Key takeaways
- Measure a two-week baseline before switching anything on; without it, every result is a guess.
- Count three kinds of return: staff time released, revenue protected or added, and risk reduced, and subtract the time spent reviewing the tool.
- Cost time properly: employer National Insurance is 15% above the secondary threshold, and holiday removes 5.6 weeks a year[1][2].
- Most AI-using UK businesses report no change in headcount, so judge the tool on output and service, not on jobs removed[3].
Why most AI business cases fall apart
Supplier business cases usually multiply minutes saved by an hourly rate and a large number of tasks, then present the total as profit. Real agencies rarely see that money. The minutes are spread across many people, so nobody's job actually changes; the time spent checking the tool's output is left out; and the same benefit is counted twice, once as time saved and again as leads gained.
A useful measurement is narrower and more honest. It picks one task, records what happens today, runs the tool for long enough to get past the novelty, and compares like with like. It also asks whether the result changed anything you care about: instructions won, lets agreed, complaints avoided or hours that can be put to better use.
Three kinds of return
- Time released: hours of typing, chasing, call handling or rekeying that people no longer do. Only real if the hours go to something valuable, or reduce overtime or temporary cover.
- Revenue protected or added: enquiries answered that would have been missed, valuations booked faster, applicants converted to viewings. Count only what you can trace to the tool.
- Risk reduced: fewer errors, missed deadlines and complaints. Harder to value, but often the most important in lettings.
Against these sit the costs: the subscription, set-up time, the hours staff spend reviewing outputs, and any extra work the tool creates, such as correcting CRM entries. Review time is the cost most often left out, and in the first month it can be larger than the time saved.
Set the baseline first
Choose three to five measures
Pick measures tied to the task, such as missed calls, time to first reply, valuations booked or hours spent on write-ups.
Decide how each is captured
Use the phone system, CRM reports or a simple time log. Avoid measures that rely on memory.
Record two normal weeks
Avoid bank holidays and unusual weeks. If the market is seasonal for you, note it.
Write the baseline down
Share the numbers with the team so everyone agrees on the starting point before the tool arrives.
Agree the success threshold
Decide in advance what result would justify keeping the tool, so the decision is not made on feel.
Measures that work in agencies
| Measure | How to capture it | What to watch |
|---|---|---|
| Missed or abandoned calls | Phone system report by hour of day | Callers who rang back later are not lost leads |
| Time to first reply to portal leads | CRM timestamps | Automated acknowledgements do not count as replies |
| Valuations booked per week | Diary and CRM | Market seasonality; compare with last year too |
| Viewing requests to viewings held | CRM | Requests the tool takes still need confirming |
| Hours on write-ups and notes | Two-week time log | Include time spent correcting drafts |
| Complaints and redress cases | Complaints log | Small numbers; look over a longer period |
| Compliance deadlines missed | Compliance tracker | One miss can outweigh many hours saved |
Costing staff time properly
An hour of staff time costs more than the hourly wage. The National Living Wage for workers aged 21 and over is £12.71 an hour from April 2026, rising to £13.21 in April 2027[4], and the National Careers Service puts estate agent pay at £23,000 for starters rising to £40,000 for experienced agents[5]. On top of pay, employers pay National Insurance at 15% on earnings above the secondary threshold of £5,000 a year[1], and almost all workers are entitled to 5.6 weeks' paid holiday[2].
Example: a fictional negotiator earns £30,000. Employer National Insurance is 15% of £25,000, or £3,750. Working 37.5 hours a week for 46.4 weeks after holiday gives about 1,740 hours, so the loaded cost is roughly £19.40 an hour before pension, bank holidays and overheads. Use your own figures, but always load the wage.
Smaller employers should also check the Employment Allowance, which reduces an eligible employer's National Insurance bill by up to £10,500 in 2026 to 2027[1]. If the allowance already wipes out your bill, the marginal saving from reducing hours is lower than the headline rate suggests.
Valuing the revenue side
Revenue gains need a chain of conversion rates, each from your own data. In August 2026 a typical Propertymark sales branch registered 69 new buyers, placed about 10.8 homes on the market and agreed 9.5 sales[6]. Your numbers will differ, but the shape is the same: many enquiries, a few instructions, fewer completions. A tool that answers more calls only adds revenue if some of those calls become valuations, instructions and completed sales.
Example: a fictional branch finds that out-of-hours calls answered by its new tool produce 6 extra valuation bookings a month. Its records show it wins one instruction in three valuations and completes 70% of instructions, with an average fee of £3,200. Six bookings become two instructions and about 1.4 completions, or roughly £4,480 of fee income a month, received months later. That is the figure to compare with the tool's cost, not the number of calls answered.
A simple payback model
- Monthly time value: hours released multiplied by loaded hourly cost, counting only hours that are reused or remove overtime or cover.
- Monthly revenue value: extra completions you can trace to the tool multiplied by average fee, discounted for the delay to completion.
- Monthly risk value: a cautious estimate, such as one avoided complaint award per quarter, or leave it at zero and treat it as a bonus.
- Monthly cost: subscription, plus review hours at loaded cost, plus any extra admin.
- Payback: one-off set-up cost divided by monthly value minus monthly cost.
If the model only works with optimistic conversion rates or by counting every minute saved, the tool is not paying for itself yet. That does not always mean stop; it may mean narrowing the task or improving the set-up.
Headcount is the wrong yardstick
Owners often expect AI to remove a role. The national evidence suggests that is rare. Among UK businesses using AI to improve operations, 63% reported no change in headcount, 6% a decrease and 1% an increase[3]. The ONS also lists difficulty identifying use cases, cost and lack of expertise as the main barriers to adoption[3], which matches what agencies find: the hard part is choosing the right task, not the technology.
Judge the tool on what the same team can now do: more valuations from the same enquiries, faster replies, fewer missed deadlines and less weekend firefighting. Those gains show up in revenue and staff retention, even if the payroll stays the same.
Pitfalls that distort results
- Seasonality: comparing a quiet December baseline with a busy January pilot will flatter any tool.
- Novelty: staff use a new tool carefully in week one and less carefully by week six; measure for at least four weeks.
- Double counting: if extra valuations are counted as revenue, do not also count the time spent booking them as a saving.
- Market noise: if prices or rates move sharply during the pilot, compare against another branch or the same weeks last year.
- Survivor bias: look at the conversations the tool handled badly, not just the ones that went well.
Who should own the numbers
Give the measurement to one person who does not sell the tool internally and does not stand to lose from a bad result. In a small agency that might be the office manager or the owner's bookkeeper; in a larger group, someone from finance or operations. Their job is to collect the baseline, keep the time logs honest, chase the CRM reports each week and write the one-page summary at the end.
Agree a short weekly check-in of fifteen minutes with the branch manager to look at the numbers and the error log together. Problems found in week two can be fixed; problems found at the end of the pilot only explain why it failed.
Making the decision
At the end of the pilot, put the baseline, the results and the payback figure on one page and share it with the people who used the tool. Ask them three questions: did it save you time you could use, did it make mistakes that worried you, and would you want it to stay? If the numbers and the team agree, extend it. If they disagree, find out why before spending more.
For the pilot itself, see our 30-day plan for piloting AI in one branch, and for the wider numbers worth tracking every month, our guide to estate agency KPIs.
Frequently asked questions
How do you calculate the return on an AI tool in an agency?
Measure a baseline, run the tool on one task for at least four weeks, then compare time released at loaded staff cost, traceable revenue and reduced risk against the tool's cost and the time spent reviewing it.
How long should an AI pilot run before deciding?
At least four weeks after set-up, and longer if the measure is seasonal or rare, such as complaints. Two weeks of baseline beforehand is the minimum.
Should time saved be counted as money saved?
Only if the time is reused for valuable work or reduces overtime, temporary cover or recruitment. Minutes spread across a team often disappear without changing anything.
What if the tool helps but does not pay back yet?
Narrow the task, fix the set-up issues the pilot revealed and measure again. If it still does not pay back after that, stop and record why.
References
HM Revenue & Customs (GOV.UK) (2026). Rates and thresholds for employers 2026 to 2027 (opens in a new tab) ↑
GOV.UK (2026). Holiday entitlement: Entitlement (opens in a new tab) ↑
Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026 (opens in a new tab) ↑
GOV.UK (2026). National Minimum Wage and National Living Wage rates (opens in a new tab) ↑
National Careers Service (GOV.UK) (2026). Estate agent job profile (opens in a new tab) ↑
Propertymark (2026). Housing Insight Report, August 2026 (opens in a new tab) ↑