AI Ledger
Is your AI spend paying off, and how should each team's budget work? Powered by NVIDIA Nemotron on Nebius Token Factory.
Data and assumptions (sliders change every number) custom value definition custom rules
What to do first
Plain-English summary
Where the waste is
Unused = a seat with under 3 sessions a month ⓘ. An unused seat is usually a fit or awareness problem, not a people problem: ask the team before reassigning. Savings count only the share of a seat you can actually cancel (contract minimums and notice periods)ⓘ. High-spend = token spend above the 75th percentile of ROI-positive peers in the same department and levelⓘ, on seats whose ROI is negative. Start with training and workflow fit; this is about how tools are set up, not about anyone's effort.
Proposed actions: approve what you want
Watch it over time
Monthly AI spend and drift alerts
Policies, simulated on this data
Assumptions behind these figures: pay-per-use list priceⓘ, rationingⓘ, elasticityⓘ, seat savingsⓘ, fixed capⓘ, learning floorⓘ.
What each policy assumes
Recommended budget per department
"Expected under plan" is the average cost per head after reassigning unused seats. "Ceiling" is the most one active person may use, so it is higher than an average on purpose. Caps are peer norms: what colleagues of the same role and level who get a positive return spend today. They describe use, not need. Interns and juniors are never set below the department cap, because they iterate more while learning. Treat every cap as a starting point to adjust with the team.
Cost by model: route routine work to a cheaper one
Prices and shares: frontier priceⓘ, cheaper modelⓘ, quality penaltyⓘ, routine shareⓘ.
Define what AI output is worth to you
AI value is: extra output (measured from your data) x what one unit of output is worth x the share you count as real (the realization sliderⓘ). The tool measures the first part. You decide the second, per department. By default a unit is worth the hours it takes times an hourly rateⓘ.
Ways to value a unit: time saved at a loaded hourly rate (the default); the handling cost avoided per ticket; the expected margin of a qualified sales opportunity; what a contract review costs externally. Use your own figures: the tool does not know them.
Blank fields use the data's own figures (shown in gray). "All departments" fills in for any department you leave alone. An extra benefit per user is for value that is not extra output, such as retention or quality; it is counted in full and is not reduced by the realization share.
What counts as one unit of output?
Name the unit your team really produces, and if you track several things, say how much each counts. For example, one pull request = 1, one code review = 0.3, one commit = 0.1. The tool then measures the AI effect on that mix.
Your CSV supplies the metrics: the required outcome_units column, plus any extra outcome_<name> columns. A weighted mix is a new unit, so also set what one is worth in the table above (dollars per unit, or hours per unit). Blank weights mean 1 for the main metric and 0 for the others.
Speed. If you track the time to finish a task (lower is better), choose it under "Measure the effect by". The AI effect is then the fall in time per task, applied to each person's baseline task volume and priced as freed-up hours at their hourly rate. Do not weight a time column as output.
What your definition changed
Advanced: the definition as JSON
Keys are department names or "*". Fields: mode ("time" or "unit"), value_per_unit, hours_per_unit, hourly_rate, benefit_per_user.
The tool's judgement calls, as settings
These are rules of thumb built into the tool: when a seat counts as unused, what counts as high spend, how many months to measure, what a fixed budget caps. Change them to match how your organization works. Each shows its default and where it comes from (ⓘ). Risky settings are allowed but come with a warning that says what the stress tests found.
The sliders in "Data and assumptions" cover the value share, elasticity, the cheaper model's price and the pay-per-use price. Value and output are on the Your value tab.
Share of tokens that are routine work, by department
Used by the model-routing estimate: how much of each department's use is summaries, drafts and lookups that a smaller model could do. Measure it on your own traffic before relying on it.
Why we use a control group
Baseline and measured monthsⓘ; pre-rollout checkⓘ.
Output rose for people who never used AI too (busy season, new product). A plain before/after would credit all of it to AI. Each person is compared with their own first two months, as a ratio, so an intern and a senior are comparable. Limit: people choose whether to adopt. We test whether adopters were already pulling ahead before rollout and remove that gap, but differences that only start at rollout cannot be seen.
ROI by department, with 95% range
"Uncertain" means the range includes zero: the data cannot tell this department from break-even. "Break-even credit" is the share of the estimated gain that must be real for the department to pay for itself.
ROI by level
Value is time saved times hourly rate, so lower-paid levels show lower dollar ROI on the same seat. This is not a verdict on their work, and it ignores learning and ramp-up. Peer comparisons for high spend are made within level, never across.
How much of this is opinion?
Computing 180 assumption scenarios…
Does the analysis find what is really there? (synthetic data only)
This company is synthetic, so we know the truth. Hidden labels (usage patterns built into the generator) are compared with what the analysis flagged. Messy cases are included: part-timers, people on leave, new hires with no baseline.
Show the math
Budget policies
Does the published research support these numbers?
Published range usedⓘ; perception gapⓘ; calibration of the demo companyⓘⓘ.
What effect does this spend need to pay off?
By department
Break-even effect = total monthly cost / dollars of value per 100% effect, if every user gained the same share of their baseline output (at your current realization). Published effects are for people who used AI, so each scenario applies one to the people counted as users. What-ifs, not forecasts.
Do less experienced people gain more?
Published studies say they do. If your data agrees, that is a reason not to cap juniors below the department, and the tool already floors junior and intern caps at the department cap.
Do employees' own estimates match the measurement?
The studies, with what was checked
Left out on purpose
Where every value comes from
Every number the tool uses is listed here with its origin. Most of the dollar figures are assumptions, and the page says so: a few are anchored to public price lists or published studies, some are our rules of thumb, and the demo company's figures are invented. Where you see ⓘ elsewhere, it opens the matching entry.
The values are read from the running code, so this page cannot disagree with the tool. Price lists and studies were read on the date shown; a source marked "not confirmed" is a third-party listing.
Plain definitions with a worked example. Where an example says "in this data", it uses this company's live numbers and changes with the sliders.