Eight findings from 49,320 council-stage planning decisions across Dublin, analysed at factor level using LLM extraction from planner reports. The first dataset of its kind in Ireland.
We built three machine-learning models on 18,066 Dublin adjudications. One uses only the 50 most-cited policy references. One uses only the 76 planning factors. The third combines both.
Once you know what the planner actually weighed up, the specific policy clauses they pointed to add nothing. The real hierarchy is a handful of concrete concerns.
The factors that do the heavy lifting are: privacy (raised as a negative in 7,365 decisions), visual impact (6,038), outlook (4,605), architectural quality (4,870), overlooking (4,026), massing (3,242), height (3,217), and overdevelopment (2,494). Three themes dominate: residential amenity, design quality, and scale. Everything else — transport, heritage, drainage, ecology — matters at the margins but rarely drives the outcome.
Planning decisions are supposed to involve balancing — weighing positives against negatives. We tested whether the system actually balances on 33,662 Dublin council decisions.
Each extracted factor has a polarity (supports grant or supports refusal) and a strength (1–3). A single strong negative drops grant probability from 95 per cent to 31 per cent. The number of positives barely matters.
With 1 positive and 0 negatives, the grant rate is 97 per cent. Add 5 negatives and it drops to 51 per cent — regardless of how many positives you stack up. The system is closer to veto logic than balancing.
DCC, DLR, Fingal and SDCC all operate under the same national legislation. We fitted identical models for each council — same 15 factors predicting outcomes — and compared the results.
The rank correlation of factor weights averages 0.55 across the six council pairs. They broadly agree on which concerns matter. But DCC grants 86 per cent of applications and SDCC grants 60 per cent. The intercept range in the logistic regression is 2.2 — a massive gap in where each council sets the refusal trigger.
| Factor | DCC | DLR | Fingal | SDCC |
|---|---|---|---|---|
| Privacy | 58% | 34% | 39% | 30% |
| Visual impact | 51% | 26% | 29% | 43% |
| Architectural quality | 49% | 25% | 27% | 48% |
| Height | 52% | 28% | 33% | 47% |
| Overdevelopment | 33% | 18% | 8% | 13% |
| Precedent | 31% | 21% | 13% | 11% |
When a DCC planner raises a privacy concern, the application still has a 58 per cent chance. In DLR: 34 per cent. Overdevelopment in Fingal is near-lethal at 8 per cent; in DCC, 33 per cent.
If planning outcomes were driven by place — character, infrastructure, zoning — then grant rates should be similar on both sides of a council border. The geography is identical. Only the decision-maker changes.
The DCC/DLR border is the starkest. DCC side: 95 per cent. DLR side: 62 per cent. Same neighbourhood. Average boundary gap: 21 percentage points.
Planning agents are repeat players. They learn the unwritten rules, know the planners, and understand which concerns to pre-empt. We tested this across 1,619 agents with 20 or more Dublin applications, controlling for development type, council and year.
The bottom-decile agent gets 73 per cent of applications through. The top-decile gets 96 per cent. That’s a 23-percentage-point spread that persists after controlling for everything observable about the case. The knowledge of how to navigate the system is real, privately held, and worth a lot of money.
Do individual planners make different calls on similar cases? The raw data suggests yes. Among 35 planners with 30 or more decisions in our dataset (concentrated in DLR and SDCC, where planner names are reliably extracted), the grant rate ranges from 75 per cent to 97 per cent. A 22-percentage-point spread.
| Council | Planners | Mean rate | Std dev | Min | Max |
|---|---|---|---|---|---|
| DLR | 13 | 82.3% | 4.2pp | 74.6% | 88.9% |
| SDCC | 25 | 89.4% | 6.1pp | 74.7% | 97.3% |
But the raw spread overstates genuine planner discretion. Planners don’t get random cases. Some handle more extensions, others get more apartments. Some cover areas with more contentious zoning. When we control for development type, geography, year, and the applicant’s agent, planner coefficients shrink by 27 per cent. The average absolute planner effect drops from 0.68 to 0.50 in the logistic regression.
That still leaves meaningful residual variation. A planner one standard deviation above average in DLR (4.2 percentage points stricter than the council mean) will refuse cases that an average DLR planner would grant. Over a year’s caseload of 50–100 applications, that’s several extra refusals traceable to the individual, not the case.
In SDCC, where the standard deviation is wider (6.1pp), the planner effect is larger. The most permissive SDCC planner grants at 97 per cent; the strictest at 75 per cent. Even after controls shrink this gap, a chunk of it reflects genuine differences in how planners read “residential amenity” or “overdevelopment” — the same subjective factors that dominate the system overall.
When a council refuses an application, it cites specific development plan policies. But not all those policies carry weight on appeal. An Bord Plean√°la overturns council refusals 37 per cent of the time in Dublin. Some policies fare far worse.
| Policy clause | Cited | Overturned |
|---|---|---|
| Section 8.2.3.4 — Additional Accommodation | 12 | 83% |
| Section 8.2.3 — Residential Development | 13 | 69% |
| Section 8.2 — Development Management | 16 | 69% |
| DLR Dev Plan 2016-22 | 11 | 64% |
| Project Ireland 2040 NPF | 17 | 59% |
| Zoning Objective A | 30 | 53% |
Section 8.2.3.4 of the Dublin City Development Plan (Additional Accommodation in Existing Built-up Areas) gets overturned 83 per cent of the time. Councils cite it in 12 refusals; ABP reverses 10. Each reversal costs the applicant 7–8 months and thousands in appeal fees.
Nobody applies for a 10-storey building in Rathmines. Not because nobody wants to, but because everyone knows what would happen. The system’s heaviest hand isn’t in its refusals — it’s in the proposals that never materialise.
1,939 of 3,430 proposals are for a single dwelling. The drop to 2 units (343 proposals) is 82 per cent. By 10 units, 29 proposals. The refusal rate barely changes — the system doesn’t punish larger schemes more. Applicants have already self-selected out.
Appeals cost money, take time, and require knowledge of the system. We tested whether contestation is socially patterned by matching 76,755 Dublin applications with their deprivation score (Pobal HP Index) and measuring appeal rates across the spectrum.
The most deprived decile has an appeal rate of 1.4 per cent. The most affluent areas appeal at 3.1 per cent — more than twice as often. Third-party appeals (objections against grants) follow the same pattern: 1.5 per cent in affluent areas, 0.8 per cent in deprived areas.
The capacity to contest planning decisions is unevenly distributed. Affluent areas are better equipped to challenge outcomes they dislike — which helps explain why those areas stay low-density while areas with fewer resources to object absorb more change.