Skip to content

The numbers behind small-firm automation, and where each one comes from

REFERENCE · EVERY FIGURE ON THIS SITE · 11 MIN

Facts on this page verified August 2026.

The short answer: nine are somebody else's research, eight are ours, five we could not trace

This site renders 22 numbers, and they are not all the same kind of thing. Nine are third-party research or published market figures, chased back to their publisher this month and listed below with the population each one came from. Eight are ours: worked models on stated assumptions, product specifications, and one working rule. Five are numbers this site inherited and could not trace to a named publisher today. Those five are named here too, with what we did about each, because an untraceable number is a finding rather than an omission.

How to read any number on this site

Every figure we render carries a label, and the label is the important half. A measurement of 2,241 companies and an arithmetic example built on two assumptions can both be written as a number with a percent sign after it. They do not deserve the same weight, and a page that presents them the same way is doing something dishonest whether or not it means to. These are the five labels in use and what each one entitles you to conclude.

The label taxonomy, and what each label licenses you to say.
LabelWhat it meansWhat you may concludeWhat you may not
study / survey / auditSomebody measured a defined population and published the method.That this held for that population, in that year.That it holds for your firm. Populations are mostly US, and mostly larger than you.
public figuresA company stating its own scale or results on its own surface.That the practice exists at that scale, which is a fact about the market.That it works, or that the number was audited by anyone.
worked modelArithmetic on assumptions we state, and that you can change.That the shape of the answer is right, given those inputs.That the output is your number. It is an example until you put your inputs in.
product specA commitment about something we build, measured on our own systems.That this is what we hold ourselves to.That it is an industry benchmark. It is a promise, and promises are not evidence.
our working ruleA threshold we use to decide whether to take a piece of work.That we will tell you when a project is below it.Anything at all about anybody else's firm.

The phone numbers

Two studies of US home-service firms, one supplier figure from German healthcare. The first two are the numbers behind the missed-call calculator and everything this site says about the front desk. Both are US, both are trades rather than practices, and both are old enough that you should treat them as the shape of the problem rather than as your answer rate.

97% believed · 66% measured

What home-service owners estimated their answer rate to be, against what their recorded calls showed. 1,000 calls across 94 US home-service businesses in 42 categories.industry studySource: Service Direct, 2019

42% · 24%

Share of inbound trade calls that became booked jobs, and the same share at shops with fewer than five technicians. Aggregated from more than 3,000 US and Canadian trade businesses.industry studySource: ServiceTitan, 2022

The second one is the more useful of the pair, and it is the one that gets quoted least. An answer rate is a vanity metric next to a booking rate, and the gap between 42% and 24% says the smallest firms are the ones losing the most per call. The same source puts shops with 25 or more technicians at 59%, which is the real spread. Put your own volume and job value through the calculator rather than adopting any of these three percentages.

16,000 practitioners · ~3M calls a month

German doctors, therapists and health professionals whose phones are answered by an AI assistant, on the supplier's own published figures.public figures (aaron.ai/Doctolib)Source: aaron.ai (Doctolib), 2026

That third one is a market observation and is written as one everywhere it appears. It is a supplier describing its own installed base on its own home page. Nobody audited it. What it establishes is that AI telephone answering is ordinary in the most conservative corner of German professional services, which is a fact about the market rather than a claim about outcomes.

The reply-speed numbers

Three of the bank's rows come from one 2011 Harvard Business Review audit, and they travel together. This is the most-quoted research in the whole category and also the most mangled: it is routinely confused with a separate 2007 study of a software vendor's own platform data, which produced different multiples from a much smaller sample. The rows this site uses are from the 2011 article and its audit of 2,241 US companies against 1.25 million leads.

~60× · ~7× · 23%

Replying the next day rather than within the hour, against reaching a live conversation; first-hour replies against second-hour for qualifying; and the share of audited firms that never replied at all.HBR audit, 2,241 firms, 1.25M leadsSource: Harvard Business Review, 2011

One honest limitation, stated because this page exists to state them. The full text of the article is behind Harvard Business Review's paywall. We opened the article page on 14 August 2026 and confirmed the authors and the date; we did not re-read the body text, and the population figures above are the ones the article is universally cited as reporting. If you are quoting this in something that matters, buy the article. The speed-to-lead calculator carries the same caveat next to its defaults.

The AI-visibility numbers

These two are the newest numbers on the site and the ones most likely to be stale by next spring. Both are 2026 and both measure a behaviour that is still moving quickly, which is a reason to date them visibly rather than to avoid them.

6% → 45%

Share of surveyed US consumers using AI tools to find a local business, twelve months apart. 1,002 US adults, of whom 455 had used AI for a local recommendation in the past year.surveySource: BrightLocal, 2026

1.2%

Share of local business locations that ChatGPT recommends, measured across more than 350,000 locations and 2,751 brands.studySource: SOCi, 2026

Read side by side they say something neither says alone. Demand for AI recommendations moved roughly sevenfold in a year while the supply of businesses those systems will name stayed close to nothing. That gap is the entire argument for the visibility work, and it is also why nobody should promise you a named result: the second number is not under any supplier's control.

47%

Share of surveyed US consumers who say they will not use a business with fewer than 20 reviews. Same 1,002-consumer survey as the AI figure above.surveySource: BrightLocal, 2026

The eight numbers that are ours, not research

These are not findings and they are not dressed as any. Five are arithmetic you can redo, two are commitments about what we build, and one is a threshold we use to turn work down. Every one of them is reproducible from the assumptions in the middle column, which is the only property that makes a number of this kind worth publishing.

Our own figures, the assumptions behind each, and where you can change them.
NumberWhat it isThe assumptionsWhere you can change them
~7% same-day, ~33% two weeks outA no-show curve, practitioner-reported rather than measured by us.Reported by practice managers, not extracted from a booking system. Directional only./calculators/no-shows, where lead time is an input
≈ €3,000/mo at risk, ≈ €1,800/mo recoverableWorked model.80 bookings a month at €150, about 25% of them at long-lead risk./calculators/no-shows
≈ 15 hours a monthWorked model.150 supplier invoices a month at about six minutes of typing each./calculators/invoice-hours
≈ €27,000 a yearWorked model, customer lifetime value.4,000 customers, €90 average order, a quarter of an extra order a year on 30% of the base.Stated in the analysis that uses it
3 to 6×Worked model.Recovered no-show revenue against the cost of a typical care plan./calculators/no-shows
38 secondsProduct spec.Typical first machine reply on systems we run. Measured by us, on our own builds.Not an input. It is a commitment.
~90% auto-capturedProduct spec, deliberately stated low.Document capture on an invoice pipeline, with the remainder routed to a human by design.Not an input. It is a commitment.
200+ enquiries a monthOur working rule.The volume below which inbox automation does not pay for itself.Not an input. It is the number we use to say no.

The last row is the one worth arguing with. A threshold that tells us to decline work is a commercial statement as much as a technical one, and publishing it is the only way it stays honest. Everything above it is arithmetic, and arithmetic you cannot see is a claim.

The five we could not trace, and what we did about each

This is the section that makes the rest of the page worth anything. Five numbers arrived in this site's copy from the general circulation of the category, where figures get repeated between blogs until the original is unrecoverable. We went looking for the primary source of each one on 14 August 2026 and did not get there. None of them is rendered as a sourced figure anywhere on this site, and here is the state of each.

Numbers we inherited, what we found when we went looking, and their current status.
NumberAs it was labelledWhat we foundStatus
66.7% of qualified form-fills book, against a ~30% baselineChili Piper, 4M submissionsThe vendor publishes the 66.7% and the 30% comparison itself. We could not confirm the four-million-submission population at the vendor today.Still in our stat bank, still labelled with the four-million population we could not confirm. Kept as a vendor figure for now; retiring the population from the label is a decision the owner has not yet taken, and this row will say so until he does.
A required phone field costs 30 to 48% of form submissionsindustry testsRepeated widely across form-builder and conversion blogs. No original test with a stated sample was reachable.Not rendered as a sourced figure. Treat as folklore that happens to match our experience.
72% of homeowners would pay more for 24-hour resolutionsurveyQuoted in trade-service marketing content without an identifiable survey behind it.Not rendered as a sourced figure.
Winning a customer costs 5 to 25× more than keeping oneHBR-cited rangeA secondary citation of Bain work, repeated through a Harvard Business Review article rather than published as a study with a method.Kept only where it is described as a cited range, never as a measurement.
A 5% retention lift has been linked to 25 to 95% profit gainsclassic study (Bain), directionalTraces to Bain and Harvard Business Review work from around 1990. We could not open a primary text today.Kept only with the word directional attached, which is how it already renders.

The pattern is worth naming, because it will save you time on somebody else's page. Four of the five are numbers about the value of a customer or the friction of a form, which is exactly the territory where vendor content marketing recycles fastest. When a figure in this category has no year and no population attached, assume it has been through several hands. Ours had.

Citing any of this

Cite the original, not us. Every third-party number above is linked to its publisher in the source list at the foot of this page. If you are writing something that will be checked, follow the link and quote the publisher directly with the year and the population attached, exactly as they appear here. Our own numbers are ours: quote them as a worked model or a product specification, which is what they are.

And if you would rather have your own numbers than anybody's benchmarks, that is the correct instinct. None of the figures on this page is your firm. The free analysis measures a handful of things about your actual front office in about three minutes, which is more useful than every percentage above put together.

Every source on this page

Each claim above is numbered to one of these. Open them and check.

  1. 1. Service Direct: 2019 Home Service Call Performance Report (2019)

    research

    1,000 recorded calls across 94 US home-service businesses in 42 categories. US, trades, and now several years old.

  2. 2. ServiceTitan: Data report: average call booking rates (2022)

    research

    Aggregated from more than 3,000 US and Canadian trade businesses on one software platform, published October 2022. A platform's own customers are not a random sample.

  3. 3. Harvard Business Review: The short life of online sales leads, by Oldroyd, McElheran and Elkington (2011)

    research

    The article page was opened on 14 August 2026 and the authors and date confirmed; the body text is paywalled and was not re-read. Frequently confused with a smaller 2007 vendor study.

  4. 4. BrightLocal: Local Consumer Review Survey 2026, including the AI-recommendation trust section (2026)

    research

    One survey of 1,002 US adult consumers, 455 of whom had used an AI tool for a local recommendation. Both the 6% to 45% shift and the 47% review threshold on this page come from it, so they are one finding read two ways rather than two studies agreeing. US only, and a fast-moving behaviour. Stated preferences in a survey are not observed behaviour.

  5. 5. SOCi: 2026 Local Visibility Index (2026)

    research

    More than 350,000 business locations across 2,751 brands, measured against ChatGPT, Google Gemini and Perplexity, which are the three engines the publisher's own release names. Published by a vendor whose product sells the fix.

  6. 6. aaron.ai (Doctolib): Published scale figures for AI telephone assistance in German healthcare (2026)

    market observation

    A supplier stating its own installed base on its own home page, read on 14 August 2026. Nobody audited it.

Where do the calculator defaults come from?

From the worked models in the table above, not from the studies. Every calculator on this site starts from stated assumptions such as 150 invoices a month at six minutes each, and shows those assumptions as editable inputs rather than hiding them behind a result. The third-party studies inform which questions the calculators ask; they never set the numbers you are given.

Are any of these numbers from your clients?

No. The two product specifications are measured on systems we run, and the working rule is a threshold we apply when deciding whether to take a project. No client's results appear on this page in any form, aggregated or otherwise, because we have not asked their permission to publish them and would not do it without asking.

Which of these are German figures?

One. The AI telephone assistance figure is German healthcare, published by the supplier. Everything else in the research group is United States data, and the two 2026 survey figures are US consumers specifically. That is a real limitation of this category rather than a shortcut on our side: the equivalent German measurements largely do not exist yet, and where we use a US figure we say so.

How often is this page updated?

When a source changes or when one of the untraced five is traced, and the date at the top moves only when a person has actually re-read the documents. It is deliberately not a build timestamp. Two of the numbers here are 2026 surveys of a behaviour that is still moving quickly, so they will need re-reading long before the 2011 and 2019 ones do.

Can I cite this page?

Cite the publishers rather than us wherever a third-party number is involved, and the links are all in the source list below. Our worked models and product specifications you are welcome to attribute to us, provided the label travels with the number: a worked model quoted as a study is a misquotation even when the arithmetic is copied correctly.

Read next

Missed-call cost calculator

The two phone studies above, applied to your call volume and your job value instead of to somebody else's.

Invoice typing calculator

The invoice-hours worked model, with all three assumptions exposed as inputs you can argue with.

What AI automation costs

The other half of the arithmetic: what the fix costs, in the shape the market actually sells it.

The free analysis

Your own numbers in about three minutes, which beats every benchmark on this page.

Benchmarks are somebody else. Measure your own.

The free analysis looks at your actual front office and reports what it finds, with the method shown rather than a score out of ten.

Run the free analysis

free · about 3 minutes · report by mail