Time-Saved Maths for AI Email That Survives Scrutiny

Every AI email vendor has a calculator on their site. Enter your headcount, drag a slider, watch a large annual number appear. Then take that number to finance and watch it die. Finance is not being difficult — they can see the calculator asserts a per-user minute saving with no baseline, no measurement, and no discount for what actually happens to the recovered time. This post is the methodology that survives that scrutiny.

Time saved is a real thing. It can be measured, it can be defended, and it can be converted into money that a CFO will let you count. But only if you do three things vendor calculators skip: measure a baseline before rollout, measure the same tasks the same way after, and apply an absorption factor for whether the recovered minutes became output or just more inbox.

The short answer#

The calculation finance accepts has four inputs and one honest adjustment. Before rollout, you measure how long the target tasks take today — timed, not remembered. After rollout, you measure the same tasks the same way. You take the difference and multiply by frequency, so a two-minute-per-thread saving on 40 threads a day is not confused with a 4-hour weekly saving on one big task. Then you apply an absorption factor.

The absorption factor is the piece nobody wants to say out loud. If your operator saves 45 minutes of triage a day and spends those 45 minutes answering more email, the labour cost per week did not change — you shipped more customer contact, which may be valuable, but it is not cash back. If they spent it on billable work you would otherwise have hired for, the saving converts to money at their loaded cost. Most real rollouts sit between those two extremes, and the honest number is somewhere between zero and the raw minutes-times-rate.

Everything else in this guide is how to do those steps rigorously enough that finance signs the number rather than adjusting it downward on your behalf.

Criteria that actually matter#

A measurement plan is defensible or it is not. The dimensions below are the ones a skeptical reader — an analyst, a CFO, the board deck reviewer — will actually push on. Anything absent from your write-up is the thing they will ask about first.

• Baseline exists and was measured. "Our team spends about two hours a day on email" from a survey is not a baseline; it is a self-report of a feeling. A baseline is timed observation, telemetry, or a diary study run before the tool went in.
• The unit of measurement is a task, not a person. Minutes per triaged thread, minutes per drafted reply, minutes from arrival to first response. "Hours saved per user per week" is derived from those, not assumed.
• Self-report is discounted or triangulated. Users overestimate time saved on tools they like and underestimate on tools they dislike — the direction depends on how the survey is framed. Never publish a self-reported number as the number.
• Attribution is isolated from confounders. If you rolled out the AI the same month volume dropped 12%, you did not save time; volume did. Hold volume, staffing, and workflow constant, or model them out.
• Absorption is stated explicitly. What actually happened to the recovered minutes? Did headcount drop? Did billable output rise? Did the team simply handle more email? Each answer converts to a different dollar figure.
• The horizon covers the full workflow. A tool that saves two minutes on the draft but adds one minute of approval-review saves one minute, not two. Measure the loop end to end.

One more criterion is easy to forget until an auditor asks: the measurement must be reproducible by someone who was not in the room. Write down the sample, the definition of the task, the timer discipline, the exclusion rules. If the reader cannot rerun your study on a different team and get a comparable number, you have an anecdote, not a measurement.

The measurement methods, ranked by how much finance trusts them#

There are essentially four ways to measure minutes saved. They are not equivalent, and a mix usually beats any single one. The table below is the ranking finance actually applies, whether they say so or not.

In practice, the plan that survives review pairs timed observation for the baseline with telemetry for the after-state, uses a control group where team size allows, and treats the self-report survey as the sanity check rather than the answer. When the four methods disagree, the disagreement itself is the finding — it means the effect is smaller than any single method suggested, and the honest write-up says so.

A worked example — one team, one month#

A sales team of ten inbox operators handles a mix of inbound leads and account follow-ups. Before rollout, a two-week diary study finds each operator spends a median of 92 minutes a day on email triage and initial reply drafting, across roughly 60 threads. That is the baseline: 92 minutes per operator per day, sample of ten, two weeks.

The team rolls out an AI email tool. Adoption stabilises by week three. In weeks four and five, telemetry shows the median time from thread-open to send has dropped by 40%. A parallel diary study on a random subset confirms the direction and gives a slightly smaller figure — 34%. The team publishes the smaller number, because that is what the honest overlap of the two methods supports.

Applied to the baseline, a 34% reduction on 92 minutes is 31 minutes per operator per day, or roughly 2.6 hours per operator per week. Across ten operators, that is 26 hours per week of recovered time. So far, so good.

Now the absorption question. In interviews, seven of the ten operators say they used the recovered time to answer more email — the queue simply refills. Two say they used it on outbound calls that were previously deferred. One says it disappeared into meetings that expanded to fill it. The team's manager confirms outbound call volume rose modestly in the study window.

The honest write-up looks like this: 26 recovered hours per week, of which roughly 20% converted to a new activity finance can price (the additional outbound calls at loaded cost), and 80% was absorbed by increased inbound handling capacity. The capacity increase is real and worth naming — it means the team can absorb 30% more inbound volume without hiring — but it is not a dollar saving to book against headcount.

The number that goes in the deck is the small one: the priced portion, plus a clearly labelled capacity figure for the rest. It will not blow anyone's mind. It will also not be adjusted downward by finance, which is the entire point.

Red flags in vendor time-saved calculators#

The calculator on a vendor's marketing page is a lead-generation tool, not a measurement tool. That does not automatically make its output wrong, but it does mean the incentives are one-directional. The patterns below are the ones that most reliably inflate the answer.

• A per-user minute saving is asserted, not measured. The number "saves 90 minutes per user per day" appears with no citation to a study, no sample size, no methodology. It is a marketing assumption dressed as an input.
• The saving is quoted per day and then multiplied by 250 working days. This is arithmetically fine and rhetorically dishonest — it turns a two-minute per-thread saving into an eye-watering annual number that no team member has ever felt.
• The absorption factor is zero. The calculator assumes every recovered minute becomes a billable minute at the user's fully loaded cost, including benefits and overhead. This is almost never true.
• Confidence percentages appear with no source. "85% confidence", "90% accuracy" — numbers precise enough to look measured, vague enough to be unfalsifiable. If the vendor cannot cite the study, the number is theatre.
• No baseline is required to run the calculator. If you can generate an ROI figure without ever telling the tool what your team spends on email today, the figure was decided before you arrived.
• The unit switches mid-argument. Minutes per email in the input, hours per week in the middle, dollars per year in the output — with a rate assumption buried in the fine print. Each conversion is a chance to inflate.

None of this means the tools do not save time. Many of them do, and the honest numbers are still large enough to justify the purchase. The problem is that the theatrical numbers are so large they make the honest ones sound disappointing by comparison, and a buyer who quotes the theatrical figure to finance loses credibility they then cannot rebuild with the real one.

What we'd pick and why (honest)#

This is a methodology post, not a product roundup — the reader came here to build a defensible number, not to buy anything. But if you are building that number for an AI email rollout and the shortlist includes us, here is how we would rank ourselves on this specific question, honestly.

For observational time-tracking of desktop-wide activity — actual keystrokes, app focus minutes, meeting-versus-email splits — a dedicated workforce analytics tool like RescueTime or a category equivalent will give you cleaner raw data than any email client can, ours included. That is a real category and we do not compete in it. If your measurement plan leans heavily on desktop telemetry, use that category for it.

Where an AI email tool actively contributes to the measurement — not just to the outcome — is the audit log for agent actions. We build AI Emaily as an AI-native email client with a full audit trail on every drafted reply, every send, every Autopilot action, and every undo. That log is the cleanest possible after-state telemetry for the specific workflow the tool changed, because it is generated by the tool itself and cannot be reconstructed by an external stopwatch. Pair it with a two-week diary baseline and you have both halves of the calculation from primary sources.

The honest scoping: we are the right measurement partner when the workflow you are measuring is email triage, drafting, and reply — the work that runs through the inbox. We are not the right measurement partner when the question is what happens to the recovered time across a workday, because that leaves the inbox and enters territory a mail client legitimately cannot see. Use both.

On the commercial side: we do not have a free plan. There is a seven-day free trial on Pro or Autopilot with the card taken and no charge if you cancel before day seven, and the current numbers live on the AI Emaily pricing page rather than in this post, because pricing changes and posts do not. If you want to run the measurement plan above against our audit log during that trial window, that is exactly the shape of use it was designed for. We build AI Emaily and we are naming ourselves in this section because the post's method leans on the audit-log capability we ship.