicts compare AI-assisted runs against non-AI runs of the same workflow. Metadata only: apps, order, time per step. Team level by default.
02 · What AI produced, outside engineering
Sales, support and finance do not leave receipts. So Larridin observes the work.
A merged PR is a receipt for engineering work. The rest of the company has none, so a browser extension and a desktop agent record how work moves through the apps your admins enable: which app is active, in what order, how long each step takes, where the hand-offs happen. Metadata only. No screenshots, no keystrokes, no content, and every view rolls up to the team.
Larridin groups those sequences into workflows, such as lead qualification in Salesforce or month-end close, and notes when an AI tool was in the loop. It then compares AI-assisted runs against non-AI runs of the same workflow on duration and outcome. The difference, multiplied by how often the workflow runs, is hours returned. Each workflow gets a verdict: AI helps, AI hurts, or unclear until there is enough evidence. A time-and-motion study that runs itself , on metadata, every week.
One workflow, one sales team, 6.5 hours a month. The example company runs 847 workflows.
Explore Workflow Intelligence03 · What AI produced, in engineering
Engineering does leave a record. Larridin reads all of it.
Connect GitHub or GitLab, CI/CD and the coding agents your teams run (Claude Code, Codex, Cursor, Copilot). Every merged change is scored as Engineering Output: complexity-weighted, with penalties for missing tests and for AI slop, so a large AI-generated diff that gets rewritten next month is not counted as full progress. Added lines are attributed as AI-assisted, agent-authored or human. That share is AI Code Share.
Each agent session carries its share of the provider bill, cache reads counted apart. That yields Output per $1K of AI spend, the AI Quality Score and defect rate on AI-assisted code, PR cycle time against the team’s pre-rollout baseline, and AI ROI for engineering on its own. In the illustrative four-week example: 1,920 Output points for $28,400 of spend, 67.6 points per $1K, 4.8× estimated net ROI.
AI Impact
Engineering Output / $1K
67.6 pts +31% ↑
1,920 Output points per $28,400 of invoiced AI spend
AI Quality Score
92 / 100 +4 ↑
defect rate 2.1% · AI Slop Index 14, lower is better
Velocity impact
18h −38% ↓
median PR cycle, AI-assisted vs pre-rollout baseline
AI ROI, engineering
4.8× +0.6 ↑
$136K estimated net value over four weeks
Computed from invoiced spend, merged PRs, CI and session telemetry. Team level by default; individual views are coaching context for the engineer.
04 · The ledger
An hour of AI work that costs $31.65 and replaces one that costs $110.
Every hour AI gave back, whether from a sales workflow or a merged PR, lands in the same ledger as the spend that produced it. Spend divided by hours is cost per AI hour. Hours valued at loaded cost, minus spend, over spend, is AI ROI. The board shows both department by department, week by week, against the quarter before, so the slide finance sees on Thursday and the ledger they audit on Friday say the same thing.
Cost per AI hour
$546,960
AI spend, 12 weeks, priced from the invoices
17,280 hrs
hours AI gave back · 1,440 a week × 12
$31.65
per human-equivalent hour of AI work
AI ROI
(17,280 hrs × $110
hours back at blended loaded cost · $1,900,800
$546,960)
AI spend
$546,960
AI spend
2.5×
net value over spend · $1.35M net in 12 weeks
Loaded cost is set per department in your settings; the example uses one blended rate. Estimated capacity value, not realized cash savings. Billed spend stays separate from usage priced at list.
Read the measurement methodologyWhy both sides
Spend tools stop at the invoice. Engineering tools stop at the PR.
In the example company, engineering is 310 of 2,000 employees. A tool that measures AI from pull requests sees 15% of the workforce and none of the $310K spent outside engineering. A FinOps tool sees every dollar and none of the work. Process consultants map workflows from interviews, once, months after the fact, for $500K to $1M an engagement. Surveys record what people remember saving. Larridin is the only platform that reads the invoice, the observed workflow and the merged change together, for every department, every week.
| Measurement approach | Sees every AI dollar | Sees engineering output | Sees work outside engineering | Reconciles to invoices | Cadence |
|---|---|---|---|---|---|
| FinOps and spend tools | Yes | No | No | Yes | Continuous |
| Engineering analytics | Seats only | Yes | No | No | Continuous |
| Process mining and consulting | No | No | Interview snapshot | No | Once |
| Surveys | No | Self-reported | Self-reported | No | Quarterly |
| Larridin | Yes | Yes | Observed | Yes | Weekly |
What changes
Reallocate before the renewal.
Each figure on the board ends in a move, with the evidence one click away.
Reclaim the seats nobody uses.
38 of 240 ChatGPT Enterprise seats idle for 30 days, renewal in 41 days: the board shows it a quarter early.
Spend IntelligenceCoach the teams that adopted without fluency.
Sales at 58% weekly adoption and 4.8 fluency is a training cue for its managers, never a mark against a rep.
AI FluencyAutomate the next workflow by hours recoverable.
195 candidates, ranked. Invoice processing and reconciliation first, at $164K a year.
Workflow IntelligenceRoute engineering tokens to the model each task needs.
Cost per task $2.68 against $4.64 pinned, quality within noise.
Larridin RouterWho uses it
One question per seat, answered from the same evidence.
For the CFO
“What is the AI spend buying?”
Cost per AI hour and AI ROI by department, reconciled to the invoices you already have.
See Spend IntelligenceFor the CIO and Head of AI
“Where is the rollout working, and what do we automate next?”
Departments ranked by hours back, 195 candidates ranked by hours recoverable.
See Workflow IntelligenceFor the CTO
“What is engineering shipping with AI, and at what quality?”
Output per $1K, AI Quality Score, cycle time against baseline, agent effectiveness.
See Developer IntelligenceFor the CHRO and enablement leads
“Who needs help using it?”
Fluency by team and role, with the power users to learn from.
See AI FluencyAcross Larridin customers
47%
average lift in measurable team output
$2.4M
median AI spend optimized in year one
60 days
from rollout to board-ready impact reporting
1.4M
agent sessions in the Larridin benchmark
A customer perspective
“If you don’t know what people are actually using, you don’t know what to buy next.”
Larry Hill · Gainsight
Gainsight used Larridin to understand AI tool adoption and inform its first enterprise LLM purchase.
Deep dives
Every number on this page has its own page.
Frequently asked questions
How AI ROI is measured, answered first.
How do you measure the ROI of AI?
From both sides of the ratio. The denominator is AI spend, priced from your invoices: seat licences, token and API usage, cloud model calls and gateway traffic, mapped to the team that used them. The numerator is what AI produced: hours returned in observed workflows across every department, plus Engineering Output, quality and cycle time from your engineering systems. AI ROI is hours returned valued at loaded cost, minus spend, over spend. Cost per AI hour is spend divided by hours. Both are reported by department, weekly, against the prior quarter.
How do you measure AI impact outside engineering?
By observing the work rather than surveying people about it. A browser extension and a desktop agent record which apps are active, in what order and for how long, on the apps your admins enable. Larridin groups that activity into workflows, then compares AI-assisted runs against non-AI runs of the same workflow. The time difference multiplied by occurrences is hours returned, and every workflow gets a verdict: AI helps, AI hurts, or unclear. No pull requests required.
What counts as AI spend?
Everything that appears on an invoice for AI, and a labelled estimate for what does not. Seat licences such as ChatGPT Enterprise, Microsoft 365 Copilot and Cursor. Provider invoices from Anthropic and OpenAI. Cloud model calls through Bedrock, Vertex AI and Azure OpenAI. Gateway traffic through LiteLLM, Portkey and Cloudflare AI Gateway. Coding-agent sessions from Claude Code, Codex and Cursor. Personal subscriptions are priced at list as estimated API value and kept separate from billed spend.
Do the numbers reconcile with finance?
Yes. Spend is allocated from the invoices finance already has, so the total on the board equals the total in the ledger. Where Larridin estimates rather than invoices, such as a personal plan priced at list, the figure is labelled as an estimate and reported separately. Cache reads and cache creation are priced apart so provider bills match to the line.
What is cost per AI hour?
AI spend divided by the human-equivalent hours of work AI performed in the same period. In the example company it is $31.65: $546,960 of spend over twelve weeks against 1,440 hours a week. Compare it with the loaded cost of the human hour in each department and the return is visible without a model.
Does Larridin measure AI agents as well as people?
Yes, and it keeps them apart. Every agent is traced to an owner, a team and a cost centre. Sessions and spend are split human versus agent, so you can see that 27% of sessions and 72% of variable spend are agent-driven in the example. Agents with no accountable owner are flagged as agent debt before they run unchecked.
Is this employee surveillance?
No. Larridin maps the work, not the worker. Only metadata is collected, only from apps an admin configures, and every view rolls up to the team by default. Message bodies, document contents, screen recordings and keystrokes are never collected. Where individual views exist, they are coaching context for the person and their manager, and they are off unless an admin turns them on.
How is AI ROI different from AI adoption?
Adoption is who used AI this week. Fluency is how well they used it. Impact is what changed in the work because they did. ROI is that impact, valued, over what it cost. Larridin reports all four because a department can score high on adoption and low on return, and the fix for each is different.
How long until we have a number?
Connect the spend sources and repositories on day one. The first spend and engineering report lands on day 14. Workflow verdicts arrive within weeks of the sensors rolling out, as workflows accumulate enough runs to compare. Board-ready impact reporting, benchmarked against 1,000+ engineering teams, is the 60-day milestone.
Walk into the budget review with the return already computed.
Connect your spend sources on day one. First report on day 14. Board-ready in 60 days.