Build vs. Buy AI in Real Estate: How Investment Firms Choose an AI Partner, Score the Evidence, and Model the Three-Year Cost

Why Build vs. Buy AI Has Five Answers for Real Estate Firms

If you lead AI at a real estate investment firm, you are probably talking to several groups at once. One sells a finished product priced per seat. Another sells a data platform. A consulting firm offers to build whatever you need, priced per project. A recruiter has a machine learning engineer who wants to build it in-house. The build vs. buy AI question sounds like a choice between two things. You are comparing at least five, each with a different pricing model, which makes it easy to rank them by whoever gave the best demo.

The stakes are real. MIT's NANDA initiative, as reported by Fortune, found that about 5% of AI pilot programs achieve rapid revenue acceleration while the vast majority stall with little to no measurable impact on P&L. The same research found that purchasing AI tools from specialized vendors and building partnerships succeed about 67% of the time, while internal builds succeed only one-third as often. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value. Read together, the findings name failure modes you can test for before you sign.

This article walks through one method you can apply to every option: sort them into types, evaluate them on the same dimensions, score only what you have verified, and model the cost over three years instead of one.

Define Your Needs Before the First Vendor Call

Write one page before your next call, then measure every option against that page instead of against each other's demos. It should answer seven questions:

• Outcome. What specific result, by when. "Asset management data in production by end of Q1" is a usable answer. "Use AI in asset management" is not.
• Priority workflows. The two or three workflows to improve first, and how long each takes today.
• Data reality. Your sources, how many operating partners report to you, in what formats, how often, and at what quality.
• Internal capacity. Who is available, for how much time, with what technical skills.
• Constraints. Security, compliance, budget range, and how approvals work at your firm.
• Ownership stance. What you are willing to rent, and what you must own.
• Success at six months. The result that would make this a good decision.

The ownership stance question shapes everything that follows. A firm that treats its underwriting logic and portfolio data as a competitive asset will weight exit terms and data location heavily. A firm that wants a single standard workflow fixed this quarter can accept more vendor control.

The Five Option Types: From Off-the-Shelf to Fully In-House

Most options fall into one of five types. Compare within and across types instead of product by product.

• A. Off-the-shelf SaaS. A ready-made product. It fits when the workflow is standard and you value speed and predictable cost over a perfect fit.
• B. Data platform. Centralizes and manages your data with strong integrations. It fits when you have many sources and simple analysis needs.
• C. Build partner. A consultant or services firm builds a system from scratch, and you own it. It fits when your needs are unique and you have a long-term budget.
• D. Hybrid. You own the system, and a partner provides the core engine under a subscription. It fits when you want ownership without starting from zero.
• E. Fully in-house. You hire a team and build it yourself. It fits when you have ample budget and a long time horizon.

You do not have to pick one type for the whole firm. A common pattern: keep the existing fund administration platform for LP reporting, buy a standard product for a commodity workflow, and take an ownership-heavy approach, hybrid or build, for underwriting and asset management, where your methodology is the edge. The question shifts from "build or buy" to "which type for which workflow." The guide includes a side-by-side table of each type's strengths, weaknesses and best-fit conditions.

AI Vendor Evaluation Criteria: The 11 Dimensions

Knowing how to evaluate AI vendors starts with fixing the dimensions before the first demo, so every option answers the same questions. Use these 11:

1. Outcome fit
2. Accuracy and trust
3. Ownership and control
4. Real estate depth
5. Production readiness
6. Integration and architecture
7. Adoption
8. Security and compliance
9. Long-term operations and cost
10. Economics (three-year total cost)
11. Team and viability

Strategic upside, meaning what else the system makes possible beyond the immediate problem, such as institutional memory for the investment committee or a stronger story for LPs, works as a tie-breaker. All 11 matter. For an investment firm, four deserve a closer look.

Accuracy and trust on your own documents

Accuracy figures quoted on a vendor's sample files tell you little about accuracy on yours. Ask three things: what accuracy they achieve on your documents, whether you can trace every extracted number back to its page and cell in the source, and how the system catches its own errors. A good answer names an error-detection mechanism, such as cross-checks between a rent roll total and the T12's rental revenue line, or a flag when a partner report's ending cash does not tie to the prior quarter. A red flag is no accuracy figure at all, no source tracing, or any version of "the AI doesn't make mistakes." In asset management, one wrong NOI figure that reaches an LP letter costs more credibility than months of manual work saved.

Ownership and control, including the exit

Ask where your data lives, who owns the code and configuration, what you can take with you if you leave, and whether the vendor offers source code escrow. Then ask for a walkthrough of the exit: step by step, what happens if you do not renew next year. Data held only on their platform, vague exit terms, and exports that lose structure are all red flags. Dependence on a primary AI provider is already common. In Zapier survey data reported by TechTarget, nearly three out of four respondents said losing their primary AI source would hurt day-to-day operations, and only 6% said they could walk away without disruption. Ownership matters in real estate because your extraction rules, partner mappings and underwriting conventions accumulate over years. If they live only inside a vendor's product, a change of partner sends you back to zero.

Real estate depth: JV structures, sponsor models, partner reports

Ask whether the team understands JV structures, sponsor models and partner reporting, and where their methodology comes from. The verification step is direct: hand them a real sponsor model or a quarterly partner report and ask them to interpret it. A team with depth will spot a promote tier that resets on a refinancing, or a capex reserve that the sponsor nets against distributable cash. A team without it will ask you to explain the business logic, and you will end up as their unpaid domain expert for the life of the contract.

Long-term operations and cost

Ask what the system costs to run after launch and why, who pays for model usage and how they optimize it, and what cost controls exist. Request a typical client's monthly run-cost breakdown. Red flags: no explanation of post-launch costs, model usage passed through with no caps or optimization, and maintenance that depends on one or two people. This dimension and the three-year cost model below are where options that looked similar in a demo start to diverge.

The guide covers all 11 dimensions with what to ask, how to verify each answer, and the red flags to watch for.

How to Score AI Vendors on Verified Evidence

Use a 1 to 5 scale for each dimension, and apply one rule without exception: anything not yet verified scores 3 at most. Higher scores require stronger evidence, and the top of the scale is reserved for results verified on your own data.

The rule exists because evaluations tend to reward presentation. The vendor with the best slides and the smoothest sample-file demo collects 4s and 5s on dimensions nobody has tested, while a stronger option with a plainer pitch scores lower. Capping unverified claims at 3 means the only way to score high is to show the work on your files. It also tells you where to spend your verification time: every dimension stuck at 3 is a test you have not run yet.

Pair the scale with pass/fail gates drawn from your constraints, such as "data must stay in our environment," so an option that fails a hard requirement is eliminated before weighting. The guide includes the scoring scale, example gates, a suggested weighting for a mid-market manager that wants to own its system, and a scorecard template.

Five Tests to Run Before You Sign

Run the same five tests on every option:

1. Messiest file test. Hand over your most irregular partner report and watch what happens live.
2. Source trace test. Pick three numbers at random from the output and ask to see where each one came from.
3. Exit walkthrough. Ask them to walk through, step by step, what happens if you do not renew.
4. Run-cost breakdown. Ask for the real annual cost after launch, line by line.
5. Reference call. Twenty minutes with a current client of similar size.

The messiest file test, in depth

Most demos run on files the vendor has seen many times. The messiest file test removes that advantage. Pick the document your analysts dread: the quarterly report from the operating partner who changes the layout every quarter, a scanned rent roll with handwritten move-out dates, a T12 that uses the sponsor's own chart of accounts, or an offering memorandum where the real rent roll sits in an appendix as an image. Pick one file instead of a curated set, and do not clean it first.

Then watch for five things:

• Live or sent away. If the vendor needs to take the file back for a week, you are testing their services team. That can be legitimate, so price and score it as a service.
• Structure. Extracting the words is easy. Does the output know that a line labeled "R&M" in one partner's report maps to "Repairs and Maintenance" in another's, and that a subtotal row is not a line item?
• Behavior on the unfamiliar. When the system meets a format it has not seen, does it flag uncertainty, or does it return a confident wrong number? The second is the dangerous failure.
• Reconciliation. Does it check internal consistency, such as unit counts against occupancy, or line items against stated totals, and tell you when they do not tie?
• Correction path. When you point out an error, how does the fix happen, and does it stick for the next file from the same partner?

Follow it with the source trace test on the same output. Three random numbers, three trips back to the source page. If any number cannot be traced, accuracy and trust stays unverified, and the score should show it. The guide includes the reference call questions to ask a current client about life after launch.

Where Real Estate Breaks Generic AI Tools

General-purpose AI evaluations tend to treat documents as interchangeable. Real estate investment work runs on a set of document types with their own structure, and the differences decide whether a system works in production. If you are evaluating AI for a commercial real estate investment firm, test against these:

• Partner and sponsor reports. Every operating partner formats differently, and formats drift over time. A system that needs a new template for each partner will not scale past a handful of JVs.
• Rent rolls. Unit-level data with concessions, lease dates, and charge codes that vary by property management system. See the rent roll red flags guide for what a reviewer checks.
• T12 operating statements. Sponsor-specific charts of accounts, one-time items buried in recurring lines, and months that need normalization. The T12 normalization guide covers the adjustments.
• Offering memoranda. Marketing narrative mixed with the data you need, often with the most important tables embedded as images. See how to analyze an OM.
• JV waterfalls. Promote tiers, catch-ups and clawbacks that differ deal by deal. A system that misreads a hurdle definition produces numbers that look plausible and are wrong. The American vs. European waterfall guide explains the structures.
• IC memos. Every figure in a memo should trace to the model. AI that drafts memos needs the same tie-out discipline described in how to write an IC memo.
• LP reporting. The highest-stakes output. Errors here reach investors, so accuracy, source tracing and an audit trail matter more than speed.

Build your test set from these document types, weighted toward the ones your priority workflows touch. A system that handles clean PDFs well and struggles with a sponsor's custom Excel layout will struggle in exactly the place your analysts spend their time.

AI Total Cost of Ownership: Why Most Cost Sits After Launch

Proposals show the visible part of the cost: build or implementation fees and the license or subscription. Once a system is live, most of the cost and risk sits in running it, and those costs rarely appear in a first-year quote.

The costs below the waterline:

• Model usage (AI tokens). Every document processed and every question asked consumes model capacity. The more people use the system, the more it costs.
• Cloud infrastructure. Storage, compute and the services the system runs on.
• Maintenance and operations. Keeping the system running as partners change formats, models change versions, and your workflows evolve.
• Quality assurance. Evaluation sets and the human review that confirms accuracy has held.
• Incident response. Someone has to respond when an extraction breaks the week before quarterly reporting.
• Internal staff time. Your people reviewing outputs, answering questions and managing the partner.
• Additional hires. Roles you will need to add, especially for build and in-house options.
• Training and rollout. Getting the team to use it, which per-seat pricing can discourage.

Add a line most cost models leave out: the cost of delay, or what you lose for each month later to production. An option with a lower three-year total that reaches production nine months later may cost more once you price the delay.

Model three adoption scenarios, low, base and high, for every option, because usage drives model and operating costs. Then track unit cost: cost per partner report processed or per deal underwritten. Unit cost measures efficiency and compares directly to the cost of doing the same work by hand. The three-year cost worksheet is in the guide, along with a hidden-costs checklist and the contract terms worth asking for.

What a Production AI System Needs After Launch

Launch is the start. A production system that your firm relies on for LP reporting or underwriting needs the same disciplines as any other critical software, whoever builds it:

• Code management and environments. Version control, code review, an architecture that stays coherent over time, and separate development, testing and production environments.
• Testing with an evaluation set. Automated tests, plus a set of your documents with known correct answers, re-run after every change to confirm accuracy has held.
• Releases and monitoring. Automated deployment with rollback and a record of every change, plus monitoring of uptime, accuracy and cost.
• Incident response, security, and backup. Who responds and how fast, how secrets, access and dependencies are managed, and how data and systems are restored.
• Documentation. Architecture notes and runbooks so the system never depends on one person.
• Model and prompt management. A record of which model and prompt version produced which result, so you can trace changes in quality.

AI coding tools have cut build time, and that is why owning your system is now realistic. One person can build a system. One person cannot safely own it: without code review, testing and shared knowledge, the system gets harder to understand each year, and no one covers when that person is away or leaves. A firm-wide production system needs product, engineering, evaluation, DevOps and security, and business reviewers, whether those roles sit in-house, with a partner, or in a mix.

Many real estate firms have not reached these controls yet. Deloitte's 2027 Commercial Real Estate Outlook, a survey of executives at commercial real estate owners and investors, found that while 72% of respondents had completed preliminary data mapping and source verification for their AI systems, fewer than half had implemented more advanced process and security controls. For a shared governance reference, the NIST AI Risk Management Frameworkis a voluntary framework, released in January 2023 with a Generative AI Profile added in July 2024, that your compliance team and any partner can work from.

Common Mistakes in AI Vendor Evaluation

• Ranking on demo quality. Demos run on the vendor's files. Without the messiest file test, you are scoring rehearsal.
• Comparing first-year price. A low implementation fee with uncapped model usage can cost more by year three than a higher fee with a fixed run cost.
• Treating the decision as firm-wide. Forcing one type onto every workflow means overpaying for commodity work or under-owning your edge.
• Skipping the exit walkthrough. Exit terms are easiest to negotiate before you sign and hardest to fix after your data and rules live in someone else's product.
• Accepting per-seat pricing without a rollout plan. Per-seat fees discourage the broad usage that makes the system pay off.
• Letting one engineer own production. It works until that person takes a vacation in the week before quarterly reporting.
• Leaving business reviewers out. Engineers can confirm a system runs. Asset managers and analysts confirm the numbers are right.

Decide in Phases and Show Results at Each One

Approve a first step with a clear checkpoint instead of committing everything at once. A first phase scoped to one priority workflow, measured against the success definition you wrote before the first vendor call, lowers the risk and makes the decision easier for leadership to approve. Continue what works and fix what does not.

Whatever you choose, keep a few principles. Own the roadmap: partners advise, you set the priorities. Keep ownership of and access to your data. Insist on documentation and a handover path so a change of partner never sends you back to zero. A new AI function earns credibility inside a firm one delivered phase at a time. The guide includes a one-to-two page recommendation memo outline you can take to your investment committee or board.

Get the Build vs. Buy Guide

Related Articles

• AI Deployment: the hub for guides on choosing, deploying and running AI at a real estate investment firm.
• How to Audit a Real Estate Financial Model: the review discipline any AI-built model should pass.
• Rent Roll Red Flags: what a reviewer checks, and what an extraction system must get right.
• How to Write an IC Memo: the tie-out standard for any AI-drafted memo.
• AI Tools for Institutional CRE: a comparison of the tools you are likely to evaluate.

FAQ

Sources

• Fortune: MIT report: 95% of generative AI pilots at companies are failing (August 2025): MIT NANDA findings on pilot outcomes and on purchased or partnered AI tools versus internal builds.
• THE Journal: Gartner: 30% of Gen AI Projects Will Be Abandoned (August 2024): Gartner's prediction on generative AI projects abandoned after proof of concept, and the reasons cited.
• NIST: AI Risk Management Framework: the voluntary framework for incorporating trustworthiness into the design, development, use and evaluation of AI systems, and its Generative AI Profile (NIST-AI-600-1).
• Deloitte: 2027 Commercial Real Estate Outlook (September 2026): survey of commercial real estate owners and investors on AI data mapping and process and security controls.
• TechTarget: Build vs. buy AI: A CIO's decision matrix (April 2026): Zapier survey data on operational dependence on a primary AI vendor.