For most commercial real estate firms of 4 to 20 people, a custom document-AI pipeline is the wrong first answer. Off-the-shelf tools — or a thin workflow built on ChatGPT, Claude, or Gemini — will read your leases, pull your fields, and cost a fraction of a build, and they arrive without a maintenance bill you have no staff to pay. Custom becomes the right answer only under specific conditions: high monthly volume, a document type no vendor handles well, or a data layer you need to own. This is a decision framework, not a tool ranking. It gives you a five-question test that maps your firm’s real situation — leases per month, clause complexity, who fixes it when it breaks — to one of three paths, and it names the triggers that flip the default from buy to build.
The three paths, defined
The buy-vs-build question in commercial real estate document work is usually framed as two options. It is really three, and naming the middle one changes most small-firm decisions.
Path one — a thin LLM workflow. You use ChatGPT, Claude, or Gemini directly, with a saved prompt that returns a lease abstract in your firm’s template. No integration, no data store, no engineering. You paste in a lease or attach the PDF, the model returns your fields, and a person verifies the clauses that matter. Cost is a per-seat subscription. This is the option the tool listicles never mention because no one sells it to you.
Path two — off-the-shelf document AI. You buy a purpose-built proptech product — Prophia, Leasecake, Visual Lease, or Kira — that abstracts leases, stores the output as structured data, and gives you audit trails, confidence scores, and portfolio views. The vendor maintains the extraction, updates it for new formats, and carries the accuracy claim. Cost is a subscription, usually priced per lease or per portfolio.
Path three — a custom pipeline. You build your own system on general-purpose components: a document-AI primitive for the reading layer (Azure Document Intelligence, Google Document AI, or AWS Textract), a large language model for extraction and normalization, and code that stitches them together, validates output, and writes it where you need it. You own the logic, the data, and — this is the part that decides the whole question — the maintenance.
“Off-the-shelf document AI” and “custom pipeline” are the two poles of the title. The thin LLM workflow sits before both, and for a firm doing a handful of leases a month it is frequently the honest answer. Getting the reading layer right is the foundation under all three paths, which is the subject of our guide to turning lease stacks into structured data.
What off-the-shelf document AI actually gets you
Purpose-built tools earn their subscription on four things a raw model does not give you.
A data layer, not just an answer. A model returns text. A product like Prophia writes every abstract into a structured store — rent rolls, critical dates, clause-level terms, stacking plans — that you can query across a portfolio. If your question is “which of my leases have a co-tenancy clause that triggers next year,” a product answers it and a chat window does not.
Grounding and audit trails. The better platforms link every extracted value back to its exact location in the source document, so a reviewer checks a clause in seconds instead of hunting the PDF. Vendors report accuracy above 95% on standard fields — parties, dates, base rent — and that number is real for the easy fields. Treat it as a vendor claim, and remember it drops on the non-standard clauses where money actually lives.
Maintenance you do not perform. When a landlord changes its lease template, or a scanned document defeats last quarter’s model, the vendor’s team fixes it. For a firm with no IT department, this is the quiet core of the value: someone else is on call.
Operations features around the extraction. Leasecake, for instance, is built around renewal and critical-date alerts for retail operators, and Visual Lease leads with lease-accounting compliance. You are often buying the workflow as much as the reading.
The cost is loss of control and per-unit pricing that scales with your portfolio: you accept the vendor’s fields and a bill that grows with lease count. For a small firm that mostly needs clean abstracts and a few alerts, that trade is usually worth it. For how two of these products stack up against building your own, we compare Prophia against a custom lease-abstraction build and Leasecake against a custom-built alternative.
What a custom pipeline actually costs
The consultant blogs describe a custom pipeline as a weekend of wiring: point Textract at a PDF, hand the text to a model, get JSON back. The build genuinely is close to that simple. The build is not the cost. The maintenance tail is the cost, and it is where small firms get hurt.
A custom pipeline in production has to survive the real world. Landlords use dozens of lease templates and change them without notice. Scans arrive rotated, faxed, or annotated by hand. A model that returned perfect fields in March starts drifting in June as documents shift and the underlying model updates. None of this breaks loudly. It breaks silently — a wrong renewal date extracted with full confidence, a missed exclusive clause — and someone has to notice, diagnose, and fix it. In a firm with no engineer, that someone does not exist, and the pipeline degrades into a tool nobody trusts.
That is the honest total cost of ownership: not the initial build, which a competent developer or a capable agency delivers inside a normal custom-automation budget, but the standing obligation to keep it accurate. In the current market, a scoped custom automation project runs roughly $25,000 to $150,000 depending on complexity, and the number a small firm forgets to add is the recurring cost of the person — internal or retained — who owns it after launch. A product folds that maintenance into a subscription. A build hands it to you.
None of this means custom is wrong. It means custom is a commitment, not a purchase, and a firm should build only when the returns clearly justify carrying that commitment. The full cost comparison across all three paths is laid out in our breakdown of what automated lease abstraction actually costs in 2026.
The five-question decision framework
Run your firm through these five questions in order. The first “build” answer that is unavoidable points you toward a custom pipeline; if none is, you belong on one of the two cheaper paths.
1. How many leases do you process per month? Under roughly 10, a thin LLM workflow or an entry off-the-shelf tier is almost always right — the volume never repays a build. In the dozens, off-the-shelf earns its keep. In the hundreds, with the same document types repeating, per-lease product pricing starts to exceed what a custom pipeline would cost to run, and build enters the conversation.
2. How standard are your documents? If you handle conventional office, retail, and industrial leases, every off-the-shelf tool is trained for you. If your bread and butter is a document type vendors handle poorly — ground leases with unusual structures, a niche asset class, a non-English portfolio — a custom pipeline you tune to your exact documents can beat a generic product. Standard documents favor buying.
3. Do you need to own the structured data? If you only need abstracts to inform a deal or feed a report, a product’s data store is fine. If the extracted data is a strategic asset — you are building a proprietary market database, or you need it inside systems no vendor integrates with — owning the pipeline and the data underneath it starts to matter.
4. Who fixes it when it breaks? This is the question that ends most build fantasies. If your firm has no one who can diagnose a broken extraction, a custom pipeline is a liability regardless of the other answers, because it will break and no one will be watching. A product’s vendor is your maintenance team. Be honest about whether you have a replacement.
5. Is fluency in place first? A firm that automates lease abstraction before its people can judge an AI-generated summary has built a machine no one can quality-check. The reviewers who verify the output — on any path — need to know what a good abstract looks like and where models fail. That capability comes before the tooling decision, not after, and it is the argument running through the small-firm AI playbook for out-operating larger competitors.
If questions one through four all point to buy, buy. The framework is not trying to talk you into a build — only to make sure you build when the situation genuinely demands it.
The three paths, side by side
| Dimension | Thin LLM workflow | Off-the-shelf document AI | Custom pipeline |
|---|---|---|---|
| Best for | Under ~10 leases/month, one-off review | Dozens/month, standard documents | Hundreds/month or non-standard documents |
| Setup cost | A subscription | A subscription | Build inside a custom-automation budget |
| Ongoing cost | Per-seat | Per-lease or per-portfolio | Compute plus a maintenance owner |
| Data ownership | You keep outputs manually | Vendor’s structured store | Full — you own the data layer |
| Who maintains accuracy | You, per document | The vendor | You |
| Audit trail / grounding | Manual | Built in | You build it |
| Breaks silently? | Visible, one doc at a time | Vendor catches it | Only if you are watching |
The table makes the pattern obvious. Moving right buys you control and, at high volume, lower unit cost — and it transfers the maintenance burden onto a firm that may have no one to carry it. Moving left keeps you cheap and safe but caps how much structure and scale you get. The right seat depends on your answers to the five questions, not on which product demos best.
The default for a small firm, and what flips it
The honest default for a 4-to-20-person CRE firm is buy, not build — either a thin LLM workflow for low volume or an off-the-shelf tool once volume and the need for a data layer justify it. Most small firms never hit the conditions that make a custom pipeline pay, and the ones that build prematurely spend a custom-automation budget on a system that quietly rots because no one owns it.
Three triggers flip that default toward a custom pipeline, and it usually takes more than one:
- Volume plus repetition. Hundreds of similar documents a month, where per-lease product pricing has grown past what a maintained pipeline would cost to run.
- A document type off-the-shelf tools handle badly. When no vendor is trained on your core asset class and the accuracy gap is costing real money.
- Data as strategy. When the structured output is a proprietary asset you must own and embed, not a convenience you rent.
Absent at least one of these, a build is expensive insurance against a problem you do not have. This is the same logic that governs when off-the-shelf proptech is genuinely enough: the small-firm edge is speed and low overhead, and a maintenance obligation you cannot staff quietly erases both.
How to verify a path before you commit
Whichever path the framework points to, prove it on your own documents before you sign or build. The verification is the same shape in every case.
Take 10 real leases that represent your actual mix — including the messy, scanned, non-standard ones, not just clean recent PDFs. Run them through the candidate: the product’s trial, the thin workflow’s prompt, or a prototype of the pipeline. Then have someone who knows leases check every extracted value 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 that decide money. A tool that nails base rent and misreads a co-tenancy trigger is not ready, and the demo will never show you that. This ten-lease test costs a day and saves a firm from a subscription or a build it will regret — and it is the moment you learn whether your reviewers can actually judge the output, because if they cannot, no path is safe yet.
Frequently asked questions
What is the difference between off-the-shelf document AI and a custom pipeline?
Off-the-shelf document AI is a purpose-built product — Prophia, Leasecake, Visual Lease, or Kira — that abstracts leases and stores the output as structured data, with the vendor maintaining accuracy and carrying audit trails. A custom pipeline is a system you build yourself on general components — a document-AI reading layer such as Azure Document Intelligence, Google Document AI, or AWS Textract, plus a language model and your own code — where you own the logic, the data, and the ongoing maintenance.
When should a small CRE firm build a custom pipeline instead of buying?
Only when at least one specific trigger is present: hundreds of similar documents a month where per-lease product pricing exceeds a maintained pipeline’s running cost, a core document type that off-the-shelf tools handle poorly, or a strategic need to own the structured data. Absent those, and absent someone who can fix a broken extraction, buying is the right call for a 4-to-20-person firm.
Can I just use ChatGPT or Claude to abstract leases?
Yes, for low volume. A thin workflow — a saved prompt in ChatGPT, Claude, or Gemini that returns your template fields, with a person verifying the important clauses — handles a handful of leases a month at the cost of a subscription. What you give up is a queryable data store, built-in audit trails, and confidence scoring, which is why firms doing dozens of leases a month usually move to a purpose-built tool.
How accurate is AI lease abstraction?
Vendors commonly report accuracy above 95% on standard fields such as parties, dates, and base rent, and that holds for the easy fields. Accuracy drops on non-standard clauses — co-tenancy, exclusives, unusual escalations — which is exactly where the financial risk sits. Every credible approach keeps a human verifying those clauses against the source document, so treat headline accuracy numbers as vendor claims about the easy fields, not a reason to skip review.
What does a custom lease-abstraction pipeline cost to build?
A scoped custom automation project generally runs $25,000 to $150,000 in the current market, depending on complexity. The number small firms miss is the maintenance tail: the recurring cost of an internal or retained person who keeps the pipeline accurate as lease templates and models change. A product subscription folds that maintenance into the price; a build hands it to you.
Is our confidential deal data safe with these tools?
It can be, but you have to verify the specific plan. The safe pattern is a business-tier account where inputs are not used for training by default, plus a rule to anonymize NDA material where possible. Confirm the terms of the exact plan you buy or the exact API you build on, because terms change and a firm handling rent rolls and tenant financials cannot assume the default is safe.
How many leases per month justify moving off ChatGPT to a real tool?
Under roughly 10 leases a month, a thin workflow is usually enough; in the dozens, an off-the-shelf tool earns its subscription through the data store and time saved; in the hundreds of repeating documents, a custom pipeline starts to compete on unit cost. Volume, not brand preference, is the variable that should decide.
What is the biggest mistake small CRE firms make with document AI?
Building a custom pipeline they cannot maintain. The build is cheap and quick; the standing obligation to keep it accurate as documents and models drift is neither, and a firm with no one watching ends up with a system that fails silently and loses trust. The second mistake is automating before the team is fluent enough to catch the errors any path will produce.
Do I still need a person reviewing the output if the tool is accurate?
Yes, on every path. AI is a fast first-pass reader, not a substitute for reading the clause that decides money. The reviewer’s job shrinks — grounding features and confidence scores make verification quick — but the clauses where accuracy drops are the clauses with financial consequences, so a human confirms them against the source before anyone relies on the abstract.
Where to start
The first question is not which product to buy or whether to build. It is which of the three paths your firm’s real numbers point 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 lease volume, document mix, and workflows, and returns an honest recommendation for whether a thin workflow, an off-the-shelf tool, a custom pipeline, or a month of fundamentals first is the right next move. Book a free AI-readiness assessment before you commit a dollar to either side of the buy-vs-build line.
Dirk Jan van Veen, PhD
