Key takeaways
- Measure a baseline first: calls received, answered and missed by hour, bookings, conversion and average job value, for at least four normal weeks.
- Compare like with like. Heating demand is seasonal, so compare with the same weeks last year or run the test across matched periods.
- Return = extra booked jobs multiplied by gross profit per job, plus time saved, minus the full cost of the tool, divided by that cost.
- Construction firms are cautious adopters: 81% used no AI in autumn 2026[1], so build the case on your own numbers rather than vendor claims.
- Check quality as well as volume; most businesses using AI keep some human checking of its outputs[2].
What you are really buying
Whether it is a person at an answering service, a virtual receptionist, an AI phone assistant or a part-time office hire, a call-handling tool is bought for one reason: to turn more of the calls you already receive into booked, profitable work, and to give engineers and owners back time they currently spend on the phone. Everything in the measurement should tie back to those two outcomes.
The sector is still early in adopting AI tools. ONS survey data show the share of construction businesses using no AI fell from 91.9% in September 2023 to 81% in autumn 2026[3]. Of construction businesses using AI, 65.1% did so to improve business operations[4], and across all firms the most common reason for adopting AI was to increase efficiency or productivity[5]. That caution is healthy. A clear before-and-after test protects you from paying for a tool that does not move your numbers.
Source: Office for National Statistics, 2023[3]
Show the data
| Item | Value |
|---|---|
| September 2023 | 91.9% |
| Autumn 2026 | 81% |
Questions to settle before the trial
Agree the rules before the test starts, so the result is not argued about afterwards.
- Which calls go to the tool: all calls, only overflow when nobody answers in a set number of rings, or only out of hours?
- What it is allowed to do: take details, book into set slots, give your standard prices, or only pass messages on.
- How safety-critical calls are handled, and who checks them.
- What result would make you keep it, and what would make you stop. Write the threshold down.
- Who reviews the numbers each week, and when the final decision is made.
An overflow-only test is often the fairest place to start. It measures exactly the calls you were missing, and it leaves the calls your team already answers untouched.
Step one: the baseline
Before changing anything, record at least four normal weeks. Avoid the Christmas period or the first cold snap of autumn unless that is exactly what you want to test. You need the following, by week and, for calls, by hour of the day.
- Inbound calls received, answered live, sent to voicemail and abandoned. Your phone provider's call log or portal shows these.
- How many missed callers were called back, and how long the callback took.
- Genuine new enquiries, separated from suppliers, sales calls and existing jobs.
- Enquiries that became booked visits or quote appointments.
- Average gross profit per booked job, by work type, from your job system or accounts.
- Time engineers and the owner spent on the phone during the working day, estimated honestly if not measured.
- Out-of-hours calls and what happened to them.
Tag every enquiry at the moment it arrives: type of work, new or existing customer, how it came in. Tagging from memory at the end of the week is where most measurement goes wrong.
Step two: change one thing
Introduce the tool and change nothing else you can avoid: same marketing, same prices, same opening hours. If you also launch a new website and a spring promotion in the same month, you will not know what caused the difference.
Run the test for at least the same length as the baseline, ideally longer. If seasonality makes a straight comparison unfair, compare the change in your numbers with the change over the same weeks last year. For example, if enquiries rose 10% year on year before the tool and 25% year on year during the test, the tool's effect is closer to 15 points than 25.
Step three: the calculation
Extra answered calls
Answered calls per week during the test, minus the baseline, adjusted for any seasonal change.
Extra booked jobs
Booked jobs from new enquiries per week during the test, minus the baseline. This is the number that matters most.
Gross profit gained
Extra booked jobs multiplied by average gross profit per job. Use gross profit, not turnover.
Time value
Hours of engineer and owner time saved per week, multiplied by what that time is worth. For an engineer, use the gross profit of an extra billable hour; for the owner, be realistic about what they do with the time.
Full cost
Subscription or service fees, call charges, set-up, your time configuring it and any extra staff time checking its work.
Return
Return = (gross profit gained + time value - full cost) divided by full cost. Work it out weekly and over the full test period.
Example: an eight-week test
Example: a fictional three-engineer firm in Reading measures four baseline weeks in May. It receives 160 inbound calls a week, answers 104 live (65%) and books 38 new jobs a week. Average gross profit per new job is £95. Engineers estimate they spend four hours a week between them answering calls on site.
Example test: it adds a call-handling tool for eight weeks in June and July. Answered calls rise to 150 of 162 a week (93%) and new bookings to 46 a week. Last year, bookings in the same weeks were about 3% higher than in May, so the firm credits the tool with about 7 extra jobs a week rather than 8.
Example result: 7 extra jobs multiplied by £95 is £665 of gross profit a week. Engineer time saved, four hours at £40 of gross profit per billable hour, adds £160. If the full cost of the tool, including the owner's set-up time, works out at £150 a week, the return is (£665 + £160 - £150) divided by £150, about 4.5 times the cost. Even counting only the extra jobs, it is more than three times.
Example caution: the test ran in summer. The firm repeats the measurement in November, when calls are heavier and more urgent, before deciding whether to keep the tool through winter.
Check the quality, not just the volume
More answered calls only help if the calls are handled well. Across UK businesses using AI, 84% reported at least some human checking of its outputs and 67% reported significant checking[2]. Apply the same discipline to any call handling, human or automated.
- Listen to or read a sample of handled calls each week: were details complete, was the job described accurately, was the caller treated well?
- Check every gas-smell or carbon monoxide call was handled with the correct safety message and routed to the right emergency line.
- Track jobs where the engineer arrived without the right information or parts because of how the call was taken.
- Ask a handful of new customers how the first call felt.
- Note any complaints that mention the phone.
Ava, for example, answers every call, takes down the job, handles gas-smell calls with a safety-first message and sends the job to your team, which makes each of these checks straightforward to run against its call records.
Value beyond the first job
A new customer who books a repair may go on to take a service plan and eventually a replacement. If your customer lifetime value for a serviced homeowner is well above the first job's profit, a short test understates the return. Include lifetime value as a second, clearly labelled figure rather than mixing it into the main calculation, so the decision does not rest on optimistic assumptions.
Recurring customers matter too. A plan holder whose breakdown call goes unanswered may cancel; protecting recurring revenue is part of the return even though it never shows up as a new booking.
Common mistakes
- No baseline, so any improvement is a guess.
- Counting all answered calls as new business, including suppliers and existing jobs.
- Ignoring seasonality and crediting the tool with the arrival of autumn.
- Using turnover instead of gross profit, which inflates the return.
- Leaving out the owner's set-up and checking time from the cost.
- Deciding after one good week.
Ease of use matters to adoption: 19% of construction businesses named ease of integration and use as something that would support their use of AI[6]. Factor in how much of your time a tool needs before the numbers come in.
Frequently asked questions
How long should a call-handling trial run?
At least four weeks of baseline and four weeks of test, ideally longer, and repeated in a different season. Heating demand changes so much through the year that one short test can mislead.
What is the most important number to measure?
Extra booked jobs from new enquiries, multiplied by average gross profit per job. Answered calls are a useful signal, but bookings are what pay for the tool.
Should time saved count in the return?
Yes, but value it realistically. Engineer time is worth the gross profit of an extra billable hour; owner time is only worth something if it goes into work that earns.
How do we check an AI phone assistant is handling calls properly?
Review a sample of calls every week, check every safety-critical call, and track jobs where the engineer arrived without the right information because of how the call was taken.
References
Office for National Statistics (2026). Business insights and impact on the UK economy: wave 165 dataset (AI tables) (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑
Office for National Statistics (2023). Business insights and impact on the UK economy: wave 165 dataset (AI tables) (opens in a new tab) ↑
Office for National Statistics (2026). Business insights and impact on the UK economy: wave 165 dataset (AI Current Purpose) (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑
Office for National Statistics (2026). Business insights and impact on the UK economy: wave 165 dataset (AI Use Factors) (opens in a new tab) ↑