For a commercial real estate acquisition shop of 4 to 20 people, building a custom AI system for due-diligence review is almost always the wrong first move — not because a build is hard, but because diligence is episodic. A custom lease-abstraction pipeline running on a steady portfolio can amortize its cost across constant volume. A diligence system does not: it fires for a few weeks inside a deal window, then sits idle until the next acquisition, decaying between deals while no one watches. That idle time is what breaks the usual build-vs-buy math. The honest field has four paths, not two, and the variable that decides between them is your deal cadence, not your monthly document count. This is that framework: four options, one test keyed to how often you buy, and the single condition under which building actually pays.
The four paths, defined
Most coverage of AI due diligence frames the choice as buy a tool or keep doing it by hand. That framing hides two options that are frequently the right answer for a small shop. Name all four and the decision gets easier.
Path one — a thin LLM workflow. You use ChatGPT, Claude, or Gemini directly, with saved prompts that read the documents in the data room and return the summaries you need: a lease abstract, an estoppel comparison, a rent-roll reconciliation, a service-contract flag list. No integration, no data store, no engineering. A person verifies the terms that decide money. Cost is a per-seat subscription you already pay for other work. This is the path no vendor sells you, which is why it rarely appears in a tool list.
Path two — an off-the-shelf tool, rented per deal. You buy access to a purpose-built product — a lease-and-document intelligence tool such as Prophia, Leasecake, or Trullion for the reading layer, or a legal-grade contract reviewer such as Kira or Luminance for clause-heavy deals — and point it at the diligence set. The vendor maintains the extraction, carries the accuracy claim, and gives you audit trails. Because diligence is deal-driven, the honest way to think about the cost is per engagement: you turn it on when a deal is live.
Path three — an outsourced diligence or abstraction service. You hand the lease stack and financials to a specialist provider that combines AI extraction with human review and returns finished abstracts and exception reports. You buy the outcome, not the tool. Pricing is per document or per deal, so cost scales with the size of the room and you pay only when you have a live transaction. For a shop doing two or three deals a year, this is often the lowest-total-cost path and the invisible option in every AI listicle.
Path four — a custom build. You commission a system on general-purpose components: a document-AI reading layer, a language model for extraction and normalization, and code that routes documents, validates output, and writes it into your underwriting model or data room. You own the logic, the data, and — the part that decides the whole question — the maintenance, including the maintenance of a system that runs a few weeks a year.
Getting the reading layer right sits under all four paths, and it is the subject of our guide to turning lease stacks into structured data. The four-path question is which wrapper around that reading layer your shop should actually pay for.
Why diligence breaks the usual build-vs-buy math
The standard argument for building custom automation is amortization. If you process hundreds of similar documents every month, a maintained pipeline can beat per-unit product pricing, and the build repays itself through constant use. That logic is sound for continuous work like portfolio lease administration. It falls apart for diligence, and the reason is timing.
Diligence is time-boxed and episodic. A purchase-and-sale agreement grants a defined diligence window — commonly around 30 to 60 days — in which the buyer reviews leases, estoppels, rent rolls, service contracts, title and survey, environmental reports, and financials, then decides whether to proceed, renegotiate, or walk. The work is intense, then it stops. For a small shop closing a handful of acquisitions a year, the machine you would build runs for a few weeks and then goes quiet for months.
Idle systems do not stay healthy. A custom extraction pipeline drifts as document formats change and the underlying models update, and none of that drift announces itself. It surfaces the week your next deal opens, when the tool you have not touched since the last closing returns a wrong renewal date with full confidence and no one on staff can diagnose it. A portfolio pipeline gets exercised constantly, so problems show up early and get fixed. A diligence pipeline gets exercised on exactly the days you cannot afford a surprise.
That is the core distinction from ordinary document automation. Build for steady, high-volume work and use itself keeps the system honest and amortizes the cost; build for episodic, high-stakes work and you pay the standing maintenance bill without the volume that would justify it — then discover its failures at the worst moment. The full cost of a build, including that tail, is laid out in our breakdown of what a custom document-automation project actually costs.
What AI actually does well inside a diligence window
Before choosing a path, be precise about which diligence tasks AI genuinely accelerates, because the answer bounds how much any of these options is worth.
AI is strong on the document-reading workstreams. It abstracts leases into your fields — parties, term, base rent, escalations, options, co-tenancy, exclusives — fast enough to clear a large stack inside the window. It reconciles a rent roll against the underlying leases and flags mismatches. It compares executed estoppels and SNDAs against the lease terms and surfaces discrepancies. It reads service and vendor contracts for assignment, termination, and change-of-control clauses. It extracts line items from operating statements and pulls them into a normalized structure. These are the tasks where a fast first-pass reader saves the most hours, and they are exactly the tasks the tools and workflows above are built for. Our guide to the AI tools that fit a diligence stack covers which product handles which of these workstreams.
AI is weak, or simply absent, on other parts of diligence. A Phase I environmental assessment is a licensed professional’s site work, not a document-reading task. Title and survey review turns on legal judgment and local practice. The negotiation of a discovered problem is human. And on the document work itself, accuracy is uneven: vendors report figures above 95% on standard fields, and that holds for parties, dates, and base rent, but it drops on the non-standard clauses — an unusual co-tenancy trigger, a bespoke exclusive, a hand-annotated amendment — which is precisely where diligence risk concentrates. Treat headline accuracy as a claim about the easy fields, and keep a human confirming the clauses that decide the deal. This is the same trap that separates a general assistant from a purpose-built reviewer, which we examine in where ChatGPT and purpose-built document AI each break on lease review.
So AI compresses the document-reading half of diligence and leaves the judgment half where it was. No path — bought, rented, outsourced, or built — removes the reviewer who checks the clauses that matter.
The four-question decision test
Run your shop through these four questions in order. The variable that should drive the answer is deal cadence, not monthly document volume.
1. How many acquisitions do you close a year? This is the amortization question. At a few deals a year, no build repays itself, and the fixed cost of owning a system dwarfs the diligence hours it saves. In that range you belong on a thin workflow, a per-deal tool, or an outsourced service. Only a shop running a genuine acquisition program — many deals a year, a repeatable diligence process — generates the frequency that makes a standing system worth maintaining.
2. How standard and repeatable is your diligence set? If every deal brings a different asset class, a different document mix, and a different set of questions, you want the flexibility of a general workflow or a service that adapts per deal. If you buy the same kind of asset over and over — the same lease structures, the same rent-roll format, the same checklist — repetition is what a custom system needs to earn its keep, and it starts to matter alongside frequency.
3. Who owns it between deals? This is the question that ends most build cases. A custom diligence system needs someone to keep it accurate through the quiet months so it works when the next deal opens. If your shop has no one who can diagnose a broken extraction — and most 4-to-20-person shops do not — a build is a liability that will surface at the worst time. A vendor or a service provider is your maintenance team on the paths that rent the capability.
4. Can you meet your confidentiality obligations on this path? Diligence runs on material under NDA — tenant financials, rent rolls, draft agreements. Any path has to satisfy that. The workable pattern is a business-tier account or a contracted provider where inputs are not used for training by default, plus a rule to limit what leaves your control. Confirm the terms of the exact plan or contract, because a shop handling confidential deal data cannot assume the default setting is safe.
If questions one and two do not both point hard toward frequency and repetition, you are on one of the three buy-or-rent paths. The test is not trying to argue you into a build; it is making sure you only build when cadence genuinely demands it.
The four paths, side by side
| Dimension | Thin LLM workflow | Off-the-shelf, per deal | Outsourced service | Custom build |
|---|---|---|---|---|
| Best for | 1–2 deals/year, ad hoc review | A few deals/year, standard documents | A few deals/year, want the finished outcome | A frequent, repeatable acquisition program |
| When you pay | Per-seat, already owned | Per engagement, deal-driven | Per document or per deal | Build cost plus standing maintenance |
| Who maintains accuracy | You, per document | The vendor | The provider | You, including between deals |
| Setup before first deal | None | Light onboarding | A briefing | A full build |
| Data ownership | You keep outputs manually | Vendor’s store | Provider delivers files | Full — you own the layer |
| Idle-time risk | None | None — off when no deal | None | High — decays between deals |
| Audit trail | Manual | Built in | Provider supplies | You build it |
Reading left to right, the first three paths share one property that suits episodic work: you pay when a deal is live and nothing rots while you wait for the next one. The build is the only column carrying idle-time risk, and for a small shop that column also carries a maintenance burden no one on staff can hold. The right seat is set by how often you transact, not by which product gives the best demo.
The default for a small shop, and the one trigger that flips it
The honest default for a 4-to-20-person acquisition shop is to rent the capability, not own it — a thin LLM workflow when volume is low and the team is fluent, an off-the-shelf tool turned on per deal when the documents are standard, or an outsourced service when you would rather buy the finished abstracts than run the tool. All three match the episodic shape of diligence: they cost you money only when you have a deal, and none of them decays in the gaps.
One condition flips that default toward a custom build, and it is a combination, not a single factor: a frequent, repeatable acquisition program with a proprietary diligence thesis. You are closing many deals a year, in a consistent asset class, against a checklist and a data model you have refined and want to own — and the extracted data feeds an underwriting engine or a portfolio system that no vendor integrates with. At that cadence the system is exercised often enough to stay healthy, the repetition lets it specialize, and owning the data layer becomes a real edge rather than a convenience. Below that threshold, a build is expensive insurance against a problem your deal volume does not create.
This is the same discipline that governs the small-firm advantage generally: the edge is speed and low overhead, and a standing obligation you cannot staff quietly erases both. That argument runs through the playbook on how small CRE firms out-operate larger competitors. Owning a diligence machine you use three weeks a year is the opposite of the low-overhead posture that makes a small shop fast.
How to prove a path on one real deal
Whichever path the test points to, prove it on a live or recent deal before you commit to it as your standard. The verification is the same shape every time.
Take the document set from one real acquisition — including the messy parts: the scanned estoppels, the hand-marked amendments, the non-standard lease, not just the clean recent leases. Run it through the candidate: the workflow’s prompts, the product’s trial, or a paid pilot with the service. Then have someone who knows diligence check the output against the source and count two things — how often it is right on the standard fields, and how often it is right on the clauses and reconciliations that would change your price or your decision. A tool that nails base rent and misses a co-tenancy trigger has not passed, and no demo will show you that failure. The test costs a day against a deal you have already worked, and it tells you two things at once: whether the path is accurate enough, and whether your reviewers can actually judge the output.
That second reading is the one shops skip. A team that adopts AI diligence before its people can spot a wrong abstract has bought speed it cannot quality-check. Fluency to read and correct an AI-generated summary comes first; the choice among these four paths comes second.
Frequently asked questions
Should a small acquisition shop build or buy AI for due-diligence review?
Buy or rent, in almost every case. Due diligence is episodic — it fires for a few weeks inside a deal window, then stops — so a custom system runs a few weeks a year while carrying a full-time maintenance obligation, and it decays between deals when no one exercises it. A thin LLM workflow, an off-the-shelf tool turned on per deal, or an outsourced service all match that episodic shape and cost you money only when a deal is live. Building is defensible only for a shop running a frequent, repeatable acquisition program with a proprietary diligence thesis.
What does AI actually do in commercial real estate due diligence?
It accelerates the document-reading workstreams: abstracting leases into your fields, reconciling a rent roll against the underlying leases, comparing estoppels and SNDAs to lease terms, flagging assignment and change-of-control clauses in service contracts, and extracting line items from operating statements. It does not perform the judgment work — Phase I environmental assessments, title and survey review, and negotiation stay human.
How accurate is AI at reviewing leases and diligence documents?
Vendors commonly report accuracy above 95% on standard fields such as parties, dates, and base rent, and that holds for the easy fields. It drops on non-standard clauses — unusual co-tenancy triggers, bespoke exclusives, hand-annotated amendments — which is exactly where diligence risk concentrates. Treat the headline number as a claim about the easy fields and keep a human verifying the money clauses against the source.
Can I just use ChatGPT or Claude for diligence document review?
For a shop doing one or two deals a year, often yes. A thin workflow — saved prompts in ChatGPT, Claude, or Gemini that return your abstracts and reconciliations, with a person verifying the important terms — clears a diligence set at the cost of a subscription you likely already have. What you give up is a queryable data store, built-in audit trails, and confidence scoring, which is why standardized, higher-frequency shops move to a purpose-built tool or an outsourced service.
When is an outsourced diligence service the right choice?
When you would rather buy finished abstracts and exception reports than operate a tool, and your deal cadence is low enough that per-deal pricing beats owning anything. A specialist combines AI extraction with human review and returns the output, priced per document or per deal, so you pay only when a transaction is live. For a shop closing two or three acquisitions a year, this is frequently the lowest-total-cost path and the option most AI tool comparisons never mention.
What is the biggest mistake small shops make with AI due diligence?
Commissioning a custom system for work that only happens a few weeks a year. The build is cheap and quick; the standing obligation to keep it accurate through the idle months is neither, and a shop with no engineer finds the system broken exactly when the next deal opens.
How much does custom due-diligence automation cost to build?
A scoped custom automation project generally runs $25,000 to $150,000 in the current market, depending on complexity, and for diligence you add a maintenance tail on a system that runs only during deal windows. Because it is exercised so rarely, that idle-time upkeep is harder to justify than for a continuously used pipeline, which is why renting the capability wins for most small shops.
Is our confidential deal data safe with these AI tools?
It can be, but you must verify the specific plan or contract. Diligence documents are under NDA, so the workable pattern is a business-tier account or a contracted provider where inputs are not used for training by default, plus a rule limiting what leaves your control. Confirm the terms of the exact plan or service, because a shop handling tenant financials and draft agreements cannot assume the default setting is safe.
Do I still need people reviewing the output if the AI is accurate?
Yes, on every path. AI is a fast first-pass reader of diligence documents, not a substitute for judging the clause or reconciliation that would change your price. Grounding features and confidence scores shrink the reviewer’s job, but the fields where accuracy drops are the fields with financial consequences, so a person confirms them against the source before anyone relies on the abstract.
Where to start
The first question is not which diligence tool to buy or whether to build one. It is which of the four paths your deal cadence points to — and whether your team is fluent enough that any of them is safe yet. A free AI-readiness assessment produces that read: a short working session that maps your acquisition frequency, your document mix, and your review workflow, and returns an honest recommendation for whether a thin workflow, a per-deal tool, an outsourced service, or a month of fundamentals first is the right next move. Book a free AI-readiness assessment before you commit a dollar to the build-vs-buy question.
Arthur Wandzel
