Most development teams do not lose time on analysis. They lose it on retrieval. A land team in Dublin or Manchester spends the bulk of a site appraisal cycle finding documents, not reading them: pulling the development plan, checking the zoning objective, tracing whether a site sits inside a Strategic Development Zone, confirming flood classification, working out who owns the adjoining strip, and reconciling three different sources that disagree about site area.
That is the part agents are genuinely good at. Not judgement. Retrieval, reconciliation, and the first pass of arithmetic.
Where the fourteen weeks go
The figure below models a mid-sized developer running a residential or mixed-use scheme through appraisal. The stage durations come from a composite of how these teams typically sequence work: an initial long list, a manual planning and constraints review, a preliminary financial appraisal, and then a partner review that sends roughly a third of sites back for rework.
The last row matters more than the others. Partner review does not compress much, and it should not. What changes is that the review happens against six well-evidenced sites rather than sixty thinly-evidenced ones, so the rework loop mostly disappears.
What the agent actually does
A site selection agent is not a chatbot with a property database bolted on. It is a sequence of tool calls with a verification step after each one, and the verification step is where most implementations fail.
- 1
Parse the brief
Use class, target unit count, budget envelope, minimum site area, acceptable travel time to transit.
- 2
Assemble candidates
Cross-reference land registry, agent listings and off-market records into a deduplicated candidate set.
- 3
Pull the planning position
Development plan zoning objective, planning history on the folio, live applications on adjoining land.
- 4
Screen constraints
Flood classification, protected structures, archaeological zones, tree preservation, services capacity.
- 5
Run the appraisal
Residual land value against build cost assumptions, sales rates and a finance cost curve.
- 6
Rank and evidence
Ordered shortlist where every figure carries a link back to the document it came from.
The last step is the commercial one. An agent that produces a ranked list is interesting. An agent that produces a ranked list where a director can click any number and land on the page of the development plan it came from is something a firm will actually put money behind. Evidence links are what move an output from advisory to usable.
The Irish and UK specifics that break naive automation
Generic property AI tends to be built against United States data conventions, and it degrades quickly here. Four things cause most of the failures.
- Development plans are PDFs with maps, not APIs. Zoning objectives sit in a written statement, the boundaries sit in a separate map layer, and the two are frequently ambiguous at the edges. An agent has to read both and flag its own uncertainty rather than guessing.
- Folio and title data does not align cleanly with mapped site boundaries. A candidate site is often several folios with different registered owners, and the arithmetic on developable area changes materially depending on how that resolves.
- Planning history is the strongest single predictor of consent risk, and it is the least structured data in the process. A refused application on an adjoining site three years ago tells you more than any generic model score.
- Part L and the wider Building Regulations changed the cost floor. Any appraisal running pre-2022 build cost assumptions is producing a residual land value that will not survive contact with a quantity surveyor.
The value is not that the agent is clever. It is that it never gets bored on the fortieth site.
What this is worth across a pipeline
Cycle time compression only pays if it changes a commercial outcome. In development it does, through three routes: more sites screened for the same team cost, faster movement on off-market opportunities, and fewer schemes carried into expensive due diligence before a fatal constraint is found.
Where to start
The instinct is to build the whole pipeline. That is usually the wrong first move, because the constraint screening step is the one that carries the most risk and the least tolerance for error. Start instead with planning position retrieval on sites the team has already appraised. You get a clean accuracy benchmark against known answers, the team builds trust in the evidence links, and you find out early how badly the source documents are going to fight you.
Once retrieval is trusted, the appraisal arithmetic is comparatively easy. Most firms have the model already. It just runs on a spreadsheet that one person maintains.