Case Study
A London-based letting agency with 28 staff and 6 branches automated 71% of new tenant enquiries with a custom AI chatbot integrated to Reapit and Alto, cutting average first-response time from 4 hours 18 minutes to 60 seconds and capturing 38% more after-hours leads in the first quarter after launch.
Independent London letting agency, 28 staff, 6 branches
A London-based letting agency with 28 staff and 6 branches automated 71% of new tenant enquiries with a custom AI chatbot integrated to Reapit and Alto, cutting average first-response time from 4 hours 18 minutes to 60 seconds and capturing 38% more after-hours leads in the first quarter after launch.
An independent London letting agency operating six branches across Zone 2 and Zone 3 came to Softomate with a problem they could see in their own data but had no clear way to solve. Their average response time to a new tenant enquiry, measured from the moment an enquiry left a property portal to the moment a human at the agency replied with a viewing slot, was 4 hours 18 minutes during weekday office hours and over 17 hours outside of them. They knew, from their own conversion analysis, that the probability of converting a portal enquiry into a booked viewing fell off a cliff after the first 5 minutes. By 30 minutes, conversion was less than half what it could be at 60 seconds. By 4 hours, almost two-thirds of enquiries had either booked a viewing with a competitor or gone cold.
The agency was not negligent. Their negotiators were genuinely busy. Each of the 14 negotiators across the six branches handled an average of 9 active tenancies, 18 ongoing viewing arrangements, and 22 inbound enquiries per day. Manually triaging a Rightmove or Zoopla enquiry took on average 7 minutes: open the email, copy the property reference, log into Reapit, check current availability, check whether the property had been let, check the viewing calendar for that branch, draft a personalised reply, send it. Multiplied across 22 enquiries per day per negotiator, that workload alone consumed 2 hours 34 minutes of focused time, before any showings, contract work, landlord calls, or tenant referencing.
The agency had tried two interim solutions, neither of which worked. The first was a third-party live-chat tool placed on their website, manned during office hours by a virtual receptionist service. It generated transcripts but rarely produced qualified leads, because the receptionists could not see the property database, could not check availability, and could not book viewings. They could only collect contact details, which negotiators then had to follow up on manually, recreating the original delay. The agency cancelled the contract after four months.
The second was an off-the-shelf chatbot widget marketed to estate and letting agencies, which the agency trialled for three months. The widget was effectively a glorified FAQ tool. It answered general questions about the company's services but had no live connection to Reapit, no awareness of which properties were still on the market, and no ability to suggest available viewing times. A prospect asking about a specific listed property would get a generic reply about contacting the office. Conversion through the widget was so low it was effectively zero, and the widget was removed.
What the agency needed, but did not initially have the words to describe, was a chatbot that behaved like an extension of the negotiator team. It needed to recognise which property a prospect was enquiring about, know whether that property was still available, suggest specific viewing slots from the branch calendar, and write back to the prospect in a tone that matched the agency's brand. It needed to escalate properly when a query required human judgment, and it needed to feed the conversation context back into Reapit so that when a negotiator picked up the lead the next morning, they had the full picture without having to re-read three days of chat transcript.
A second, less visible problem compounded the response-time issue: after-hours enquiry capture. The agency's data showed that 41% of all portal enquiries arrived outside of office hours, with a strong peak between 7pm and 11pm on weekday evenings and across Sunday afternoons. Every one of those enquiries was answered at best the next morning, often the following Monday, by which point the prospect had moved on. The agency was effectively giving away two days of conversion opportunity each week with no realistic prospect of recovering it without either a 24/7 staffing model that would cost roughly £72,000 a year in additional salaries, or an automated response system that could acknowledge and qualify enquiries at the moment they arrived.
The managing director's brief to Softomate was deliberately commercial: cut response time to under 5 minutes on average across all hours, capture and qualify a meaningful share of after-hours enquiries automatically, do not damage the brand voice, and integrate cleanly with Reapit and Alto so the negotiators kept a single source of truth. The build had to be invisible to landlords. The chatbot needed to feel like a competent member of the team to prospects, not a clearly-labelled automation that prompted them to wait for a human.
Two seasonal pressure points shaped the brief further. London's lettings market produces a clear demand peak each September as students and graduate tenants relocate, and a second peak from January through March as job movers reset. During those two windows, daily enquiry volume across the six branches rose to over 380 enquiries on the heaviest days, a 70% spike on baseline. In the previous September peak, the agency had recorded a 51% missed-enquiry rate over the first ten days of the academic term, and the lettings director was clear that without an automated layer, the September peak in the year of the engagement would be impossible to handle without either turning away landlord instructions or signing a temporary staffing contract that historically had cost the agency around £11,400 per month for the eight-week peak window.
The competitive landscape was a third pressure. The London lettings market is unusually transparent because most prospects send the same enquiry to three or four agencies simultaneously through the portal multi-enquiry feature. The agency that replies first with a qualified, time-specific viewing slot wins disproportionately often, regardless of whether their property is objectively the best fit. Foxtons, Hamptons, and several of the larger independents had already begun investing in automated qualification at the top of the funnel, and the lettings director had monitored several enquiries internally in which a Foxtons reply landed within 11 minutes while the agency's own reply landed 6 hours later. The market direction was clear; the only question was whether the agency would lead or follow.
Softomate began with a two-week discovery phase, working onsite with one branch and remotely with the other five. The first task was to map the actual enquiry lifecycle, from arrival in a portal to a confirmed viewing or a closed lead. The team analysed 4 weeks of historical enquiry data, 1,847 enquiries in total, categorising each by source, query type, property type, time of arrival, response time, and eventual outcome. This produced a clear priority matrix: 71% of all enquiries could in principle be answered by an automated system with live read access to Reapit, viewing calendar API access, and a structured set of qualifying questions. The remaining 29% genuinely required negotiator judgment: complex referencing questions, landlord-side decisions, multi-property comparisons across portfolios, and any query that required calling a third party.
The architecture Softomate designed was specifically built around the fact that a letting enquiry is a data-rich query masquerading as a chat message. A prospect typing “Is the flat on Acton Lane still available for the 14th?” is not really asking a conversational question. They are asking the agency to query its property database with two filters, check the viewing calendar of one branch, and reply with a clear yes or no plus a confirmed time slot. The chatbot needed to do exactly that, fast, and write back in a way that sounded like a real negotiator answering the phone, not a robot acknowledging a form.
The chatbot itself was built on GPT-4o, with a retrieval-augmented generation layer that gave it live access to three data sources: the agency's full property listings via the Reapit Foundations API, branch-level viewing calendars via a custom integration with their viewing diary, and a knowledge base of agency policies (rent guarantees, holding deposit terms, referencing process, deposit protection scheme, fees policy, pet policy by branch). The vector store was structured so that semantic search returned the correct policy snippet even when prospects phrased questions in unusual ways. A prospect asking “Do I need to put money down to hold a viewing?” and a prospect asking “What's the deposit policy?” were both routed to the same retrieved policy passage, then reframed in the brand voice.
Brand voice calibration was a critical part of the build and took longer than any other aspect of training. Softomate collected 240 of the agency's highest-rated historical email responses to prospect enquiries, sampled across the six branches, and used them as the tone calibration corpus for the model. Before launch, the agency's lettings director reviewed 300 sample chatbot responses to varied enquiry types and either approved them or sent them back with specific tonal corrections. By the end of the third review round, approval rate exceeded 96% on first generation.
The chatbot was embedded as a floating widget on the agency's main website and on each branch microsite, and was also wired to receive WhatsApp inbound messages routed through Twilio for the agency's listed business number. When a prospect started a conversation, the chatbot first identified which property they were enquiring about (either from the URL context if they came from a property listing page, or by asking a clarifying question if not), pulled the live availability from Reapit, and either offered specific viewing time slots from the relevant branch's calendar or, if the property was let, suggested up to three similar available properties pulled from the same vector store ranked by relevance.
For escalations, the chatbot was trained to recognise four categories of query that required human handover: anything involving a landlord-side decision, complex referencing questions, multi-property comparison requests across portfolios, and any query where the prospect explicitly asked for a human. On escalation, the chatbot generated a structured summary of the conversation including the prospect's name, contact details, the property of interest, the qualifying information already gathered, and the specific outstanding question. This summary was written into a Reapit case note and sent simultaneously by SMS to the on-duty branch negotiator and posted to the agency's Slack lettings channel. By the time a negotiator picked up the lead, they had every piece of context they needed in front of them, with no need to re-read transcripts.
Deployment was deliberately staged. Softomate ran a single-branch pilot through week 3, monitoring every conversation in real time and reviewing escalations daily with the branch manager. Eleven targeted refinements were made during the pilot, mostly to the policy knowledge base where the chatbot had given technically correct but tonally awkward answers about deposit protection. By the end of week 4, the chatbot was live across all six branches.
The technical model choice deserves a note, because it was a deliberate decision rather than a default. Softomate evaluated three options in the first week: GPT-4o, GPT-4o-mini, and Claude 3.5 Sonnet. GPT-4o-mini cost roughly a tenth of GPT-4o per million tokens and would have been the obvious default for a high-volume use case, but in side-by-side testing on 80 representative letting enquiries it produced tonally inconsistent replies and was noticeably weaker at recognising when a query genuinely required escalation versus when a policy answer was sufficient. Claude 3.5 Sonnet matched GPT-4o on tone but introduced an additional vendor relationship and a higher complexity in the RAG layer for this particular use case. GPT-4o was selected on tone reliability and integration simplicity, with GPT-4o-mini retained as a fallback model for the lower-stakes follow-up question flow where cost matters more than tonal nuance. Per-interaction model cost at typical conversation length came out to roughly £0.04, well inside the agency's economic threshold for the service.
GDPR and data residency considerations were addressed before any live data flowed. Tenant enquiries contain personal data under UK GDPR, and the agency had a clear preference for keeping that data within UK or EU jurisdictions wherever possible. Softomate routed all model calls through the OpenAI Europe endpoint, configured the OpenAI organisation account to opt out of training data use on the prompts and completions, and held the vector store of property and policy data on a UK-region instance of the underlying database. A formal Data Processing Agreement was executed between the agency and Softomate, and a separate DPA was held with OpenAI as the sub-processor. The agency's Data Protection Officer reviewed and approved the data flow before pilot launch. Prospect chat transcripts are retained for 90 days in line with the agency's existing privacy policy, then permanently deleted via an automated cleanup job.
Within the first 90 days of full deployment across all six branches, the measurable outcomes were unambiguous. Average first-response time across all enquiry sources fell from 4 hours 18 minutes to 60 seconds, a reduction of more than 99.6%. The chatbot resolved 71% of all inbound enquiries to a fully qualified, viewing-booked state without any human intervention at any point. The remaining 29% were escalated to negotiators with full context, and negotiator average response time on those escalations also improved from 2 hours 11 minutes to 23 minutes, because they no longer had to wade through unread enquiry inboxes to triage what needed attention.
The after-hours capture metric showed perhaps the most striking shift. Before deployment, the agency captured contact details on roughly 8% of after-hours portal enquiries, almost all of which were re-engaged the following morning with mixed results. After deployment, the chatbot qualified and booked viewings for 38% of after-hours enquiries directly, without any human involvement. Many of those bookings were for properties prospects had viewed within minutes of the listing going live. In several cases, prospects booked viewings for the following morning before the branch had opened, completed referencing during the next working day, and signed within 48 hours of first contact, a sequence that would have been impossible with the old workflow.
The missed-leads metric, defined by the agency as enquiries that went 24 hours or more without a response, fell from 62% pre-deployment to 11% post-deployment. The remaining 11% were almost entirely escalations that fell during weekend periods where the on-duty negotiator coverage gap was not fully closed by the chatbot escalation policy. The agency has since refined the escalation rules to forward Saturday evening complex enquiries to a Sunday morning duty negotiator, and that figure is expected to fall further in the next quarter.
The conversion metric the lettings director was watching most closely, viewing booking rate per enquiry, rose from 19% to 42%. This was the metric that translated most directly into commercial impact. At the agency's average viewing-to-let conversion rate of 33%, the lift in viewing rate alone added an estimated 84 additional confirmed tenancies in the 90-day window, worth around £680,000 in attached management fees across the full tenancy term, against a total Softomate engagement cost recovered within the first 71 days.
The chatbot also surfaced operational intelligence the agency had not previously captured. Because every conversation was logged with structured intent tagging, the lettings director could see clearly which property categories generated the most enquiry friction. A pattern of repeated questions about pet policy at one branch led to a decision to update the listing copy with explicit pet policy lines on every property in that area, which removed those queries entirely from the chatbot's workload within two weeks.
The branch managers reported, in a structured review session 60 days after full deployment, that the change in their negotiators' working day was more significant than they had expected. The two hours and 34 minutes that each negotiator had previously spent on enquiry triage was now spent on viewings, follow-ups, and tenant referencing. Two branches that had been planning to hire an additional junior negotiator put those hires on indefinite hold; the existing team had effective capacity for the volume the chatbot was now passing through.
One outcome the lettings director called out in the post-launch review session was tonal consistency. Before the chatbot, response quality across the six branches varied significantly depending on which negotiator handled a given enquiry. After deployment, every prospect received a first response in the same brand voice, calibrated to the agency's house style, regardless of which property they enquired about or which branch held the stock. The lettings director described it as the first time the agency had a genuinely consistent prospect experience at the top of the funnel.
The chatbot has since been extended to handle landlord-side enquiries about new instructions, a workflow that adds approximately £14,000 of estimated annual time savings on top of the original tenant-enquiry savings. Softomate provides quarterly review sessions and adds new intent categories as the agency's needs evolve. The total engagement cost was fully recovered within 71 days of go-live, well inside the six-month payback the managing director had originally been prepared to accept.
Three things, with the benefit of hindsight, would have made the build faster or smoother. First, the agency's property data in Reapit had three years of accumulated inconsistencies in the postcode and parking fields that the chatbot initially struggled with, returning generic responses for genuinely answerable queries. Softomate added a data hygiene pass in week 5 that cleaned the worst-affected 460 records, and the chatbot's first-response accuracy on those properties immediately rose. Future engagements should include data quality as an explicit week-one workstream rather than a discovered issue. Second, the brand-voice corpus would have been stronger with senior negotiator pull-quotes about why their best replies worked, not just the replies themselves. The model improved noticeably once Softomate added a short rationale to each calibration example. Third, weekend escalation coverage was under-scoped. The original policy assumed Saturday morning coverage by default; that has since been extended to a full weekend rota for escalations of any kind, and the missed-leads metric is expected to compress further as a result.
The next phase of work, already in scope for the following quarter, will extend the chatbot to handle property valuations for landlord prospects, a workflow that historically converted at around 4% from first enquiry to instructed valuation appointment. The hypothesis is that the same speed-and-context advantage the chatbot demonstrated on tenant enquiries will translate to landlord enquiries, where time-to-first-response is also a strong predictor of instruction conversion. Softomate is also building a parallel WhatsApp-only flow for the agency's growing share of repeat tenants who prefer to deal in WhatsApp messages rather than email, which Reapit's native channels do not currently support cleanly. A second engagement is in active scoping with the agency's landlord-services arm to apply the same approach to inbound rent guarantee claim queries, where the existing process averages 3 working days to acknowledge and the agency suspects an automated triage layer would compress that to under an hour for the 60% of cases that follow standard patterns.
Related service:AI Chatbot Development Service London. Further reading:AI Chatbots for London Estate Agents and Estate Agency Software UK Comparison.
Anonymised client engagement. Identifying details modified for confidentiality. Outcome ranges reflect typical results from similar projects.
Names withheld to preserve confidentiality.
Work with us
Every project we take on has a measurable outcome. Talk to our London team and we will show you exactly how we would approach your challenge.
Softomate Solutions Limited
Deen Dayal Yadav
Online