The Intelligent Letting Agency
The AI-native firm thesis applied to residential lettings — for a colleague or an investor. This sets out the vision and the ceiling. The conservative deal underwriting — the floor we would actually pay against — lives separately in the Glenn Flegg deal memo and the acquisition model, and is deliberately left untouched here.
1The thesis, in a paragraph
A letting agency is, stripped back, a knowledge-work factory wrapped around a handful of physical visits and one regulated promise: that someone competent and accountable is looking after the landlord’s asset and the tenant’s home. Almost all of its cost is brain-work — rent reconciliation, statements, compliance tracking, document drafting, arrears chasing, referencing, communications. Most of that work does not need to happen in any particular place, and increasingly does not need to be done by a person at all. Project Barbara’s thesis is to acquire established lettings books and re-found them as Intelligent Letting Agencies: businesses with a single AI-managed data spine at the centre, where the machine executes the knowledge work and humans are kept only where humans are irreducible — physical presence, accountability, and judgement. The result is a services business that runs at software margins and compounds as it consolidates a fragmented market.
2Why lettings — and why now
It is not a coincidence that lettings scores a near clean sweep against our framework for where AI-native operating firms win first (the Nine Conditions):
| Condition | Lettings |
|---|---|
| 1 · Heavy knowledge-work cost | ✅ ~60–70% of cost is brain-work — property managers, accounts, admin, plus regulated suppliers (solicitors, accountants). |
| 2 · Fragmented, no escape velocity | ✅ ~25,000 UK agencies, 92%+ independent; no incumbent with the scale or talent to transform first. |
| 3 · Fast revenue acquisition | ✅ Buy a book, inherit sticky recurring relationships, take cost out. |
| 4 · Fast capital cycles | ✅ Monthly recurring fee income; weeks-to-months feedback, not years. |
| 5 · Zero-marginal-cost adjacencies | ✅ The schema already knows the property, landlord and tenancy — insurance, maintenance, rent reviews flow at near-zero cost. |
| 6 · Structurally slow competitors | ✅ Run on email, spreadsheets and institutional memory; no CTO, let alone frontier AI. |
| 7 · Compounding process intelligence | ✅ Every automation refined and edge case handled is a moat a competitor must re-earn from years behind. |
| 8 · Multi-domain, high-consequence need | ✅ Regulation + money handling + legal obligations + maintenance, with catastrophic downside if done wrong — so the customer still pays someone to own the risk. |
| 9 · Undervalued physical/legal asset | ✅ This is the Barbara thesis — a recurring lettings book priced at a service multiple, as if AI did not exist. |
Why now: frontier models that clear the bar on most lettings knowledge-work; a fragmented market of retiring owners; and the Renters’ Rights Act resetting every process at once — a once-in-a-generation moment when rebuilding the workflow from scratch is not a disadvantage but the whole point.
3What an Intelligent Letting Agency is
- A central brain. One unified, AI-managed database is the cornerstone asset — the Labs pattern. Veco, the portals, open banking, the deposit schemes and council registers become thin connectors that ingest into our schema; they stop being systems of record. Kill the file: a file in a folder is data the AI cannot see.
- The hopper / centurion operating model. The AI processes first, with maximum context; a human governs a review queue — approve, edit or reject in seconds; corrections feed back and the system improves every cycle. People guard the integrity of the system that does the work; they do not produce the work.
- AI-first decision routing, with human governance reserved for the high-stakes points only.
- Supplier subsumption. The external bookkeeper, the compliance consultant, much of the solicitor — absorbed into the brain. As the regulated principal we carry the accountability and capture the margin.
- Zero-marginal-cost adjacencies. Insurance referral, maintenance margin, rent-review service, ancillary income — computational, not new hires.
- The proof metric is RPE (revenue per employee), not headcount or branches.
4What’s possible — and how
Every process sorts on three axes. Where the work must physically be: physical visit / local knowledge / anywhere knowledge. Who owns the outcome: A — the AI owns it (verifiable, low-stakes, fast to ground truth), or B — a human owns it (slow or contested truth, or unrecoverable stakes) with AI as a tool — the Verifiable Frontier: who carries the risk of the AI being wrong? And — the third axis, added once we accept a competent agency already runs a modern CRM (Alto-class) — what we actually build: is the core delivered by the CRM (a hygiene factor we buy), an AI outcome (we build the automation), an AI copilot (AI assembles, a human decides), or a physical body; and what net-new AI adds on top. (All 63 processes are classified this way in the process catalogue.)
| A — AI delivers the outcome | B — AI is a tool; a human is accountable | |
|---|---|---|
| 3 · Knowledge work, ANYWHERE (most of today’s salaried cost) | Rent reconciliation · statements, payments, tax packs · client-money recon · deposit protection + prescribed info · tenancy drafting · arrears reminders & letters · compliance tracking · listing copy + portal publish · referencing parse · notice drafting (S13/S8) · AML · legislative monitoring | Possession strategy & hearings · deposit-dispute & disrepair narratives · complaints / ombudsman · vulnerable-tenant arrears judgement · landlord advice · acquisition valuation |
| 2 · Knowledge work, LOCAL (the operator’s patch) | Comparables & data that feed the appraisal | Rent-level judgement · landlord relationships & trust · contractor network & triage · council-scheme nuance · fee negotiation |
| 1 · PHYSICAL visit (local presence — AI coordinates and arms, never performs) | — (AI schedules, routes, chases access, parses the report, and arms the visit) | Viewings · check-in/out & inventories · inspections · emergency make-safe · gas/EICR/EPC certs · keys & photos |
The third axis re-bases the thesis honestly. Measured against a well-run modern CRM — not the email-and-spreadsheets caricature — the operational work splits almost exactly in thirds:
- ~⅓ the CRM already does (Base, ~736 hrs/100 props) — statements, payments, rent matching, reconciliation, VAT, portal publish. Use the tool properly; don’t build. Tier 2 — and the adoption gap is real (agencies own the CRM yet pay people to do its job by hand).
- ~⅓ AI delivers (~659 hrs) — the cognitive work the CRM only logs: maintenance triage, tenant comms, drafting, arrears, rent-review, referencing decisioning. This is what we build — Tier 3.
- ~⅓ physical (~653 hrs) — armed, not untouched. The body is irreducible, but AI makes it 2–3× more effective: the contractor arrives knowing the exact part on the exact appliance (the digital twin) on a batched, keyed run; the valuer turns up with an AI-built pack of comparables and a defensible recommendation — a cyber-enhanced expert. Tier 3 on the floor we used to write off.
- a sliver of AI copilot (~61 hrs) — disputes, possession: AI assembles the file, a human decides.
So the edge over a competent CRM agency is not “automation vs none” — the CRM increasingly automates the back office — it is the AI-build third + the armed physical third + the owned spine, concentrated in maintenance, comms, drafting, rent-review and disputes, and in the cross-process intelligence (the twin, the proxy model) a rented, siloed CRM structurally cannot hold.
The engine — and where the edge is. Verifiability is really a question of τ, the time to ground truth. You disaggregate each process and engineer it toward τ ≈ 0: turn regulation into a procedural oracle, design self-checking structure (reconcile-to-zero, registry gates), and build a private proxy model from the book’s own outcomes. The alpha is buying work the market prices as human-only and re-engineering it to τ ≈ 0 — on a spine we own. The build order matters and most people get it backwards: fix the data → live ingestion → copilot → automate. The work is the data, not the automation.
5The economics of the architecture — illustrative, NOT the underwriting
This is the end-state ceiling: what an established ~290-managed-property book becomes once it has been re-built as an Intelligent Letting Agency and the adjacencies are switched on. It is not the deal underwriting. The conservative, day-one-grounded numbers we would actually pay against live in the acquisition model — see the Glenn Flegg memo. Figures illustrative and round.
| Annual P&L — Intelligent Letting Agency (single ~290-property book) | £ |
|---|---|
| Recurring management & rent-collection fees | 450,000 |
| Transactional (set-up, let-only, renewals) | 100,000 |
| Zero-marginal-cost adjacencies (insurance, maintenance margin, rent review) | 50,000 |
| Total revenue | 600,000 |
| Operator / governor (1.0) | (85,000) |
| Physical operations (~1.5 FTE — inspections, emergencies, keys) | (55,000) |
| Client services / escalations (~0.5 FTE — the human face) | (30,000) |
| Outsourced inventories & viewings | (25,000) |
| AI & data platform — the central brain | (30,000) |
| Property portals (Rightmove / Zoopla) | (10,000) |
| Regulatory & PI insurance (PI, CMP, redress, ICO, AML) | (15,000) |
| Premises (back-office / lock-up) | (10,000) |
| Accountancy & sundries | (15,000) |
| Total operating cost | (275,000) |
| EBITDA | 325,000 (~54%) |
Revenue per employee ≈ £200k (~3 employed FTE), against a traditional agency’s ~£50k. The shape is the point: the cost base is now dominated by the human residual (an operator, a lean physical function, a human face) plus the platform and the regulated licences — not a floor of desk staff.
6This opportunity, in this architecture
Glenn Flegg is a near-textbook platform anchor: ~289 managed properties, ~85% recurring, a clean diversified book (no landlord over 10 units), a retiring owner, in a Crossrail-corridor patch. The single-book economics above are good. But the transformational economics are the roll-up: one governor and one central brain spanning several books. Physical operations and client services scale with property count; the brain, the platform and the governance layer are shared. Run ~1,000 managed properties across three or four acquired books on that one spine and RPE compounds past £300k while margin pushes through 60% — and the process intelligence accumulated along the way is the moat a fresh entrant cannot time-travel to acquire. Glenn Flegg is not the prize; it is the first room of the house.
7What’s real now, and the trajectory
Honesty about sequencing — the ceiling is built, not bought:
- Buildable now (2026). The τ ≈ 0 back office — reconciliation, statements, compliance tracking, document generation, deposit / prescribed-info, arrears drafting, referencing. These clear the gate today. On an acquired book they are a data migration and build (months off legacy Veco), not a switch you flip on completion.
- The Mythos era (end-2026 and beyond). The medium-τ, judgement-adjacent layer — fully-autonomous tenant comms, maintenance triage at scale, dispute bundling, possession workflows — moves from tool to outcome as the models improve and our private data matures. Wild levels of automation become accessible to those willing to build for them.
- Never. The residual below. Do not model it away.
The discipline: underwrite on the floor, build toward the ceiling, let the roll-up bridge the two.
8What only humans do
Three things survive, permanently, and they are why the customer cannot disintermediate us:
- Physical presence — bodies at properties. Coordinated by the brain; executed by a lean operative and an outsourced network.
- Accountability — the regulated principal the customer pays precisely so that someone else owns the risk. This is a feature, not a cost.
- Taste, judgement and trust — the operator’s relationships and strategic calls; the contested, τ = ∞ slice no oracle will ever settle.
And the human role does not merely shrink — it compounds. People in an Intelligent Business are managed at the level of intent, not tasks: given an outcome and a quality bar, they own the method, and they get better through exposure to better judgement at the checkpoints that matter. Fewer people; each one worth far more. That is the business we are building.