Brokerage AI Build vs. Buy: A Decision Framework for Real Estate Leaders

Every brokerage leader reaches the brokerage AI build vs buy question in the same week that two things happen: a vendor demo comes in priced higher than expected, and someone asks whether you could just build it yourselves. Both reactions are reasonable. Neither is a decision.

It carries more weight in residential real estate than in most industries, because margins are thin and the data is licensed. AccountTECH's EBITDA Margin Index, reported by HousingWire in July 2025, put the average brokerage EBITDA margin at 3.50% for May 2025 across more than 150 firms, against a seven-year May average of 4.69%. A six-figure technology mistake does not get absorbed quietly at those margins.

What follows is the arithmetic for each path, the evidence on which one reaches production, and a six-question score you can run this month.

Table of Contents

1. Brokerage AI Build vs. Buy: What the Three Options Mean
2. What the Evidence Says About Building In-House
3. What Building Actually Costs a 45-Agent Brokerage
4. The Real Estate Constraints That Change the Math
5. A Six-Question Score That Settles It
6. What No Brokerage Should Build
7. A 60-Day Path From Question to Decision
8. Frequently Asked Questions
9. Ready to Settle Your Build or Buy Decision?

Brokerage AI Build vs. Buy: What the Three Options Mean

Build means owning the code, the roadmap, and every bug for as long as the tool exists. Buy means licensing a finished product and accepting someone else's roadmap. The third option, which most brokerages skip and which is usually the right first move, is configure: assembling what you need inside software you already pay for.

Buy is a per-seat subscription to a platform handling lead routing, follow-up, listing content, or transaction workflow. You get maintenance, compliance updates, and support. You give up control over priorities and, often, easy export of your own data.

Build is a custom application: a developer, a product owner, hosting, a model provider, and maintenance that does not end. Building is not writing a tool. It is agreeing to operate one.

Configure is the middle path: automation inside your CRM, a licensed team workspace for a general AI assistant, a no-code layer joining systems you already run. No new codebase, no new vendor, and a two-week feedback loop instead of four months. Much of what brokerages call building is configuring, which is cheaper and reversible. These are not equal choices: buy and configure are the default, and build is the exception you justify.

What the Evidence Says About Building In-House

The largest public dataset on this question favors buying by a wide margin. MIT's Project NANDA report The GenAI Divide: State of AI in Business 2025, published July 2025, found AI tools sourced through external partnerships reached deployment about 67% of the time against about 33% for internally built tools, with employee usage of external tools roughly double. Only 5% of custom enterprise AI tools reach production at all.

Two caveats, which the report states itself: the figures are self-reported, from 52 interviews and 153 survey responses, and the gap may reflect organizational capability rather than the choice itself. That second caveat is the useful part. A firm with a real engineering function is not the firm in the 33%, and almost no residential brokerage has one. The 33% is your number, not the 67%.

Brokerage behavior is moving the same way. In HousingWire's August 2026 coverage of Delta Media Group's Real Estate AI and Leadership Survey, a three-year study of more than 100 brokerage leaders, the share rating all-in-one AI and automation platforms 8 or higher out of 10 rose to 51.5% from 41.6% a year earlier. The direction is consolidation onto platforms.

What Building Actually Costs a 45-Agent Brokerage

A custom AI tool for a mid-sized brokerage runs roughly $78,000 in year one and $49,000 every year after, against about $48,000 and $45,000 for a bought platform — and the build carries the entire risk of never shipping. Here is the arithmetic for a hypothetical 45-agent brokerage wanting AI-assisted lead triage and listing content, with every assumption stated so you can substitute your own.

Build. The U.S. Bureau of Labor Statistics puts the median annual wage for software developers at $135,980 as of May 2025. Loaded at 1.3 times for taxes, benefits, and equipment, that is about $176,800, so a four-month first version costs $58,900. Add a quarter of an operations leader's time over those months at a loaded $95,000 salary ($7,900), $400 a month for model calls and hosting ($4,800), and $6,000 for legal review and the MLS data agreement work described below. Year one: $77,600. Year two is not zero. Hold the developer at 25% for fixes, model changes, and MLS schema updates, plus hosting: $49,000 a year, indefinitely.

Buy. At $80 per agent per month, 45 seats cost $43,200 a year, plus 60 hours of implementation and training at a loaded $45 an hour ($2,700) and 40 hours of annual admin ($1,800). Year one: $47,700. Year two: about $45,000.

Configure. AI workspace seats at $30 per agent per month, $16,200. An automation platform, $1,800. Roughly 120 hours of operations time to design the workflows, $5,400. Year one: $23,400.

Two numbers put that in context. At a 3.50% EBITDA margin, a $77,600 build has to be funded from the margin on roughly $2.2 million in revenue. And applying MIT's 33% deployment rate for internal builds, the expected cost per tool that actually ships is closer to $235,000. That second figure is illustrative, since the 33% comes from enterprises rather than brokerages, but it is the right shape of risk to bring to the conversation. Our breakdown of what AI implementation costs a small business covers the same hidden lines.

The Real Estate Constraints That Change the Math

Three things make brokerage software different, and all three push toward buying.

Your listing data is licensed, not owned. MLS data reaches you through a specific feed under a specific agreement. San Diego MLS, as one example, documents three access types: IDX for public website display, VOW for registered users, and a broker back-office feed for internal operations and analytics, the last two requiring a signed Broker Data Access Agreement and delivered in RESO Data Dictionary 2.0 format. Operate across three MLSs and you maintain three sets of rules, three schemas, three compliance reviews. A vendor serving hundreds of markets has absorbed that cost; you would absorb it alone.

Your agents are not employees. You cannot mandate adoption the way an enterprise IT department can. NAR's 2026 REALTORS Technology Report, a survey of 1,200 agents released September 2026, found that among agents using AI, 91% use ChatGPT, 42% Gemini, and 25% Claude. They already chose. Anything you build competes with a free habit.

You carry the supervisory liability either way. Florida Realtors' March 2026 guidance on why every brokerage needs an AI use policy is direct: brokers already have supervisory responsibilities tied to advertising, disclosures, and client communications, and when inaccurate AI output lands in a listing, the agent and the brokerage are still responsible. It flags fair housing specifically: descriptions of who a home is "perfect for" can drift into steering, and the violation counts even when unintentional. Buying does not transfer that liability, but building means you own the model's behavior as well as the review process. Our overview of AI for real estate brokerages covers that oversight.

A Six-Question Score That Settles It

Score each question 0, 1, or 2 for the project in front of you. The total points to a path, and two questions can veto a build outright.

1. Does your brokerage do this differently from every competitor in your market? Standard practice: 0. Somewhat: 1. A differentiator you can state in one sentence: 2.
2. Have you trialed a vendor product with real agents and real data? One worked acceptably: 0. All fell short on a named gap: 1. Nothing on the market does it: 2.
3. Who owns this in 24 months? Nobody named: 0. Someone already full-time elsewhere: 1. A named role with budgeted time: 2.
4. Does it need data no vendor can access on your behalf? No: 0. Partly: 1. Yes: 2.
5. Can you state the outcome metric and its baseline today? No baseline: 0. Rough: 1. Measured: 2.
6. How reversible is it? Switching away would be slow and painful: 0. Moderate: 1. Abandonable in a week: 2.

Reading the score. Zero to four: buy. Five to eight: configure inside what you own. Nine to twelve: a build is defensible, so write the business case.

The two vetoes. A zero on question 3 means do not build, whatever the total; an unowned tool becomes an unmaintained one. A zero on question 5 means do not buy yet either; you are shopping before you have defined the problem, the most expensive mistake here. Process first, technology second. Our twelve questions to ask before signing with an AI vendor picks up where this score leaves off.

What No Brokerage Should Build

Some categories are settled, and building them in-house moves risk onto your license for no gain.

• Consumer-facing text with no human reviewer. Fair housing exposure is not a problem you can engineer away. AI enhances human judgment here; it does not replace the person who signs off.
• Valuation or pricing logic. You would compete with firms whose entire business is AVM accuracy, and a bad number costs a listing or a lawsuit.
• Transaction management and forms. Compliance updates are the whole product. Vendors maintain them; you would not.
• Anything storing client financial data. The security obligations exceed a brokerage IT budget.
• Anything whose only champion is one enthusiastic agent. Enthusiasm is not an owner. See question 3.

Buying can be wrong too. If a vendor cannot export your data, cannot say what happens to information you enter, or is selling a capability your CRM already has, the answer is not a new contract.

A 60-Day Path From Question to Decision

Sixty days settles this without a quarter of leadership debate.

• Days 1 to 10. Survey what agents already pay for personally. Pick one problem and write down its baseline: response time, listings per agent, days from contract to close.
• Days 11 to 25. Score your top three candidate problems. Most will land in configure. Start there.
• Days 26 to 45. Put configure workflows in front of ten agents; run buy candidates as two parallel trials with the same agents and metric.
• Days 46 to 60. Decide against a kill criterion set in advance: no movement in the baseline, or fewer than six of the ten using it weekly, means stop. A clean stop costs a few thousand dollars; a quiet failure costs a year.

Frequently Asked Questions

Ready to Settle Your Build or Buy Decision?

Scoring a project honestly, pricing a build against the margin it comes out of, and designing a trial with real kill criteria are strategy questions that belong before any contract. ANtreliX is a strategy-first AI and business intelligence consultancy in North Myrtle Beach, South Carolina, working with brokerages across the Grand Strand and remotely. See our case studies or Talk to ANtreliX.