Key takeaways
- Only 21% of UK businesses had adopted some AI tools in 2025/26, while 45% said AI was not relevant to them, so a careful pilot puts you ahead without betting the business[1].
- Pick one task in one branch with high volume, low legal risk and an easy measure, such as out-of-hours calls or inspection write-ups.
- Do the groundwork first: a named owner, a data protection screening, supplier terms, a staff briefing and wording for clients[2].
- Decide go or no-go on day 30 against thresholds you wrote down before day 1, and listen to the people who used the tool.
Why one branch, one task and thirty days
AI projects in agencies usually fail in one of two ways. Either the owner buys a tool, switches it on everywhere and spends months dealing with the fallout, or the tool sits half-configured because nobody owns it. A tightly scoped pilot avoids both. One branch keeps the impact contained and makes the results easy to compare with others. One task keeps the measures clear. Thirty days is long enough to get past the novelty and short enough to keep attention.
The aim is a decision, not a demonstration. At the end you should know whether the tool does the job well enough to extend, needs adjusting, or should be stopped, and you should be able to explain why with numbers and with the team's experience.
Where most firms are starting from
The government's 2025/26 cyber security breaches survey found that 21% of UK businesses had adopted some AI tools, while 45% said AI was not relevant to their organisation[1]. The ONS reports that 41% of businesses with ten or more staff saw no barriers to adoption in mid-2026, and that the most common barriers were difficulty identifying use cases, cost and lack of expertise[3]. A pilot deals with all three: it names the use case, limits the cost and builds expertise inside the team.
Source: Department for Science, Innovation and Technology (GOV.UK), 2026[1]
Show the data
| Item | Value |
|---|---|
| Adopted some AI tools | 21% |
| Said AI was not relevant | 45% |
Choosing the task
| Candidate task | Volume | Legal risk | Easy to measure? | Pilot verdict |
|---|---|---|---|---|
| Out-of-hours call answering and logging | High | Low to medium | Yes: answered calls, bookings, callbacks | Good first pilot |
| Inspection and valuation write-ups from voice notes | Medium | Low | Yes: time per report, corrections | Good first pilot |
| Listing copy from verified fact sheets | Medium | Medium | Partly: time saved, corrections | Good with strict checks |
| Applicant enquiry replies | High | Medium | Yes: response time, viewings booked | Good with approved facts |
| Maintenance triage for managed lets | Medium | Medium to high | Yes: response times, re-tiering | Second pilot, after experience |
| Tenant scoring or selection | Medium | High | Hard: fairness needs long testing | Not a pilot task |
Choose the task that causes the most visible pain today and scores well on the other columns. If the branch manager cannot describe the problem in one sentence, such as 'we miss most calls after half past five', the task is not ready for a pilot.
Groundwork before day one
- Name a pilot owner in the branch with time set aside each day, and a sponsor who will make the final decision.
- Screen the processing for a data protection impact assessment; UK GDPR requires one before new technology processing that is likely to result in a high risk to people[2].
- Get processor terms from the supplier, confirm where data is held and switch off any use of your data for training if you can.
- Prepare the approved facts the tool will use: property sheets, office hours, fees page links and escalation contacts.
- Brief the branch team on what will change, what will not, and how to report a problem.
- Write the client-facing wording, such as a call greeting or a line in the email footer.
Get pilot-friendly contract terms
A pilot is only a pilot if you can walk away. Before signing, check that the agreement matches a 30-day decision rather than locking you into a year.
- A monthly or pilot-length term, with no automatic roll into an annual contract.
- Set-up help and a named contact included for the pilot period.
- The right to export your data in a usable format and have the rest deleted if you stop.
- Clear limits on any usage-based charges, so a busy week does not produce a surprise bill.
- Agreement that pilot results, including errors, can be shared inside your business.
The 30-day plan
Days minus 14 to 0: baseline
Record the measures for two normal weeks: for call answering, calls by hour, answered, missed and returned, plus valuations and viewings booked by phone.
Days 1 to 3: set-up and shadow mode
Configure the tool and run it alongside the current process without clients relying on it, or on test calls only. Fix obvious errors.
Days 4 to 10: live with full review
Switch on for real clients in the agreed hours. The owner reviews every output or call record daily and logs every error.
Days 11 to 20: spot checks
If the error log is short, move to reviewing a daily sample. Adjust scripts and approved facts weekly.
Days 21 to 28: steady state
Run as you would after rollout. Keep the daily sample and record time spent on review.
Days 29 and 30: decide
Compare results with the baseline and thresholds, gather the team's views and make the go, adjust or stop decision.
Daily and weekly checks
- Daily: read the overnight records, confirm anything urgent was handed to a person, and log errors by type.
- Daily: check that records reached the CRM against the right contact and property.
- Weekly: update the measures against the baseline and share them with the branch team.
- Weekly: listen to or read five complete interactions start to finish, including at least one the tool handled badly.
- Weekly: review client feedback and any complaint that mentions the new process.
Bring the team with you
Expect mixed feelings. In the government's public attitudes survey, around four in ten UK adults expected AI to have a positive impact and three in ten a negative one[4]. Your team is likely to split in a similar way. Involve sceptics in writing the scripts and reviewing outputs; they find problems faster than enthusiasts. Make clear that the pilot is about removing the work people dislike, not about cutting jobs, and keep that promise in how results are used.
Ask for a two-minute note from each person at the end of each week: what helped, what went wrong and what they would change. Those notes often explain the numbers better than the numbers explain themselves.
Stop rules
Write down in advance what would pause the pilot immediately. That protects clients and gives the owner permission to act without a debate.
- Any safety-related failure, such as an emergency repair or gas report not passed to a person.
- A data protection incident, such as client data sent to the wrong person.
- A pattern of wrong information given to clients about properties, prices or rents.
- A complaint that the tool refused or obstructed a request to speak to a person.
- Review time rising instead of falling after the first fortnight.
The go/no-go scorecard
| Criterion | Threshold set before day 1 | Example result |
|---|---|---|
| Out-of-hours calls answered | At least 90%, from a baseline of 35% | Example: 97% |
| Valuations booked from those calls | At least 4 more a month than baseline | Example: 6 more |
| Urgent issues passed to a person | 100% within 15 minutes | Example: 100% |
| Errors needing correction | Under 1 in 20 records by week 4 | Example: 1 in 32 |
| Daily review time | Under 30 minutes by week 4 | Example: 20 minutes |
| Team view | Majority want to keep it | Example: 5 of 6 |
If every threshold is met, extend to a second branch with the same measures. If most are met, adjust and run another 30 days. If the safety or error thresholds fail, stop and record why. For turning results into money, see measuring AI return in an agency.
Worked example: out-of-hours calls in one branch
Example: a fictional branch finds in its baseline that 41% of calls arrive after half past five or at weekends and that only about a third of those are answered or returned the same day. It pilots AI call answering for evenings and Sundays, with valuations booked into the valuer's diary, viewing requests passed to negotiators and repairs logged with urgent flags to the out-of-hours contact.
Example, continued: in week one the owner corrects several call notes where property names were misheard, and adds missing parking details to three property sheets. By week four the error rate is low, review takes twenty minutes a day and the branch has booked noticeably more valuations from evening calls. The team votes to keep it. The sponsor extends the pilot to the second branch, keeping the same scorecard. Before switching on, both branches work through an AI call answering checklist.
Write it up
Whatever the decision, write a one-page summary: the task, the baseline, the results against each threshold, the main errors and fixes, the team's view and the decision. Store it with the data protection assessment. When the next tool comes along, the summary is the template, and the agency gets better at piloting each time. The same discipline helps when you review bigger systems, such as an estate agent CRM.
Frequently asked questions
How long should an AI pilot in an estate agency last?
About 30 days of live use after two weeks of baseline measurement. That is long enough to get past early teething problems and short enough to keep the team focused on the decision.
Which task should we pilot first?
A high-volume, low-risk task that is easy to measure, such as out-of-hours call answering or inspection write-ups. Avoid tenant selection or anything that makes significant decisions about people.
Do we need a DPIA for a pilot?
Screen the processing before you start. If it is likely to result in a high risk to people, for example recording every call, complete an impact assessment before going live, even for a pilot.
What if the results are mixed?
Adjust the set-up, such as scripts, approved facts or hours, and run a second 30 days with the same scorecard. Stop if safety or error thresholds fail.
References
Department for Science, Innovation and Technology (GOV.UK) (2026). Cyber security breaches survey 2025/2026 (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2026). UK GDPR, Article 35: Data protection impact assessment (opens in a new tab) ↑
Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026 (opens in a new tab) ↑
Department for Science, Innovation and Technology (GOV.UK) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (opens in a new tab) ↑