Property due diligence in Ireland is mostly pattern recognition under time pressure. You're given a folio, a contract, a bundle of replies to requisitions, maybe a planning history pulled from the local authority, and you need to decide what's missing, what's wrong, and what's going to cost the buyer money in three years. The work itself isn't intellectually exotic — but it's high-volume, the source documents are inconsistent, and the cost of missing something is borne by the client. This is exactly the kind of work that benefits from a private intelligence layer sitting beside the conveyancer, not replacing them.
Why property DD breaks generic AI tools
If you've tried running a contract for sale through a general-purpose chatbot, you'll know the failure modes. It hallucinates folio numbers. It confuses a Land Registry folio with a Registry of Deeds memorial. It treats a Section 72 burden as if it were a registered charge. It doesn't know what a Lis Pendens looks like in an Irish search, or why a "Note: Mapping" entry on a folio is sometimes the most important line on the page.
The reason isn't model intelligence — it's grounding. A consumer model has read a lot of US real estate forums and very little Irish conveyancing. It has no access to the Law Society's pre-contract enquiries, no access to the firm's own precedent bank, no access to the prior transactions where a similar issue came up and was resolved a particular way. Without that grounding, every answer is plausible-but-unverified, which is the worst possible state for due diligence work.
An intelligence brain is the opposite of that. It's a controlled corpus — your firm's documents, your firm's prior matters, your firm's templates, plus the structured public sources you actually rely on — wrapped in a retrieval layer that forces every answer to cite where it came from. The model becomes a reasoning engine over your evidence, not a guesser pretending to be one.
What the document set actually looks like
Walk through a typical residential purchase. You'll have:
- A contract for sale, usually the Law Society standard form with special conditions bolted on
- Copy folio and filed plan from the PRA (or a Registry of Deeds search)
- Replies to pre-contract enquiries and requisitions on title
- Planning search — grants, refusals, conditions, compliance certificates, exempted development opinions
- BER cert, NPPR and LPT history, Household Charge clearance
- Local authority searches, judgments search, bankruptcy search, company search if relevant
- For new builds: HomeBond or Premier Guarantee, architect's certificate of compliance, BCAR documents
- For second-hand: declarations on extensions, septic tank registration, septic tank inspection reports
That's twenty to forty PDFs, often scanned, often with handwritten margins, often with a critical fact buried in a special condition on page nineteen. The conveyancer's job is to read all of it, cross-reference, and produce a report on title. The honest version of the workflow is that experienced solicitors skim aggressively and rely on pattern recognition to know where to slow down. The risk is that the pattern doesn't always hold.
How an intelligence brain ingests this material
The ingestion side is more engineering than AI. Scanned PDFs go through OCR — and not just any OCR; folios have a specific layout that benefits from layout-aware extraction so that Part 1 (description), Part 2 (ownership), and Part 3 (burdens) end up as separate structured blocks rather than a wall of text. Filed plans need to be kept as image references with the folio they belong to. Special conditions in a contract should be extracted as discrete items, each linked back to the page and clause they sit on.
Then there's classification. The brain needs to know that a given document is a folio versus a planning grant versus a requisitions reply, because the questions you ask of each are different. Embeddings handle semantic search — "find me anything that mentions a right of way over the avenue" — but for property work you also want structured fields: folio number, county, registered owner, burdens, dealing numbers. A hybrid approach where structured extraction sits alongside vector search is the only way to get reliable answers to questions like "is this folio subject to a Section 72 burden that would survive a sale?"
For Irish DD specifically, the brain also benefits from a small ontology of local concepts: PRA dealings, Land Registry mapping notes, Section 49 ESB wayleaves, Section 72 burdens, the difference between absolute and possessory title, the relevance of a Land Commission charge on rural folios. These aren't things a general model knows in any depth.
The questions worth automating first
You don't deploy an intelligence brain to do conveyancing. You deploy it to do the cross-referencing that humans are bad at and computers are good at. The high-value questions in property research with an intelligence brain tend to be the ones where the answer requires three or four documents to agree:
- Does the contract description match the folio? The contract says "all that and those the dwellinghouse and lands". The folio says something specific. Are they the same property? Has anything been carved out?
- Are all the burdens on the folio addressed somewhere? If the folio shows a right of way at entry 3, do the requisitions or special conditions confirm it, vary it, or ignore it?
- Does the planning history cover every visible structure? The filed plan shows a footprint. Aerial imagery shows an extension. Is there a grant, a compliance cert, or a declaration of exempted development for it?
- Are the replies to requisitions internally consistent? A vendor saying "no disputes" in one reply and referring to a boundary issue in another is the sort of contradiction a tired solicitor misses at 6pm.
- Do the special conditions modify standard conditions in ways that hurt the buyer? Special condition drafting is where vendor's solicitors put the things they don't want you to look at twice.
Each of these is a structured question with a structured answer. The brain produces a draft answer with citations to the source documents; the solicitor reviews and signs off. The work doesn't disappear — it gets faster and more thorough at the same time.
Keeping the data on your side of the wall
Property files contain personal data, financial data, and information that's commercially sensitive to clients. Sending all of that to a cloud API run by a US hyperscaler is, at best, a Data Processing Agreement problem and, at worst, a Solicitors Accounts Regulations and confidentiality issue you don't want to argue about with the Law Society.
This is why the deployment model matters as much as the model itself. The version of the intelligence brain that I build runs on-premise or in a tenancy you control, with the document corpus never leaving your environment. Inference happens locally against open-weights models that are good enough for this kind of structured reasoning — you do not need a frontier model to compare a folio description to a contract description. You need a competent model with excellent retrieval and a clean corpus.
The other reason for on-premise is auditability. When a client asks "how did you spot this?", you want to be able to point to the documents, the extracted fields, and the reasoning trace. Cloud-only systems make that hard. A self-hosted system makes it the default.
What this changes about a property practice
The honest answer is: not the legal judgement, not the client relationship, and not the responsibility for the advice. What it changes is the floor. Junior solicitors stop missing things that experienced ones would catch. Senior solicitors stop spending two hours doing what should take twenty minutes. The report on title becomes more comprehensive without becoming more expensive to produce. And — the part that matters commercially — the firm can take on more work without hiring proportionally, or take on the same work with better margins.
It also changes how a firm holds its own knowledge. Every matter that goes through the brain becomes part of the corpus. The next time a solicitor faces an unusual covenant on a folio in Tipperary, the brain can surface the previous three matters where the firm dealt with something similar and what was negotiated. That institutional memory currently lives in the heads of a few partners. It shouldn't.
Where to start this week
Pick one closed matter from the last six months — ideally one that had a quirk in it, a special condition that mattered, or a burden that needed working around. Take the full document bundle and run a thought experiment: if a brain had been ingesting this from day one, which three questions would have saved you the most time, and which one fact would you most want flagged automatically next time? That list is the specification for a pilot. Start there, not with a procurement exercise. Property DD rewards specificity, and so does building the system that supports it.