Tenant Retention Probability Scorer

A tenant retention probability scorer ranks which occupiers are most likely to renew so an asset manager can spend scarce relationship time where it changes the outcome. The model outputs a probability per tenancy, not a notice, a heads of terms draft, or an instruction to the tenant. Asset managers still choose who to call, what to offer, and when to escalate.

The score is useful when a portfolio has more upcoming expiries than the team can work with equal depth. A high probability tenant may need a light-touch confirmation and a clean paper trail. A mid-range score often warrants a structured conversation about space, rent, and remaining term. A low score flags a real risk of vacancy, but it does not by itself justify a concession or a legal step.

The unit of scoring is a live tenancy on a named asset, not a company brand or a building-level occupancy rate. For each eligible lease, the model returns a renewal probability for the current term's remaining window, plus the main drivers that moved the score. Drivers should be readable by a property professional: late option dates, falling desk or suite utilization, unpaid or waived covenant items, complaint clusters, or a pattern of short extensions rather than a full term.

The score is a ranking aid. Two tenants at 0.62 and 0.64 are not meaningfully different for a weekly call list. Banding (for example high, watch, and at-risk) is enough for triage, provided the bands are defined in operating procedure and not treated as a credit rating. When two scores sit in the same band, the asset manager still breaks the tie with rent roll materiality, vacancy cost on that suite, and relationship knowledge the model does not have.

The scorer does not send email, book meetings, or write commercial terms. Those steps sit with staff and with neighbouring workflows such as Automated Renewal Heads of Terms and Renewal Pipeline Workflow Agent. Keeping probability separate from action prevents a noisy feature from flooding tenants with premature outreach.

Lease-event timing is a required input. That includes current expiry, break dates, option notice windows, rent review dates, and the dates of prior renewals, holds over, and short-term extensions. Without a reliable event calendar, "soon" and "safe" are guesses. The model also needs a current occupancy identity: which legal entity occupies which demised premises, and whether the lease is still the live instrument.

Space utilization is the other required input. Typical sources are access-control counts, badge or visitor logs mapped to the demise, sensor or workplace-system occupancy where the landlord already has a lawful feed, and vacancy of contiguous space the occupier historically used. Utilization does not have to be a single occupancy percentage. A collapse in after-hours use, a floor that is consistently dark while rent is paid, or a request history that shrinks rather than grows can all be valid signals if they are dated and attributable.

Churn signals are supporting, not a substitute for the two required families. Useful examples include a spike in assignment or subletting enquiries, broker introductions the landlord can see, service-charge disputes that stall, repeated delay on fit-out or reinstatement, and a pattern of "we will confirm next quarter" with no diary follow-through. These features help when lease dates and utilization are present. They should not be used to invent a score when the core inputs are missing.

Covenant health is likewise supporting. It covers rent and service-charge payment behaviour, agreed payment plans, parent or guarantor changes, and known insolvency or restructuring flags already on the credit file the asset team uses. Covenant stress can raise the chance that a tenant wants to stay but cannot, which is a different retention problem from a tenant that is already walking. The score should expose that distinction in drivers so the manager does not spend goodwill budget on a credit issue that belongs to another process.

Data quality rules belong in the same operating note as the model. Duplicate lease IDs, demises that do not match the rent roll, and utilization feeds that cover the building but not the suite should fail closed. A score built on the wrong floor is worse than no score.

If lease-event inputs are missing, incomplete, or internally inconsistent (for example two different expiries, or a break date with no associated notice mechanics), the model returns empty output for that tenancy. If space utilization inputs are missing for the demise, the model likewise returns empty output. Partial building-level occupancy, a last-known figure older than the team's freshness policy, or a feed that cannot be mapped to the lease should be treated as missing.

Empty output is a first-class result. It should appear on the worklist as "unscored" so the team knows the gap is data, not a quiet high-probability tenant. Operations then either repairs the lease diary and utilization join, or the asset manager works that file on a manual checklist. The model must not impute a default "average" probability to keep dashboards looking complete.

Unscored files still need a human owner. A missing score often coincides with the messiest assets: recent acquisitions, incomplete handover packs, or occupiers on licences and side letters that never entered the event engine. Those cases are exactly where an automatic rank would be most misleading.

The practical cadence is a rolling look-ahead across the portfolio's notice and expiry windows, not a once-a-year campaign. Each cycle, the asset manager (or the AM lead for a mandate) reviews scored tenancies in the relevant window, drops unscored rows into a data queue, and builds a short outreach list. Priority usually combines probability with economic weight: passing rent, unrecovered void cost, and whether the suite is hard to re-let.

High-probability names still get a plan, but a thinner one: confirm intent, lock the diary, and avoid last-minute surprises. Watch-band names get the scarce hours: a visit or a structured call, a check on real space need, and an early view of what would make a stay rational. At-risk names need a dual track: retention conversation where it is still plausible, and a parallel leasing plan so the asset is not betting the budget on a score that already says leave.

The list is a weekly (or fortnightly) artefact, not a permanent ranking. Utilization and covenant features move. A tenant that looked stable can empty a floor after a merger. Re-scoring on a defined schedule, and after material lease events, keeps the queue honest. Staff should be able to pin a comment ("CFO told us they are consolidating") that does not overwrite the model but does explain why this week's calls ignore a high score.

The model scores probability. Staff still run outreach. Before a tenant is contacted on the back of a score, a named asset manager (or delegated property manager with AM oversight) reviews drivers, rent-roll facts, and any relationship notes. They confirm the legal entity, the right relationship owner, and that no litigation, insolvency, or confidentiality constraint makes a "retention" call inappropriate.

Outreach content is a human decision. A low score is not a script for panic discounts. A high score is not permission to ignore the file until the notice date. Offers, options, and heads of terms remain commercial and legal work, often handed to Automated Renewal Heads of Terms only after the manager has chosen a direction. The scorer's job ends when the ranked list and drivers are current.

Auditability matters when several people share a mandate. Record who accepted or overrode a band, on which date, and whether the tenant was contacted. Overrides are expected. The model cannot see a confidential sale of the occupier's business or a handshake in a site meeting. What it can do is stop the team from spending equal effort on every expiry while the files that will actually churn sit untouched.

Related reading: Lease Break Probability Model for break-window risk on the same leases, and Renewal Pipeline Workflow Agent for routing approved files through tasks after the score has been reviewed.