To prepare your firm’s data for an AI project, you do the opposite of what most guides tell you: you do not clean everything, you clean the slice of data that one specific workflow needs, and you do it after you know what that workflow is. Every enterprise article on this subject assumes a data engineer, a warehouse, and a year. A commercial real estate firm of ten people has none of those, and the good news is that it does not need them. Preparation, sized correctly, is a half-day of honest looking, not a project. This is what “getting your data ready” actually means for a lean firm, why the boil-the-ocean version is the expensive mistake, and the concrete steps that get you to a working automation without hiring anyone.
What Preparing Your Data Really Means
Preparing data for AI means getting the specific records a task depends on into a shape a machine can read reliably, and nothing more. That last clause is where small firms get talked into a project they never needed.
The enterprise version of this advice is real for enterprises. A bank standing up a fraud model needs governed pipelines, labeled datasets, and a data catalog because it is training a system on millions of transactions. Your firm is not doing that. When you ask a current AI tool to summarize a lease, screen a rent roll, or draft a market write-up, the model is reading a handful of your documents at the moment of the task, not learning from a warehouse you built in advance.
So the honest definition of preparation for a firm your size is narrow. It is making sure the files a given workflow touches are findable, readable, and correct enough to trust. Findable means you can point to them without a treasure hunt. Readable means a machine can extract the text, which is the whole problem with a scanned PDF. Correct enough means the numbers and dates that matter are right, because an AI that reads a wrong number will confidently repeat it. That is the entire job. Everything past it is scope you are paying for and do not need.
Start With the Question, Not the Data
The first step in preparing your data is refusing to prepare your data until you know the one question you want answered. Preparation without a target is just cleaning, and cleaning an entire archive is the single most expensive thing a small firm can do to itself.
Pick one workflow. Not a strategy, a workflow: “summarize incoming leases into a one-page abstract,” or “flag which of these fifty listings match a buyer’s criteria,” or “pull the key dates out of our active leases into a calendar.” Each of those depends on a small, nameable set of records. The lease-summary task needs the lease documents and nothing else. It does not need your contact database cleaned, your fifteen-year email archive sorted, or your accounting exports reconciled. When you start from the question, the data you must prepare shrinks from “everything we own” to “these forty files,” and the whole effort becomes a morning instead of a quarter.
Choosing that first workflow well is its own skill, and it is worth being deliberate about it. The method for finding the one or two tasks in your firm that are actually worth automating is laid out in a practical guide to workflow mapping and finding automation candidates. Do that thinking first. The data preparation falls out of it almost automatically, because the workflow tells you exactly which records are in scope and which are noise.
The Two Kinds of Data in a CRE Firm
Everything your firm stores is one of two kinds, and the split decides how hard preparation will be. Knowing which kind your workflow depends on tells you the cost before anyone quotes you.
Structured data is information already living in rows and columns: a contact list in a spreadsheet, a rent roll exported from your accounting system, a pipeline tracker. A computer reads this natively. Preparing it is a matter of tidying, fixing obvious errors, standardizing how dates and dollars are written, removing duplicates. This is cheap work and often something you can do yourself in an afternoon.
Unstructured data is documents: leases, letters of intent, offering memoranda, estoppels, CAM reconciliations, appraisals. This is the valuable half of a CRE firm’s memory, and it is the hard half, because most of it is not text a machine can read. It is scanned images, some crisp, many photographed at an angle or faxed a decade ago. A machine turns those images into text through optical character recognition, and OCR on a clean document is nearly free while OCR on a skewed, low-contrast scan produces a confident wrong answer that costs more to catch than a blank would.
The reason this distinction matters for preparation is that it sets your expectations honestly. If your chosen workflow runs on structured records, preparation is light and you can likely do it yourself. If it runs on your document pile, preparation is the real work, and the mess in those documents is the number a proposal is quietly estimating. The same reality drives the cost and timeline of moving records between systems, which is worked through in detail in an explainer on what data migration involves for a small firm. Preparation and migration are the same problem seen from two angles: both are about turning an accidental archive into something a system can use.
The Half-Day Inventory
The concrete first action is not cleaning and not buying. It is an inventory, and it is the step everyone skips because it produces nothing visible. For the workflow you chose, sit down and answer five questions on paper.
Where does this data live? Name the actual locations: a folder on the shared drive, a specific inbox, a proptech tool, a spreadsheet on someone’s desktop. You will usually find it is in more places than you expected, and that scattering is itself a finding.
How much of it is there? Count roughly. Forty active leases is a different job from four hundred. Volume is what turns a morning into a week.
What format is it in? Clean digital PDFs and native spreadsheets are easy. Scans, photos, and paper are the expensive cases. Be honest about the ratio.
How consistent is it? Does every file follow a naming convention, or is the same property filed three different ways? Do rent figures appear in one column or scattered across a broker’s private shorthand?
Who knows the quirks? Somewhere in your firm one person knows that “the Henderson file” is really under a former tenant’s name and that the 2019 rent roll double-counts a unit. That knowledge is part of the data, and it walks out the door if that person leaves.
An inventory like this takes a few hours and changes every conversation that follows. It replaces “our data is a mess” with a specific, sized picture, and it is the difference between getting a real quote and getting a guess. No credible partner should give you a timeline before you can answer these five questions, and you should be suspicious of one who does.
What “Clean Enough” Actually Means
You do not need clean data. You need data that is correct where it counts and consistent enough to process, which is a far lower bar than the perfection enterprise guides imply. Chasing perfection is how a half-day job becomes a lost quarter.
Cleaning has a point of diminishing returns, and for a small firm it arrives early. There are four things worth checking, and they are not equally important. Accuracy on the fields that carry money and dates is non-negotiable: a rent, a commencement date, an expiration, a square footage. A wrong number here poisons everything downstream. Consistency matters next, because a machine handles “1,200 SF” and “1200 sqft” and “1.2k” as three different things unless you standardize them. Completeness matters less than people fear; a missing field is honest, and a good process flags it rather than inventing a value. Duplication matters mostly because it wastes effort and inflates counts.
Notice what is not on that list: making every file beautiful, backfilling every empty cell, reorganizing the whole archive. Those are cosmetic against the one question you are answering. The discipline is to clean the fields your workflow reads and leave the rest alone. If the task is summarizing leases, the tenant names, dates, rents, and terms must be right; whether the folder is tidy is irrelevant to the machine.
Modern AI tools change this bar in your favor. They tolerate messy input better than older software did, reading a document in its natural language and pulling structure out rather than demanding structure going in. What they cannot do is invent a fact that is not there or correct one they cannot see is wrong, which is why a person still verifies a sample of the output, focusing on the money and the dates. Use the machine for the labor and a human for the judgment.
Where Your Documents Travel
A preparation step that generic guides treat as a footnote is, for a CRE firm, a first-class decision: before any document is fed to an AI tool, you decide where it is allowed to go. Deal terms, client financials, and ownership details are the sensitive core of the business, and “preparing” that data includes preparing a rule for its confidentiality.
The question to answer, in writing, is simple. When this file is read by a tool, where does it physically travel, who can see it, and is it retained or used to train anything? The answers differ sharply between a consumer chat product and a business-tier tool with data-handling commitments, and they differ again for a purpose-built system that keeps your files inside your own environment. None of this requires you to become technical. It requires you to ask the question before the files leave your control rather than after, and to get a plain answer rather than a shrug.
This is not a reason to avoid AI with confidential data. Firms handle sensitive documents with these tools every day. It is a reason to make the confidentiality decision part of preparation, deliberately, so that the convenience of a tool never quietly becomes an exposure you did not choose. Treating it as a setup step rather than an afterthought is the posture that separates a firm that adopts AI safely from one that gets surprised.
The “How Much Data” Myth
The belief that you need a large, clean dataset before you can start is the most common reason small firms delay, and it is out of date. It describes an older kind of AI that had to be trained on thousands of labeled examples. That is not what you are doing.
When you ask a current AI tool to abstract a lease or screen a set of deals, it brings its general capability to your specific documents at the moment of the task. It does not need to have seen a thousand of your leases first. Forty is plenty; four is often enough to prove the workflow. The volume that matters is not how much history you can assemble but how many live documents the task touches, and that is usually small. This inverts the anxiety: you are not blocked until you have built a big clean archive, you are ready as soon as you can point at the handful of files one workflow needs.
That reframe matters for how you spend, too. The instinct to fund a large data-cleanup program before doing anything useful is backward. Start with one workflow on a small set of records, prove it earns its keep, and let the results justify the next step. Whether that next step is buying a tool or commissioning something built for your firm is a separate decision, and the trade-off is worked through in a plain look at the build-versus-buy question for CRE owners. Preparation does not commit you to either path; it just gets you to the point where you can choose one with evidence.
What to Do This Month
Preparation is a sequence, and none of the steps require a technical hire. Run them in order.
First, pick one workflow that costs your firm real hours every week. Second, run the half-day inventory on the records that workflow touches: where they live, how many, what format, how consistent, who knows the quirks. Third, decide your confidentiality rule for those files before any tool sees them. Fourth, clean only the fields the workflow reads, correcting the money and the dates and standardizing formats, and stop there. Fifth, run the workflow on a small batch and have a person verify a sample. That is a complete preparation cycle, and most firms can do it inside a month without spending anything but attention.
What you will have at the end is not a finished data platform. It is proof, on one real task, that your data is ready enough to be useful, and a clear-eyed picture of what the rest of the archive would take. That evidence is what turns “should we invest in AI” from a leap of faith into a sized decision. How lean firms sequence these small wins into a durable operating advantage over larger, slower competitors is the argument running through the case for how small commercial real estate firms out-operate institutional players, and where preparation fits inside the wider choice between off-the-shelf tools and something built for you is mapped in the playbook on when off-the-shelf proptech is enough.
The firms that get value from AI are not the ones with the cleanest data. They are the ones who prepared the right small slice of it, proved a workflow, and moved.
Frequently Asked Questions
How do I prepare my firm’s data for an AI project?
Prepare the data one specific workflow needs, not your whole archive. Choose a task that costs real hours, such as summarizing leases or screening deals. Inventory the records it touches: where they live, how many, what format, how consistent they are. Decide where those files are allowed to travel before any tool reads them. Clean only the fields the task reads, getting the money and dates right and standardizing formats. Then run the workflow on a small batch and have a person verify a sample. That is a full preparation cycle, and it takes a small firm about a month without any technical hire.
Do I need to clean all my data before starting an AI project?
No, and trying to is the most expensive mistake a small firm makes. You clean the slice of data the specific workflow uses, not everything you own. If the task is abstracting leases, the tenant names, dates, and rents in those leases must be correct; the tidiness of the rest of your archive is irrelevant to the machine. Scope preparation to the question, prove one workflow, and only expand cleanup if a later project needs it.
Do I need a data warehouse to use AI?
No. A data warehouse is infrastructure for organizations training models on large datasets with a data team. When you ask a current AI tool to read your documents, it works with the files at hand at the moment of the task. A firm running on a shared drive, Outlook, and a few spreadsheets can prepare data and get value without building any central database, which is why the enterprise advice to “stand up a warehouse first” does not apply to a lean firm.
My data is a mess of spreadsheets and PDFs. Is it too messy for AI?
Almost certainly not. Messiness sets the cost and effort, not whether it is possible. Structured records like spreadsheets are cheap to prepare. Scanned PDFs are the harder case because they must be read by optical character recognition and verified, and a poor scan produces confident wrong answers. The honest way to know is a short inventory of the files one workflow needs. That replaces the feeling of “it’s a mess” with a specific, sized picture you can act on.
How much data do I need for an AI project?
Far less than most people assume. The idea that you need thousands of clean examples describes an older kind of AI that had to be trained first. Current AI tools apply their general capability to your specific documents at the moment of the task, so forty live leases is plenty and four is often enough to prove a workflow. The volume that matters is how many documents the task actually touches, which is usually small.
What is the difference between structured and unstructured data?
Structured data is already in rows and columns a computer reads directly, like a contact list or a rent roll export, and it is cheap to prepare. Unstructured data is documents, such as leases, letters of intent, and reconciliations, which hold their information as text or images rather than fields. Unstructured data must be read and extracted before a machine can use it, and for a commercial real estate firm it is the valuable half of the archive and the source of nearly all preparation cost.
Is it safe to put confidential deal documents into AI tools?
It can be, if you decide the rule before the files leave your control. Ask, in writing, where a document travels when a tool reads it, who can see it, and whether it is retained or used to train anything. Consumer chat products, business-tier tools with data-handling commitments, and systems that keep files inside your own environment give very different answers. Make that confidentiality decision a deliberate preparation step rather than an afterthought, and handling sensitive documents with AI becomes a controlled choice instead of an accidental exposure.
Should I hire a data engineer before starting?
For a first project, no. Preparation sized to one workflow is inventory, light cleaning, and a confidentiality decision, none of which requires a technical hire. The work is mostly domain knowledge about your own records, which you already have. If preparation reveals that a large document archive needs systematic extraction, that is a scoping question for a build partner, not a reason to add headcount before you have proven a single workflow earns its keep.
How long does data preparation take?
For one scoped workflow, usually days to a few weeks, and most firms complete a full cycle within a month. Structured records are prepared in an afternoon. A pile of scanned documents takes longer because reading and verifying them is human-checked work that cannot be safely rushed. The timeline follows the inventory: once you know the volume and format of the files one workflow touches, the effort becomes predictable, which is exactly why no one should quote you before that inventory exists.
Where to Start
Preparing your firm’s data for an AI project is not the year-long cleanup the enterprise guides describe. It is choosing one workflow, taking an honest half-day look at the records it touches, deciding where those files are allowed to travel, cleaning only the fields that carry money and dates, and proving the task on a small batch. Done that way, preparation stops being the barrier and becomes the first useful step.
The move before any of it is a clear read of what you are actually sitting on and which single workflow is worth preparing for first. A free AI-readiness assessment gives you exactly that: a short working session that looks at where your records live, how tangled the document pile really is, and which task would return the most for the least preparation, then hands you a plain recommendation instead of a sales pitch. Book a free AI-readiness assessment and start with the slice of data that will actually pay you back.
Dirk Jan van Veen, PhD
