Most AI in residential real estate answers questions. The next step is agents that make and execute operational decisions inside predefined workflows, with every resident and property in one view and the asset strategy as the brief. On a patchwork stack that is dangerous; on a rental operating system with standardised data, documented workflows and human-set approval gates, it is how the levers get pulled on every tenancy, every day, and compound into NOI and asset value.
%202%20(1).png)
AI was supposed to solve all our problems. Eighteen months on, the industry has chatbots answering enquiries, models summarising documents, and copilots drafting emails, yet 70–85% of enterprise AI projects still fail to meet expectations. The boardroom pressure to "do AI" has run into the obstacle we have written about all year: the models are fine, the foundation beneath them is not.
This piece is about what happens on the other side of that obstacle. Not AI that answers questions, but agents that make and execute operational decisions inside predefined workflows, with every resident and every property in one view, and the asset strategy as the brief. It’s what happens when everything resident lives in one system.
Ask any operator what they would trust an agent to do unsupervised, and the honest answer is: nothing that touches money. Draft the email, yes. Set the renewal rent, no. Decide which void to prioritise, no. Trigger a retention incentive, no. The caution is well founded, and it has three causes, none of them the model.
The first is data. Most operational data still lives in a PMS built for accounting, where the property is the master record and the resident is a loose attachment to it: initials in name fields, four residents in one record, contact details that decayed years ago. Layer point solutions on top for listings, leads, viewings, maintenance and comms, and every handover between them leaks: fields that do not map, records that do not reconcile, timestamps that do not align. Hand a model that corpus, and it does not produce poor answers; it produces confidently wrong ones.
The second is workflow. If the process for qualifying a lead or logging a maintenance request depends on who is on shift, you can’t automate it; you can only accelerate a habit. An agent dropped into an undefined workflow scales the inconsistency at machine speed, across the whole portfolio.
The third is trust, and it follows the first two. Teams who have spent years double-checking the PMS against the spreadsheet will double-check the agent too, and they will be right to. Which is why almost every AI product in residential today is what Antony Slumbers calls H1: the same work, done faster and cheaper. Useful, real money, and competed away in fees within a quarter or two, because everyone has the same tool on the same broken stack.
The interesting question is not how to make the chatbot better. It is what becomes possible once you remove those three causes.
An agent is only as trustworthy as the boundaries it operates within. The market’s mistake is to buy an agent and point it at a stack. The right sequence is to define the workflow first, standardise the data it produces, set the rules and the approval gates, and then let the agent run inside them. The workflow is the guardrail. The asset strategy is the brief. The human sits at the gate.
Our rental operating system already has that structure, because it is how the platform has run for years. Every workflow, from lead enquiry to referencing, contract to renewal, notice to re-let, is rules-based and identical across every property, team and geography. The process cannot be followed differently in Manchester than in Birmingham.
What changes with agents is the class of decision the system can take. Automation executes a rule: when the lease is signed, send the move-in pack. An agent weighs a situation: this resident's sentiment has dipped, their maintenance ticket is a week overdue, their rent review is six weeks out, and the local market has softened, so the right move is to fix the ticket today and open the renewal conversation at a smaller uplift. That is a judgement, and in most portfolios it is made late, by a stretched property manager, on a spreadsheet.
Five conditions make that judgement safe to delegate:
Under those conditions, trust in the agent is no longer a leap of faith. It is the same trust the team already places in the workflow.
A chatbot sees one conversation. An agent that drives asset strategy must see the whole portfolio at once, and through the resident, because the resident's behaviour drives nearly every financial outcome in the building.
That is why our master record is the resident, not the property. Every resident has one golden record, built during onboarding and enriched by every interaction afterwards: lease dates, rent and payment history, maintenance requests and how quickly they were resolved, NPS and sentiment, app engagement, the moving-out FAQs they read, the referral they made, the incentive they accepted. Each record is tied to a property that carries its own history: achieved rents, void days per re-let, days to let, Rental Quality Index score, comparables, demand signals. The result is a live, comparable dataset in which a void rate means the same thing everywhere. Insights+ already turns it into leaderboards, alerts and management reporting; an agent turns it into decisions.
The brain, then, is not a model bolted to a data warehouse. It is the operating system's own view of the portfolio, continuously refreshed by the workflows that run on it, read against the fund-level KPIs and asset strategy the owner has set. It knows which 40 tenancies in the portfolio are 90 days from rent review and carry a renewal-risk score above the threshold. It knows which three schemes are letting slower than their comparables and why. It knows which voids are about to become expensive. And it knows what the workflow permits it to do about each one.
That is the difference between a tool that helps a person work faster and a system that works the portfolio.
The tenancy is a loop, not a line, and the most expensive weeks of it - the rent review, the renew-or-leave decision, the notice period, the void and the re-let - are the weeks most portfolios still run on spreadsheets and month-old reports. Here is what the loop looks like when agents run it inside the workflow, with a human at every gate that matters.
Retention: the agent acts while retention is still cheap. Four months before the rent review, the agent reads the signals the resident is already sending: NPS down two points, a maintenance ticket unresolved for nine days, app engagement fading, moving-out FAQs opened twice. It scores the tenancy for renewal risk, names the drivers, reprioritises the ticket to the top of the contractor's queue, and drafts the gesture, a voucher or a call, for the property manager to approve. Turnover moves from 35% to 32%, and every renewal is a re-let cost of roughly £1,800 that is never paid.
Smarter pricing: the agent prices the stay, not the postcode. Six weeks out, it builds the rent review from quality-scored comparables, achieved rents rather than asking, live demand, and everything it knows about the resident: payment record, affordability, sentiment. It returns a recommendation with the reasoning attached, the market rate, the suggested range, the confidence score and the explicit trade-off between a full uplift and the renewal probability that comes with it. Within the approved band, the workflow generates the Section 13 notice, polices the frequency cap and notice period, and files the written evidence pack. Outside the band, it waits for the asset manager. The arithmetic the agent is running: a fortnight of void on a £2,000 rent costs about £925; a £50 squeeze earns about £600 a year. Homes priced right first time let in 21 median days; homes that needed a cut take 78. The result is +0.75% rent captured and a notice that survives the tribunal.
Voids: the notice period becomes a marketing head-start. When a resident serves notice, the agent offers an early-notice incentive, a free end-of-tenancy clean or a week's rent for viewing access and a referral, through the resident app, and tracks what each pound returned in void days avoided. It lists the home immediately, pre-qualifies enquiries against the operator's rules, schedules viewings, and runs the application. 46% of offers agreed before the resident vacates becomes 63%; 48 median void days per re-let becomes 15; roughly 3% of NOI stops leaking.
Arrears: a conversation on day three, not a write-off on day thirty. The agent flags the missed payment the day it happens, sends the right nudge through the right channel, offers the three in-app payment routes, and escalates to a human only when the pattern, not the single event, says it should. Bad debt moves from 1.2% towards 0.8%.
Scalability: the agent prioritises the day. For every property manager, the agent orders the morning around what the asset strategy needs: renewals at risk, voids about to breach their SLA, and arrears cases that have crossed the threshold. SOPs live in the workflow, so new team members are productive in days. Admin hours per lease fall from 12 to 5, back-office hours drop about 37%, and units per head rise 30%, the number that decides whether growth improves margin or eats it.
None of these is a separate AI product. They are the same workflows the platform runs today, with judgement delegated at points where it was previously late, manual, or absent.
Delegating judgement to an agent does not remove people from the loop. It moves them to the points where their judgement matters most, and it puts every stakeholder in front of the same live record.
This is hub-and-spoke control made operational. Asset strategy is set centrally, executed locally by the team or third-party agent through the same workflows, with an agent doing the work in between and the data staying with the asset, whoever runs it.
The case for agents is not that any one of these decisions is dramatic. It is that an agent makes all of them, on every tenancy, every day, and modest improvements on every lever stacked are worth more than any single initiative. Run on separate tools, each lever is a project. Run on one system, each strengthens the next: the fixed ticket lifts the renewal score, the renewal avoids the void, the avoided void protects the achieved rent that prices the next review.
On an illustrative portfolio of 1,000 homes at £1,550 average rent and 35% turnover, valued at an unchanged 4.5% cap rate, an owner-operator taking control of the levers adds £759k of NOI in year one, +5.5% on a £13.7m base, which is £16.9m of asset value, or £16,874 a home. By year five, that is £1.2m of NOI a year and £26.8m of value, with £4.9m of cumulative NOI along the way. Hit only half of every target, and it is still +2.8% NOI, £380k a year. An owner bringing leasing in-house on the same system adds £956k in year one and £21.2m of value.
The agent does not create those numbers. The levers do. The agent changes the probability that the levers are pulled on every tenancy, not just the ones a stretched team gets to, and the speed at which they are pulled, which, in the tenancy loop, is the whole game: the earlier you act, the cheaper the fix.
Everything above depends on a property most operators don’t have and cannot buy as a feature: one connected system for everything operational, with every stakeholder in it, every workflow predefined, every record standardised, and the rules of the asset strategy encoded as guardrails. Marketing, leasing and community on one platform; Insights+ as the intelligence layer across it; Revenue Management and Yield Control as the disciplines the agent executes; the PMS and accounting backend integrated two-way, in real time, rather than replaced.
An agent on a patchwork stack inherits every leaking bucket, every undocumented workflow and every untrusted record, and then acts on them faster. An agent on a rental operating system inherits a golden record, a defined process and a human-set threshold for every decision it takes. That is not a better model. It is a different foundation, and it is the one we have been building for years, which is why we have been content to be a fast follower on the models themselves. Apple did not build the first MP3 player or the first smartphone. It built the best ones, on the right foundation, at the right moment.
The operators who will earn meaningful income from capabilities that did not exist in 2025 are the ones whose stack can carry them. Fix the chassis, and the engine has somewhere to go.
