We read the public material of 31 vendors and publishers around AI website agents, in 6 families, as of 10 October 2026. The set holds:
No vendor is named and no average is computed. The study makes no claim about Sophia's results. HumanOps has no measured results of its own yet.
Study version 2026-10-11.1. Dataset date 10 October 2026.
Printed from humanops.pro/kpi-study/, study version 2026-10-11.1, dataset date 10 October 2026. The web page is the source; this copy is derived from it.
Both paths start at the same definitions and end at the questions to ask a vendor.
The study reads public material in 6 families of tools. Vendors appear only as families, with counts.
| Family | Units | Cases | Rows |
|---|---|---|---|
| AI website agents and AI outreach agents | 11 | 64 | 378 |
| Interactive product demo platforms | 2 | 4 | 12 |
| Attribution platforms and CRM influence reports | 4 | 4 | 14 |
| Meeting booking tools | 2 | 3 | 12 |
| Industrial tools: quoting, product configuration, one implementer, one services firm | 4 | 11 | 19 |
| Executive and academic measurement research, plus one secondary vendor source | 8 | 0 | 19 |
| All families | 31 | 86 | 454 |
Each finding gives its count and its base. A market label stands in quotation marks. Where the research grouped differently named metrics under one label, the finding says so.
41 rows sit at the meeting booked stage and 3 at meeting held. None of those 3 rows states a formula or a count of held meetings.
Recounted from the evidenceNone of 378 rows from AI agent vendors is coded at meeting held. The 3 rows at meeting held all come from meeting booking tools. 1 agent vendor row reports a drop in no-shows, with no formula, coded at another stage.
Recounted from the evidenceNone of 454 rows reports how often the sales team accepted buyers the agent qualified. 27 rows report qualification, and each gives only the agent's own verdict.
Partly a research readingOf 6 vendors with qualification rows, 2 publish the criteria their agent uses, and 1 leaves them to each customer. The research found no published criteria for the rest.
Partly a research readingThe research groups metrics of 6 vendors as “engagement rate”. They name 5 different bases, and 1 vendor states no base.
Partly a research readingThe research groups 7 rows from 5 vendors as “meeting booked rate”. They name 4 different bases, and only 1 of those is defined. 3 rows state no base.
Partly a research readingIn 174 of 271 graded rows, the source does not state what the figure is divided by. 54 rows give a total or another figure that needs no base, and 43 name a base.
Recounted from the evidenceIn 151 of 271 graded rows, the source states no period for the figure. 23 more rows give only part of a period: a length without dates, a start without an end, an undefined window or a publication date.
Recounted from the evidenceThe research codes 20 rows as “influenced pipeline” claims or definitions, from 9 vendors. 6 of those rows are definitions.
Recounted from the evidence11 of the 20 rows state no rule at all. The other rows range from a full published rule to a model named without its settings.
Recounted from the evidenceOf those 9 vendors, 2 publish a full rule for “influenced pipeline”, 4 state part of a rule and 3 state none.
Partly a research reading11 of 454 rows state an “ROI” level. None shows the cost behind it in the source the research read.
Partly a research readingNone of 378 AI agent vendor rows reports a quote or RFQ outcome. Industrial tools do, in 6 of their 19 rows. No row from the other families mentions quotes.
Partly a research reading19 of 19 rows from the 4 industrial vendors sit at grade E, the weakest rung of the ladder.
Recounted from the evidenceThe research grades rows on its own ladder, from A, the strongest, to E, the weakest. The letters match How we count, but the research level A is looser.
| Level | Meaning | Rows |
|---|---|---|
| A | Controlled experiment | 11 |
| B | CRM-linked sourced outcome with an explicit attribution rule | 0 |
| C | Before and after with a defined base | 2 |
| D | Influenced or multi-touch credit | 15 |
| E | Vendor calculation, survey or anecdote | 243 |
| Definition | A product definition or documentation, with no outcome to grade | 183 |
271 of 454 rows carry a grade, and 183 are definitions.
In the research, level A means any controlled experiment. How we count asks more of HumanOps: a control group at visitor level.
The study computes no average across vendors, because it would mix different quantities and print a typical value nobody measured.
The research groups differently named metrics under one market label, and their bases differ. “Engagement rate” has 5 named bases plus 1 unstated, and “meeting booked rate” has 4 named plus 3 unstated.
In 174 of 271 graded rows, the source does not state the base. Where cases give a period for booked meetings, it runs from weeks to a year. The cases are stories the vendors chose to publish, not a sample.
HumanOps customers are not a sample of the market either, so HumanOps never shows their aggregate as a market figure.
These practices from the material are worth copying. The page names no vendor.
One AI agent vendor published its attribution rules as query text in its developer docs, with a fixed window and a contact condition.
The query is a sketch and does not run as printed. The docs carry no date, and no case with numbers under the rule was found.
One vendor versions its metric definitions, keeps the old ones beside the new ones, and states what each change does.
The wording was read through a summarising tool and waits for a re-check.
One customer test split site visitors in half, with and without the agent, over a dated window.
The design is strong, but the source gives no counts for either half and no significance test.
One attribution platform reports an influenced opportunity at its stage value. It keeps that apart from model credit, which shifts when the model changes.
Some vendors publish what stays out of their counts: auto-replies, cookie-only identities, test runs and shared company domains.
A few vendors define the base itself, not only the rate. Examples: inbound booking requests, eligible calls, contacts the agent messaged.
One observational study states its own selection bias: buyers who used the demo may already have had higher intent.
A vendor produced it, and the study lists it as an aggregate, not a case.
Each question follows from a finding on this page.
HumanOps has no measured results of its own yet. A public HumanOps figure needs level B at least: counts under a published rule, never a multiple.
The same questions, turned into definitions for Sophia: one definition, one source and one base for every number.
Read How we count On example dataEvery tile carries its data class and its definition. The numbers are seeded and illustrative, not a customer result.
Open the example cabinet ModelledA scenario model from your own inputs, kept apart from observed data.
Open the business-case calculatorBook a call about the study and how we count, or read the definitions first.