operating-model / decisions/0003-concierge-as-defining-feature.md
id: adr-0003type: adrstatus: accepted

ADR 0003 — Concierge as defining feature

Status

Accepted 2026-04-20. Replaces an earlier draft that framed the defining feature around "match quality" as a measurable output — a framing that conflated process (measurable in-flight) with outcome (measurable only post-signature, with time). This ADR separates them: the concierge mechanism is defining and is evaluated on process; outcome quality is held as a future instrumentation target.

The accepted version carries the operational content tightened through two rounds of internal review: honest brokerage collapsed into a single enforceable rule keyed on justification against the renter's current brief (with a silence clause and an allowed-extension clause, the latter constrained by preserve-and-declare); an explicit concierge→CAC chain; the perfect-fit alert framed as a falsifiable differentiation test (two-tier — reply-rate primary, positive-intent-signal secondary once sample permits); explicit escalation triggers so escalation judgement is falsifiable; and explicit failure tripwires in place of passive "drift" language. The term "sustained", which appears in the tripwire language below, is defined once in renter-loop/metrics.md · Measurement conventions and applies uniformly across this ADR.

Context

The UK BTR marketplace space is crowded with portal-style products: large inventories, good search, commoditised user experience. Competing on breadth is expensive (CAC becomes a race to the bottom) and undifferentiated.

Rentiful is a closed-loop system that converges on the right unit for each renter by tightening the profile through every interaction. This is the system identity, and the filter that every feature, hire, and investor conversation should be checked against.

Rentiful's product thesis is that renters benefit more from the right unit surfaced proactively than from a larger list to filter through. The concierge is a progressive-disclosure mechanism: it surfaces options, captures the renter's reaction, tightens the profile, and converges on the right unit faster than unaided search would. Operationally, that means Rentiful runs a concierge process for each renter — learning their needs progressively, monitoring the market in near-real-time, surfacing matches with reasoning as they appear, and capturing feedback on every option surfaced so the profile tightens with each turn of the loop.

Without a structural commitment to this, the business drifts. Under commercial pressure the easiest thing to do is widen the top of the funnel and send more digests; the product quietly converges on a portal.

Decision

Concierge is the defining feature of the Renter Loop. The concierge process has four components that must all operate together:

  • Learn. Build and maintain a profile of what the renter actually wants, progressively, through the conversation rather than in one upfront form. The profile is alive, not one-shot.
  • Monitor. Watch the market — new units, changes, price moves — in near-real-time against active profiles.
  • Share. When a match appears, surface it to the renter in the conversation with specific reasoning for why this unit fits this renter.
  • Feedback. Every option surfaced produces a signal — too small, too expensive, wrong area, right direction — that updates the profile. The loop closes itself.

This has specific structural implications.

Purpose statement

The loop's closed state includes a unit the renter genuinely wanted, not just a signed lease. A signed lease on a unit the renter settled for is a different outcome and is not the goal.

Operation shape

Profile-building is continuous, not one-off. Matching produces explicit fit reasoning. The perfect-fit alert operation is a distinct operation alongside the recurring digest, not absorbed into it.

Process quality, not output quality

The concierge is evaluated on how it operates — six observable dimensions: responsiveness, profile-updating, honest brokerage, specific reasoning, escalation judgement, and feedback-loop capture. Each dimension carries an enforceable operational threshold. The form of each threshold is fixed here; the numeric values live in renter-loop/metrics.md and are tunable based on baseline data without amending this ADR.

The conversation health operation is the review mechanism that samples conversations against these thresholds.

Honest brokerage as architectural constraint

Honest brokerage is the dimension most at risk under commercial pressure and under automation. It cannot survive as a value statement; it must be encoded as a system constraint. The concierge is required to hold one rule:

Every surfaced option must carry an inspectable justification against the renter's current brief. The brief is what Rentiful has recorded about the renter — at minimum location, bedrooms, budget, and timescale, plus whatever the conversation has surfaced (travel-time constraints, walking-distance preferences, furnishing, pets, and so on). Two operational clauses follow:

  • Silence when nothing is justifiable. If no option can be justified against the brief, the concierge stays silent; optionally a nurture touch that does not claim to be a match. The system must not push synthetic or stretch stock to hit a cadence target. Silence on no-brief-match days is correct concierge behaviour, not a failure mode, and must not be overridden by engagement or cadence targets. A sustained absence of silence on days when no new inventory meets the brief is evidence of over-promising, not of working.
  • Justified brief-extension is allowed; brief-violation is not. The concierge may introduce a dimension that unlocks an option, but any extension must (a) preserve at least one of the brief's core dimensions — typically location or budget, but the set is determined by the brief — and (b) explicitly declare the dimension being flexed in the surfaced reasoning. Example: ARK co-living at £1,599 all-inclusive against a Canary Wharf studio brief capped at £2,000 preserves location and budget and declares the format flex (studio → co-living studio). Counter-example: a studio in Glasgow against a Canary Wharf brief preserves no core dimension — the location is abandoned with no substitute and the flex cannot be declared defensibly; cheaper stock is not on its own a justification.

A match without a stated why this unit for this renter is a portal recommendation. The concierge must be able to justify itself to the renter and to the audit.

Commercial linkage — why concierge is also the CAC story

The concierge is not separable from the commercial engine. The progressive-disclosure mechanism is what makes the north-star (signed leases at CAC < £350) reachable sustainably:

  • Progressive profile-tightening produces higher-fit matches over time.
  • Higher-fit matches convert better at every later funnel stage (enquiry → conversation, conversation → viewing, viewing → signed).
  • Higher stage-to-stage conversion means lower spend per signed lease at any given channel mix.
  • The £350 guardrail is therefore reachable by the concierge mechanism working, not by cheaper spend or by scaling low-conversion channels.

The inverse is load-bearing: if CAC trends up at constant spend, the concierge process is degrading. CAC drift is a direct test of the concierge thesis, not an independent commercial problem. This is what makes the defining-feature claim and the commercial north-star the same thing rather than two separate claims.

Perfect-fit alert as the falsifiable differentiation claim

The perfect-fit alert operation is the single clearest expression of the concierge thesis. A recurring digest is what a portal does well enough. An alert that says "a unit came on the market in the last hour that fits your brief because of X, Y, and Z" is what a portal cannot do.

This makes the alert the cleanest falsifiable test of this ADR, expressed in two tiers:

  • Primary (reply rate). The reply-within-24h rate on perfect-fit alerts must exceed the equivalent rate on recurring digests by a margin defined in renter-loop/metrics.md. Reply is a weak proxy, but it is the signal with enough sample to be steerable from day one.
  • Secondary (positive intent). Once sample permits, the alert must also out-convert the digest on positive-intent signals — viewing requests, availability enquiries, specific questions about a surfaced unit. Reply rate alone can be gamed by volume or provocation; positive-intent-signal rate is the honest check on whether the alert is delivering value, not just attention. Threshold and volume gate live in metrics.md.

If the alert fails the primary test and the secondary test as they come online, the concierge differentiation is empirically absent and this ADR is wrong. Both tiers are first-class measured events, instrumented from day one — crudely if necessary (manually tagged in the spreadsheet), properly as soon as platform work allows.

Outcome quality held for later

Whether a match turned out to be the right one — measured through retention, post-move satisfaction, and referral — is a future instrumentation target, not a current metric. When sufficient post-signature signal exists (likely ≥12 months of residency data at meaningful sample), outcome quality can be added to the Renter Loop metrics. Until then, no claim is made to measure it.

Escalation triggers

Escalation judgement is listed above as a process-quality dimension; without explicit triggers, that dimension is unfalsifiable and the system can loop indefinitely while appearing healthy. The concierge must escalate — from AI to human, or across channel (e.g. email → WhatsApp) — under any of the following conditions:

  • Repeated negative signals on surfaced options. A renter who rejects successive options without the profile tightening is not being served by the loop and must be escalated. Numeric threshold (count and window) lives in metrics.md.
  • Expressed frustration or ambiguity in renter language. Detected frustration, confusion, or a contradictory signal (e.g. budget contracts and expands in the same thread) triggers a human review of the conversation before the next automated touch.
  • Stuck-conversation firing. When renter-loop.engagement.stuck-conversation identifies a conversation as stuck by its own criteria, escalation is automatic, not discretionary.

Escalations are a first-class measured event in the conversation health operation; correct timing is one of the criteria used to score the escalation-judgement dimension. A system that never escalates is failing; a system that escalates everything is also failing, but the former is the more dangerous drift.

Platform requirement

Near-real-time supply visibility is a structural prerequisite (see platform.unit-index). A concierge operation that surfaces yesterday's units is already a broken concierge operation.

Automation discipline

As the Renter Loop scales from 1/week toward 25/week, the process-quality dimensions must not degrade under automation. Responsiveness typically improves with automation. Honest brokerage is the dimension most at risk, as templated responses default toward optimism — the architectural constraints above exist specifically to protect it under automation pressure. Escalation judgement is the dimension the AI must get right, or the whole thesis fails. Each automation step is evaluated against the failure tripwires below before it ships.

Consequences

The north-star metric for the Renter Loop is signed leases at CAC < £350. Concierge process quality is the leading indicator, sampled from real conversations. Outcome quality is a future instrumentation target and is not claimed today. Diagnostic feedback work (Domain D) centres on concierge-process diagnosis; a separate match-quality-analysis operation is not required and would introduce an unmeasurable output claim. The perfect-fit alert operation creates a hard cross-loop dependency: the Operator Loop must surface new units to the Renter Loop in near-real-time.

Failure tripwires

The defining-feature claim is empirically false under any of the following sustained conditions. "Sustained" is defined in renter-loop/metrics.md · Measurement conventions. Each tripwire is paired with the dimension it tests; each tripwire's numeric threshold, tolerance band, and window length live in metrics.md and are reviewed quarterly.

  • Response latency on escalated conversations exceeds threshold. The concierge mechanism is failing its responsiveness promise. Tests: responsiveness.
  • Share of AI-generated responses containing no renter-specific reasoning exceeds threshold. Honest brokerage has degraded into template generation. Tests: honest brokerage, specific reasoning.
  • Fraction of surfaced options without an inspectable justification against the renter's current brief exceeds threshold, OR silence-rate on no-brief-match days falls below its floor. Either is evidence the concierge has drifted into portal-style recommendations. Tests: honest brokerage.
  • Perfect-fit alert reply rate / digest reply rate falls below threshold ratio, AND (once sample permits) alert positive-intent-signal rate / digest positive-intent-signal rate falls below its threshold ratio. The concierge differentiation is empirically absent; the defining-feature claim is wrong. Tests: the ADR itself.
  • Rolling 90-day CAC per channel × bedroom trends up at constant or falling spend. The concierge process is degrading, even if individual process dimensions look healthy. Tests: the commercial linkage.

If any tripwire fires, the response is not optimisation — it is investigation. These are tests that the ADR is still operationally true, not KPIs to be managed.

Alternatives considered

  • Concierge as a nice-to-have. Rejected: under commercial pressure, it would be the first thing cut, and the product would converge on a portal.
  • Concierge as a premium tier. Rejected for now: splitting the renter base into concierge and non-concierge complicates every operation and dilutes the brand promise. May be revisited if a clear commercial case emerges.
  • Match quality as a primary metric (an earlier draft's framing). Rejected. "Match quality" conflates process and outcome and implies a measurable score that Rentiful cannot produce today without sustained post-move signal. The concierge is instead evaluated on mechanism (process quality, observable in-flight), with outcome quality held as a future instrumentation target to be added when the signal exists.

Related

Comment on GitHub →