Argue with the method, not the number.
Any tool can produce a large number for an AI initiative. The questions a CFO actually asks are where it came from, who agreed to it, and what happens to it when the assumptions change. Roiva publishes the answers, so the argument happens over the assumptions instead of over the output.
How a number is made
The chain behind every dollar
Every part of it is a record you can open, not a step in a black box.
A formula, not a spreadsheet
Calculated value is produced by a named expression over metric inputs. Formulas are versioned: a change to the arithmetic writes a new version and keeps the earlier ones, and each entry records the version, expression and inputs it was calculated with, so a number approved last quarter still reproduces today. Value that is not calculated says so: a synced figure keeps the source records behind it, and one a person entered names them and the documents they cite, or shows that it cites none.
A baseline, declared as what it is
A baseline is the figure before the AI that later readings are compared against. Every initiative records one and how it was measured — from history, by a time study, by a sample timed without the AI after go-live, from a benchmark, or as an estimate — with the window and sample size where they are known. An estimate stays labeled an estimate, and so does a convention Roiva lets stand in until a figure is measured — the minutes of manual work an automated run replaces, the minutes a resolved issue saves, ten minutes for a deflected ticket. The Board Brief, the Value Evidence Register and the ROI pages show how much value rests on a reconstructed or estimated baseline.
Rules agreed, best before results
An initiative's measurement charter states, in words, what the AI gets credit for and what it doesn't, with the formulas, shares and baselines that count it. Someone other than its drafter signs it, best before results arrive. A figure calculated outside it is marked so and never approves itself, and a revision made after value was approved needs a reason and is printed on the Board Brief.
Inputs, frozen at calculation
Each entry stores a snapshot of the inputs it was calculated from and how they arrived. Change a metric next month and last month's number still explains itself.
Sign-off before it counts
A person's entry is submitted for review and then approved; a figure a formula or sync calculates goes straight to an approver. Either can be disputed and resubmitted. Only approved entries reach a report. An account can let entries from Roiva's library formulas and AI-built ones approve themselves, but never one that reads an attribution share, rests on a Roiva default or an estimated baseline, falls outside the charter, reads a time study past its re-study date, comes from a formula that doesn't adjust a count for its volume, or comes from a formula no person has approved before.
An append-only record
Nothing in Roiva edits a row of the audit log — which is what makes the chain above checkable after the fact rather than merely stated. Rows are kept for two years. Deleting a person blanks the personal details on their rows; deleting an account removes its log and leaves a record that it was deleted.
See the chain on the reports themselves: a Board Brief and Value Evidence Register made by Roiva's own report code, for a fictional company. Read the sample reports →
The arithmetic
What counts, and what does not
Six rules decide whether a figure reaches a report, in which column, and in which month. They are deliberately asymmetric, because the failure mode worth preventing is an overstated return.
- Value counts only once approved. A calculated but unreviewed entry is visible on the initiative and on ROI › Provisional, labeled as waiting, and is in no reported total, ROI, report, export or API.
- Cost counts in full, approved or not. Implementation and run costs are netted off without needing anyone's sign-off.
- Net, not gross. Reported ROI is cash value minus cost, divided by cost. Every ROI Roiva reports is net of cost; none is computed on value alone.
- Cash and capacity, never added. Freed capacity is reported beside cash, in its own column, and is in no ROI. The next section says why.
- In dollars, at the ECB's rate. An entry keeps the currency its source booked it in, and totals add its US dollars at the European Central Bank's average reference rate over the entry's period, stored beside it with the rate. Value in another currency waits for its period's rates and is in no total until it has them. A cost doesn't wait: it counts at the latest rate until its period's are published, so a known cost is never left out, and an undated one takes the rate for the day it was booked. A currency the ECB doesn't publish isn't converted, and a total that leaves one out says so.
- Booked as earned. Calculated value is booked a month at a time, and a recurring deal synced from a CRM is booked over the 12 months from its close date, never a year up front: its annual value ÷ 12, or its monthly value, for each full month that has ended. A deal that closes mid-month books the close month by its days, and the rest of that month in a thirteenth. A one-time deal books on its close date.
Two kinds of value
Cash and capacity
Most of what AI does at first is give people time back. That time is worth something, but it is not money the business stopped spending: payroll is the same afterwards. So Roiva reports two figures, side by side, and never adds them.
Cash value
Money that shows in the books: the margin on revenue gained, costs that fell, losses and spend that did not happen, and freed capacity a person showed became cash. Every ROI Roiva reports, planned or actual, is a cash ROI.
Freed capacity
Hours and output freed up, priced at what they cost: a labor rate, or a cost per ticket or per item that is mostly people's time. It is reported beside cash on the ROI pages and in every report, and is in no ROI.
How a formula is classed
By what it prices, not by what anyone would like it to be. A formula that values people's time — a labor rate, or a unit cost that is mostly staff time or lost output — books freed capacity, and Roiva will not save it as cash. Every formula in the library declares what it counts as, and the formulas reference shows it.
When capacity becomes cash
Only when a person shows it did, one month at a time. They record the amount as capacity realized as cash, say what changed — a hire not made, overtime or contractor spend that fell, growth absorbed without adding staff — and cite the document that shows it, such as a canceled requisition or a contract not renewed. An approver signs it off, and it can be no more than the freed capacity approved for that month. The amount then counts in cash and comes out of capacity, so a dollar is counted once.
Roiva never decides on its own that freed hours were redeployed or saved. That is a business decision, and the record of it is the document someone cites. We have written about why an avoided hire is capacity unless there is a dated decision and how to tell the two apart in your first six weeks.
Measuring what no system records
Time studies
Most AI value is minutes: per ticket, per invoice, per claim. No system records how long the work took before the AI, so the baseline is the weakest link in most business cases. A time study measures it with the people who do the work.
- Planned before it starts. The owner states the precision wanted and how much the task varies, and the study says how many timings each arm needs, at least 30, and asks for at least five people in each; an arm timed on fewer than three is not published.
- Days drawn, not chosen. When a study times the work both with the AI and without it, Roiva draws each person's days without it at random, so nobody saves the easy tasks for one side. Timings imported from another tool weren't drawn, and the result says how many there were.
- An interval by person. Twenty timings from one fast person are one person's pace, so the interval is worked out by person and narrows only as more people are timed. The published figure is the mean, because only the mean times the volume is the total.
- Named for what it is. Timed before go-live on the work without the AI alone, with every timing recorded in Roiva, the result is a measured baseline. A study that also times the work with the AI, runs past go-live, or imports timings gives a reconstructed one, and reports count it apart.
- Due again. A result stands until its re-study date. After it, value calculated from it is marked and waits for a person to approve it.
A time study shows how long the timed tasks took the people timed, in that window, with an interval for how much it could differ had other people been timed. It does not show that the tasks stand for every task, that people work at the same pace when they are not being timed, that typed minutes are exact, what freed minutes were used for, or that the AI caused a change in volume or outcomes. It is a measured estimate, not an experiment.
The limits
What this method does not prove
Stated here rather than discovered later. A methodology that only makes claims is a brochure.
Where judgment enters
Four places. Roiva makes each one explicit, versioned and re-runnable — it does not remove them.
- The baseline. Someone decides what "before" was. Roiva records who, when, over what window and by which method, and a time study adds an interval — but an estimated baseline is still an estimate, and a reconstructed one is a sample taken after go-live.
- The attribution rule. Someone decides how much of an outcome the AI gets credit for. Roiva makes that rule explicit and re-runnable, and the measurement charter is where it is written down and signed, best before results. Where a formula values a business-wide outcome — a loss ratio, a churn rate, a win rate — the rule is a share of the change someone sets on the initiative, with no default, whoever wrote the formula: Roiva won't save one that values the whole change. It records the figure, who set it and when, and every entry that reads it waits for a person's approval. A formula an account saved before v4 that values the whole change keeps calculating until its arithmetic is next changed. It is not a causal claim, and Roiva does not run experiments.
- Whether capacity became cash. Someone decides that freed time turned into money not spent. Roiva requires the document that shows it and caps it at the capacity approved for the month — but the decision, and its evidence, are yours.
- The confidence rating. Confidence is a judgment your team records on an initiative and on each value entry, and uses to triage review. Roiva does not compute a confidence of its own: the Board Brief and Initiative Review show an initiative's as the average of the scores on its approved entries, weighted by their dollars. No value, cost or ROI is weighted by it.
What Roiva will not tell you
- Whether the outcome would have happened anyway. A rising number during a deployment is a correlation; the attribution rule is how you choose to treat it, stated openly rather than buried in a total.
- Why a rate moved. A count that rises and falls with volume — calls, scrap cost, fleet miles, a cloud bill — is compared per unit of that volume and multiplied by the month's volume, so a busy month is not booked as a loss nor a quiet one as a saving. A formula that compares the raw count still saves, but every entry it calculates is marked as not adjusted for volume and waits for a person's approval, until the formula adjusts or someone other than its author records why it shouldn't. Adjusting takes out the volume the formula reads and nothing else: weather, mix and prices still move the rate.
- What freed time was spent on. Freed capacity says how much time was made available, not that it went to something worth having. Roiva counts it as cash only when someone shows what changed.
- That an estimated input is a measured one. The method is carried on the entry, so a number resting on estimates can be identified as such by anyone reading it.
- That your AI spend is complete. Roiva sees the providers, cards and books you connect. AI paid for anywhere you haven't connected is invisible — which is itself a finding worth surfacing.
Published
The methodologies we publish
Each is versioned and dated, and every report PDF Roiva generates carries the version that produced it.
Measurement methodology
How value is calculated, approved and audited once an initiative is live in Roiva — the chain described on this page.
v6 · October 2026 · this page
AI ROI Exposure Assessment
How the free assessment estimates your annual AI spend, how much of it is provably measured today, and which moves close the gap. Full formulas, bands and worked examples.
v1 · May 2026
Versioning
How we update this
The version above is printed on the footer of every PDF Roiva produces from your account, so a report forwarded a year from now can be matched back to the method that produced it. When the method changes we bump the version and date rather than retuning silently.
v6 — October 2026
- Value and cost booked in another currency count in dollars at the European Central Bank's average rate over the entry's period. The entry keeps its own amount and currency, and the rate is stored beside it. Converting doesn't send approved value back for review.
- Value in another currency counts once its period's rates are published. Cost counts straight away at the latest rate, marked provisional, and moves to the period's average when it can; an undated cost takes the rate for the day it was booked. What has no rate is in no total, and the report says what was left out.
v5 — October 2026
- The minutes an automated run or a resolved issue saves are measured figures — how long the task took a person, by a time study or your own timing — not a guessed number of minutes saved. Until they are measured, Roiva's convention stands in, and the value it gives counts as estimated.
- A deflected ticket or conversation priced at Roiva's ten-minute convention counts as estimated too; one priced at your own figure or from your time study does not.
v4 — October 2026
- The attribution share holds for every formula, however it was written: one that values the whole change in a business-wide outcome is refused on save.
- A formula that compares a count that rises and falls with volume without adjusting for it is flagged: its entries say so wherever they appear and wait for a person. Clearing the flag takes a recorded reason from someone other than the formula's author.
v3 — October 2026
- Every baseline is declared by how it was measured: from history, a time study before go-live, a sample timed without the AI after go-live, a benchmark, or an estimate. Reports split approved value by whether it rests on a measured, reconstructed or estimated baseline.
- An initiative's measurement charter states its attribution rule, formulas, shares and baselines, and is signed by someone other than its drafter. A figure calculated outside it never approves itself, and a revision after value was approved needs a reason and is printed on the Board Brief.
- A time study publishes the mean of timings with an interval worked out by person, from days drawn at random where it times both with and without the AI. Its result stands until a re-study date, after which value calculated from it waits for a person.
- A help desk's deflected tickets can be priced from the initiative's own time study of agent minutes, in place of Roiva's ten-minute convention.
v2 — October 2026
- Value is reported as two figures: cash, and freed capacity beside it. Reported ROI is cash ROI; freed capacity is in no ROI and is never added to cash.
- A formula that prices people's time books freed capacity. Each formula in the library declares what it counts as.
- Freed capacity becomes cash only when a person records it for a month, citing a document, within that month's approved capacity.
v1 — published September 2026
- Initial published measurement methodology: versioned formulas, declared baselines, frozen inputs, approval before value counts, append-only audit log.
- Reported ROI is net of cost; value requires approval, cost does not.
- Confidence is recorded by your team for triage, and not used to weight any reported figure.
See it against your own numbers
The free AI ROI Exposure Assessment runs the published math on your inputs — five minutes, no account, no card; your email unlocks the report.