The AI services market is full of bold percentages with no definitions: no baseline, no period, no one who checked. We take a different route, and it applies both to our own results and to the market research we cite. Next to every number you should be able to see where it comes from, what exactly it measures and how far to trust it.
This page sets out the rules we use to calculate and publish numbers: on this site, in proposals and in client reports.
Not every number is equally reliable. One was checked by an independent party, another comes from the client, a third is only a calculation. So we sort numbers into five levels, from the most reliable, V1, to V5, which we never publish. When we quote a number of our own, we state its level.
| Level | What stands behind it | Example wording | How we present it |
|---|---|---|---|
| V1 · independently verified | a client result checked by a third party; we know the before, the after, the period and how it was measured | “First response time fell from 18 to 4 minutes over three months” + a link to the check | as a fact |
| V2 · confirmed by the client | the client approved the figure or quote in writing, but nobody checked it independently | “According to the client, managers spend half as much time on requests” | marked “according to the client” |
| V3 · our own measurement | we measured it ourselves in tests or from system logs | “On 200 test questions the agent answered 180 correctly” | together with how it was measured |
| V4 · calculation | an estimate from a formula or a forecast, not a measured result | “Payback of about 4 months if the agent takes over 80 hours of work a month” | always as an “estimate”, with assumptions |
| V5 · unverified | a draft, or a number without a source or method | — | never published |
The examples show the wording only; they are not results of our clients.
Every metric has a numerator, a denominator and a period. These are the metrics we use for agents and automations.
| Metric | Formula | Watch out for |
|---|---|---|
| Resolved without a person | conversations completed by the agent with no hand-off ÷ all conversations the agent took part in | we split “confirmed” (the customer said it was solved) from “assumed” (no request for a person and no repeat contact on the same issue within 24 hours) |
| Hand-off rate | conversations handed to a person ÷ all conversations with the agent | a high rate is not always bad; sometimes that is the design |
| Corrected answer rate | answers a person corrected or marked wrong ÷ all reviewed answers | measured on a sample reviewed by hand |
| Time to first response | median time from the customer’s first message to the first reply | median, not mean, so a few long cases do not distort it |
| Manual hours per month | number of process runs × average time per run | measured the same way before and after launch |
| Flow failure rate | failed runs ÷ all runs in the period | failures that lost data are counted separately |
Why the base matters. 88% of McKinsey respondents report that their organisations regularly use AI in at least one business function (The state of AI 2025, 1,993 respondents; the 2026 edition says “nearly nine in ten”). Among EU enterprises with 10 or more employees, 20.0% used AI in 2025 (Eurostat). The two do not contradict each other: the first is a survey weighted towards large companies, the second is official statistics across all enterprises.
Why the version matters. The 2023 working paper on AI assistants in customer support reported +14% overall and +34% for novices across 5,179 agents. The peer-reviewed version (Quarterly Journal of Economics, 2025) reports +15% on average and about +30% for less-experienced workers, across 5,172 agents. We cite the latter.
Monthly saving = manual hours per month × hourly cost − support − model usage − platform. Payback period = one-off cost ÷ monthly saving.
Before launch this is a level V4 estimate with assumptions you can see and change. After launch we replace it with measured data, using the formulas above. A worked example is in our AI agent cost guide.
Every page with numbers shows the date of its last review. We re-check market data when a new edition of the source comes out.
Spotted a wrong number or a broken source? Write to boost.info.ai@gmail.com and we reply within five business days. Corrected figures carry the date of the change.
We publish a client result only with written permission, a baseline and a period (level V1 or V2). Case studies will appear here once they meet that rule.
No. We give an estimate with open assumptions and measure the actual effect after launch with the same formulas.
We report confirmed and assumed resolution separately, show absolute numbers and give you access to the data behind the calculation.
Because they measure different bases: some survey respondents, mostly from large companies; others count every enterprise in a country. We always state the base next to the percentage.
Start with a short brief — we reply with an estimate.