all research

Real Estate & Development · 8 min read

The lease portfolio that reads itself: compliance agents across commercial estates

Rent review triggers, break clauses and statutory deadlines are a data extraction problem that portfolios keep solving with calendar reminders and hope.

1,240

Leases parsed per run

94%

Clause extraction accuracy

£410k

Missed escalation exposure found

Every commercial portfolio of any size has the same quiet problem. The lease data in the management system is a summary of the lease, written by somebody who has since left, and nobody has read the underlying document since it was signed. The summary captures rent, term and break date. It does not capture the clause that says the service charge cap falls away on assignment.

That gap is where money leaks, and it leaks in both directions. Escalations that were never applied. Dilapidations positions that were never argued. Statutory obligations sitting in a schedule nobody indexed.

Extraction is the easy half

Parsing a lease into structured fields is close to a solved problem now, provided the model is given the document rather than a scan of a fax of the document. Accuracy on core commercial terms sits high enough to be operationally useful. The difficulty is not extraction. It is knowing what to do when extraction and the system of record disagree.

Extraction accuracy by clause type, measured against manually abstracted control set
percent
Term, rent, break dates98
Rent review mechanism94
Service charge provisions89
Alienation and assignment86
Repair and dilapidations79
Bespoke side letters61

Note where accuracy falls away. Side letters and negotiated variations are exactly the clauses that carry unusual commercial risk, and they are the ones a model handles worst, because they are by definition non-standard. Any credible deployment routes those to a human rather than reporting a confident answer.

The part that changes behaviour

Abstraction alone produces a database. What produces a return is the monitoring loop that runs on top of it: an agent that holds the extracted obligations, watches the calendar against them, and raises the action before the deadline rather than after.

From document to action
  1. 1

    Ingest and segment

    Layout aware parsing across leases, variations, side letters and licences.

  2. 2

    Extract obligations

    Dates, triggers, conditions and the party carrying each duty.

  3. 3

    Reconcile

    Compare against the management system and flag every disagreement rather than overwriting.

  4. 4

    Monitor

    Watch trigger dates, notice periods and statutory windows across the whole estate.

  5. 5

    Act

    Raise the review, draft the notice, open the work order, escalate what needs a decision.

Operations is the same problem wearing different clothes

The identical pattern applies to the operational side of an estate. Building management systems already produce the data. Almost nobody acts on it at the speed it arrives, because acting on it means somebody reading a dashboard at the right moment.

An operations agent watching plant data does not need to be sophisticated to be valuable. Detecting that an air handling unit has been running at a fixed damper position for nine days, cross-referencing the occupancy schedule, and raising a work order with the fault symptom already described is unglamorous work that pays continuously.

Modelled cumulative benefit across a 34 asset estate after deployment
£ thousand, cumulative
0M142M3128M6244M9391M12638M18902M24
observed range modelled projection
The compliance calendar is not a calendar problem. It is a document comprehension problem that everybody has agreed to treat as a calendar problem.

What good looks like

A deployment worth running has three properties. Every extracted obligation links back to the clause and page it came from. Disagreements with the system of record are surfaced rather than silently resolved. And low-confidence extractions are routed to a person with the specific uncertainty named, instead of being reported at the same confidence as everything else.

Portfolios that skip the third property get a system their asset managers stop trusting within a quarter, which is a more expensive outcome than never having built it.

Work with me

Run this model against your own project

I am Kanishk Kapoor, Technical Accounts Manager at AI Institute in Dublin. I build agentic AI systems with built-environment teams across Ireland and the UK. If any figure here looks wrong for your business, that is the useful conversation. Send me your assumptions and I will re-run it.

Continue reading