How to Track Your Team's AI Tool Spend, Adoption, and ROI

Your team's AI tools have quietly become a real line item. Cursor, Codex, Claude Code, Replit, ChatGPT, Claude, Gemini, and more: a seat here, a usage bill there, multiplied across teams and orgs, and in larger companies the whole enterprise.

The thing keeping you up at night is not what one engineer pasted into ChatGPT. It is that nobody can say, with a straight face, what all of this costs, who is really using it, or whether the spend is paying for itself.

Sometimes this gets filed under "tracking employee AI usage." That phrase is worth untangling, because it usually means something else. If your goal is prompt monitoring, keystroke capture, or stopping someone from pasting a customer list into a chatbot, that is a security and data-loss problem. It belongs to a separate category of monitoring software, a different market from what this post is about.

This post is about the other side: what your team's AI tools cost, who actually uses them, and whether they are paying off. You can answer all of it from metadata, without reading a single prompt.

The question really splits into three: what are we spending, who is actually using it, and is the spend producing faster, safe work? It is a common struggle. In McKinsey's 2025 State of AI survey, more than 80 percent of organizations reported no material enterprise-level impact on profit from generative AI yet, and the practice most strongly tied to bottom-line impact was simply tracking well-defined KPIs for it. The spend is real. The visibility is not, because it sits across a dozen vendor consoles that do not talk to each other.

Tracking your team's AI usage really means answering three questions, each with a different owner:

Question	What you measure	Who cares
Cost	What you spend on AI tools across the org, by team and person	Finance, engineering leadership
Adoption	Who actually uses each tool, how often, and how deeply	Engineering leadership, enablement
ROI	Whether that spend produces faster, safe output	Leadership, finance, the board

Governance and security (shadow-AI risk, data leakage, prompt policy) is a separate track, with its own tools and its own owner. It answers a different question. Conflate it with the three above and you end up buying a surveillance product to solve a budgeting problem. This guide stays on cost, adoption, and ROI.

The catch is that a seat count only measures what you bought. It says nothing about who logged in, how often, or what came back. And the seat is only half the bill: most of these tools also charge for consumption, the tokens and agent runs metered on top of the subscription, which can dwarf the flat fee and never shows up on the seat line. Tracking your team's AI spend means following both, the per-seat subscription and the usage stacked on it.

Idle and underused licenses are a well-known SaaS problem, and AI tools are no exception. Gartner estimates that organizations without centralized SaaS management overspend by at least 25 percent, on unused entitlements and overlapping tools. The flip side is how much is recoverable once you actually look: Snowflake built an internal tool to find and reclaim idle licenses and reported $5.5 million in cost avoidance in its first year. A seat nobody opens still renews every month, and that waste rarely shows up on the invoice as a problem.

Quick answer: to track your team's AI usage in a way that actually helps, skip the surveillance and measure three things.

Cost: pull every AI tool invoice plus each admin and billing API into one view, broken down by team and person.
Adoption: track daily active users and cohorts (power, casual, new, idle), not seats purchased.
ROI: compare spend per engineer to throughput against a pre-AI baseline, and watch rework and revert rates.

The rest of this guide is the step-by-step, with the metrics and the tools for each.

Seats tell you what you bought. They tell you nothing about whether anyone used it, or what it was worth.

Step 1: Inventory every AI tool your team actually uses

Start with what you pay for, then with what you cannot see. This step finds the sources. It does not yet tell you what they truly cost.

The tools you already pay for are the easy half: they show up in expense reports and each vendor's admin console. Shadow AI, the tools people put on a personal card or use on a free tier, shows up nowhere, until a reimbursement or a data-loss scare surfaces it.

To catch both, reconcile two sources you already have: finance and expense data, and each vendor's billing or admin console. If you run a single sign-on provider, its app list is a useful third signal. Anything in expenses but missing from your approved list is worth pulling in.

The output is a simple inventory, a list of every AI tool, its owner, and seat count, with the list price next to each. That list price is only the subscription, the per-seat fee that lands around $20 to $40 for standard plans, with heavier tiers above it (Cursor Teams is $40 for a Standard seat and $120 for a Premium one; chat plans like ChatGPT Business run about $20 to $25 per seat, Enterprise priced custom). Most of these tools also bill for consumption on top of the seat, agent runs, premium models, and API calls metered by tokens, and none of that shows on the rate card. So the inventory tells you what to track and roughly what it lists for, not what it actually costs. Pulling the real spend together is the next step.

Step 2: Unify cost and usage in one view

With the inventory in hand, connect to each tool and pull the real numbers, not the sticker price.

AI spend comes in two shapes. There are per-seat subscriptions, the flat monthly fee you already listed for tools like Cursor, ChatGPT, Claude, and Gemini. And there is consumption-based usage, billed by tokens or runs: Cursor usage past its seat allowance, Codex and Claude Code agent runs, and premium-model API budgets.

The second kind is the one that surprises people. Agentic workflows can use 20 to 50 times more tokens than autocomplete, so one behavior change becomes a step up on the bill, and none of it shows on the subscription line.

Each tool already gives you a dashboard with some of this. The catch is that the breakdown stops at its own walls and its own level of detail: one console slices spend by model and project, another shows little more than a monthly total, and none of them line up with each other. The granularity is rarely the exact slice you want, and you are still checking every tool separately, in its own format.

So the real work here is unglamorous: pull every tool's numbers into one place, on the same definitions, instead of opening a dozen consoles at month end. That unified view is the first number you can actually trust. Slicing it further, down to a team, person, product, or customer, comes in Step 4.

This is where a dedicated cost layer earns its place: one tool that connects to each provider's billing and usage API and reads every tool's spend into a single, normalized view, so you check one place instead of a dozen dashboards. Because it works by reading APIs out of band, it stays out of your request path and adds no latency to anything your team runs. (The same approach covers the AI spend inside the product you ship, which we cover in the best LLM cost tracking tools.)

Step 3: Measure adoption, not seats

With spend and usage in one place, you can finally ask who is actually using these tools. A seat is a license you bought. Adoption is whether anyone uses it. The gap between them is where the money quietly leaks.

So measure active use, not headcount with a license assigned: daily active users, frequency, and depth. Then segment, because one org-wide number hides everything useful. Sort people into cohorts: power users (most of their pull requests AI-assisted), casual users, new adopters, and idle holders who tried a tool once and stopped.

As a rough internal target, daily active use in the 40 to 50 percent range is healthy, and under 30 percent usually means you are paying for shelfware. Set the exact bar yourself; what matters is tracking active use over time, not seats assigned. Each cohort then needs a different move: reclaim idle seats, train the casual users, and learn what the power users do differently. None of that is visible from a seat count.

Step 4: Attribute spend to teams, people, products, and customers

Total spend is a start. Attribution is the answer. "We spent $40K on AI last month" is a number nobody can act on. The useful version always names a who and a what.

What you attribute to depends on the tool. For coding assistants, the natural unit is the engineer and the pull request. Cost per engineer and cost per merged PR put the spend right next to the work it produced, instead of leaving it to float in a billing report.

For product and API spend, the questions are different but just as concrete:

Which team owns the API, so the bill has an accountable owner instead of landing in a shared pool.
Which product or feature is driving the calls, so you know what you are actually paying for.
Which customer consumes the most, so you can see cost per customer and catch the accounts that cost more to serve than they bring in.

The goal is the same either way: every dollar of AI spend should trace back to a team, a person, a product, or a customer. Spend you cannot attribute is spend you cannot manage.

Step 5: Measure ROI, not usage

Usage is a vanity metric. Prompts per day and suggestions accepted feel like progress and prove nothing. ROI is the question finance will ask, and it is fair: in IBM's 2025 CEO study, only about 25 percent of AI initiatives had delivered the ROI expected of them.

No single metric survives an audit, so combine three:

Spend per engineer per month, the input.
Throughput against a pre-AI baseline, the output (DORA-style delivery metrics work well).
Rework or revert rate on AI-influenced pull requests, the quality check.

Picking one invites pushback. Together they form a story finance can evaluate.

One trap: do not trust perception. In METR's 2025 randomized controlled trial, developers predicted AI would make them 24 percent faster, were measured 19 percent slower, and still believed it had sped them up by 20 percent afterward. You cannot read ROI from how it feels, so capture a real baseline before you scale.

Step 6: Set budgets, alerts, and a monthly review

Tracking is a loop, not a one-time audit. Three controls keep it honest:

Per-team budgets, not one shared pool. A budget with a named owner creates accountability a shared pool never will.
Anomaly alerts. A 20-to-50x token jump, when a team flips into agentic mode, should reach you mid-month, not as a line on the invoice.
A short monthly review. Fifteen minutes with engineering leadership: idle seats reclaimed, the cost-per-PR trend, how cohorts moved.

The cadence is what turns a one-off cleanup into sustained efficiency.

Do this without surveilling your team

Most "AI monitoring" pitches get this backwards. Every question above (cost, adoption, ROI) is answered from metadata, not content: seat assignment, active use, token counts, model, and pull request links. The hard part is not reading prompts; it is pulling that metadata out of every provider, normalizing it, and tying it back to the right team, person, product, or customer. That is the real work, and it never has to touch a prompt, completion, or keystroke.

The invasive version is not just a morale cost. It is a compliance liability, and it does not even produce better cost or ROI numbers.

That is the whole thesis: track outcomes, not keystrokes. Data minimization is where regulation is heading anyway, and it is the lighter, more accurate way to get the numbers you need. SuperPenguin defaults to not storing prompt or response text, or API keys. Cost, attribution, and ROI all work without touching content. Teams that want prompt-level optimization can opt in to sampled, encrypted captures, but that stays off unless you turn it on.

Where to start

You do not have to do all six steps at once. Start where your spend and your hardest questions are.

If your situation is	Start here
One engineering team, a few tools	Right-size seats first: pull daily active users, reclaim the idle ones, then baseline cost per PR.
Multiple teams, finance starting to ask	Unify cost across every tool plus consumption, attribute by team, and start the monthly review.
The board wants the ROI number	Combine spend per engineer, throughput against a baseline, and rework rate; never ship one alone.

Most of this you can assemble from vendor consoles and a spreadsheet, and for a small team that is a perfectly good place to start. It stops scaling the moment more than one team is spending, the tools multiply, and finance wants a single number instead of eight browser tabs. That is the gap SuperPenguin was built for: it brings AI spend across 14 providers into one view, attributes it down to the team, person, and pull request, and computes cost per PR so the ROI of your team's AI coding is a number rather than a vibe, all without storing prompts or keys. If the engineering slice is where your spend and your hardest questions live, try SuperPenguin free.

Sources

Team Pricing, Cursor.
ChatGPT Pricing, OpenAI.
The State of AI: How organizations are rewiring to capture value, McKinsey, 2025.
IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, IBM Institute for Business Value, 2025.
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, 2025.
Magic Quadrant for SaaS Management Platforms, Gartner, 2025.
Software Asset Management: Predict Your SaaS Needs, Snowflake.

Frequently asked questions

What does it mean to track employee AI usage?

Two different things. One is surveillance (watching what people type, flagging risky prompts): a data-loss problem for a separate category of tools. The other, and the one most leaders actually need, is the business side: which tools your team uses, what they cost, and whether the spend pays off. You can answer that from metadata, no keystrokes required.

How do you measure AI tool adoption across a team?

Count active use, not seats: daily active users, frequency, and depth, split into cohorts (power, casual, new, idle). As a rough internal target, 40 to 50 percent daily active use is healthy and under 30 percent usually means shelfware to reclaim. Track it over time.

How do you measure the ROI of AI tools for employees?

Combine three numbers: spend per engineer (input), throughput against a pre-AI baseline (output), and rework or revert rate on AI-influenced pull requests (quality). Capture the baseline first, because perception lies: in METR's 2025 trial, developers were measured 19 percent slower yet still felt 20 percent faster.

Can you track employee AI usage without monitoring or surveillance?

Yes, and usually you should. Cost, adoption, and ROI all come from metadata: seats, active use, token counts, model, and pull request links, never prompt content. SuperPenguin defaults to storing no prompt or response text and no API keys, so the numbers work without touching what anyone typed.

How much do AI tools cost per employee?

No single number; pricing has two layers. Seats are the floor (coding tools roughly $20 to $40, Cursor Teams $40 Standard or $120 Premium; chat plans like ChatGPT Business about $20 to $25). On top, the big labs bundle coding agents into existing subscriptions (Codex, Claude Code, Gemini), each adding usage costs. That consumption layer is the surprise, so the seat is only where the bill starts.

What is shadow AI and how do you find it?

The AI tools employees use that IT and finance never approved or can see, usually on personal cards or free tiers. Find it by cross-referencing expense and reimbursement reports against vendor admin consoles (and your SSO app list if you have one). Anything in expenses or logins but not your approved inventory is shadow AI.

What metrics should engineering leaders track for AI coding tools?

Three things: adoption (daily active users and cohort mix), cost (spend per engineer, normalized across every tool and its usage-based costs), and outcome (throughput against a baseline, plus rework and revert rates). For engineering, cost per pull request ties spend directly to shipped work.

How do you measure the ROI of Cursor or another AI coding assistant?

Tie cost to output. Track cost per engineer and cost per pull request for Cursor (or Codex or Claude Code), compare throughput to a pre-AI baseline, and watch the rework rate on AI-assisted PRs. Cost per merged PR is the cleanest single number; SuperPenguin computes it by tying Cursor spend to the PRs it helped ship.

When do you need a dedicated tool to track employee AI usage?

At small scale, vendor consoles plus a spreadsheet are fine. You need a dedicated layer once more than one team is spending, the tools multiply, finance wants one number, or you need spend attributed to teams, people, products, and PRs in one view. That is when stitching consoles by hand stops being worth it.