Skip to content
All resources

Blog

Should you build AI in-house for commercial real estate, or buy a platform?

by Daniel Leong, GTM

Should your CRE firm build AI in-house or buy a platform? What building involves, the hidden costs, lock-in and exit, and a scorecard to decide.

Guide cover reading “Build vs buy”: should your CRE firm build AI in-house or buy a platform?

Short answers

Should a commercial real estate firm build its own AI or buy a platform?

There are really three options: build, buy a specialist platform, or extend a general AI tool you already license, such as Microsoft Copilot or an enterprise ChatGPT or Claude plan. Build when the AI itself is something you want to own and you have the engineers to run it for years. Buy a specialist when the goal is to get analysts, asset managers and deal teams working faster on CRE documents and data. Many firms land on a mix: buy the platform, then build the pieces that are truly specific to the firm. Nobody here is building the model itself: "build" means an application on a model provider's API, on cloud AI services, or on models you host.

What does it take to build AI in-house?

A working demo is quick. A tool your team can rely on takes much longer, because you also need document intake, links to your systems, permissions, a way to show where each answer came from, security review, rollout and ongoing upkeep. Each of those needs people, and the work does not end at launch.

What does it cost, and what is the hidden cost?

The visible cost is the build team and model usage. The hidden cost is upkeep: models change, connectors break, access rules shift and someone has to keep checking the output. This guide gives you a worksheet to put your own numbers on each line, so you can compare it with a quote.

Why do CRE firms want to build their own AI?

The reasons are sensible. Five come up most often.

1. It looks cheaper and easier
A general-purpose model is available by API, so a prototype seems like a few weeks of work.
2. It would fit the firm better
A tool built for your own workflows should match how you work.
3. It feels safer
Keeping sensitive deal and portfolio data away from a third party is a fair instinct.
4. It could become a product
Some firms want something they can white-label or offer to their own clients.
5. It avoids depending on a vendor
No lock-in, no surprise price rises, no risk that the vendor closes.

Each one holds up in part. A prototype is cheap. The fit question is real, and it is why a good platform lets you build custom skills around your processes. Data control matters, and it is something to test with any vendor (see the scorecard below), but a build on a model API still involves a provider, though running models through a cloud you already contract with can keep data inside agreements and regions you have already reviewed. A platform adds itself and, usually, the model providers it uses; ask for that list. The resale idea is a different business from running an asset management or investment firm, so treat it as its own decision. Vendor dependence is fair and testable: ask for export and exit terms before you sign.

What does the process of building AI in-house look like?

Here is a realistic sequence. The work is in stages 3 to 9, not stage 1.

  • 1. Pick the workflow

    What it involves
    Choose one task, such as reading offering memorandums or building a quarterly report. Write down what a good output looks like.
    Who you need
    Business owner, an analyst who does the task
  • 2. Get the data

    What it involves
    Get access to the documents and systems the task depends on, for example a SharePoint library or your property management system. Choose a model provider and agree its data terms (retention, training, region) and usage limits. Clean and map structured data, such as property IDs and charts of accounts, across systems.
    Who you need
    IT, data owner, data engineers
  • 3. Document intake

    What it involves
    Read PDFs, spreadsheets and scans reliably, including odd layouts and long files.
    Who you need
    Engineers
  • 4. Retrieval and reasoning

    What it involves
    Find the right passages across many files, and keep each lease read together with its amendments.
    Who you need
    AI engineers
  • 5. Evaluation

    What it involves
    Build a test set of real documents with the correct answers, measure accuracy, and rerun it every time the model or prompts change.
    Who you need
    AI engineers, analysts who know the right answers
  • 6. Evidence and review

    What it involves
    Show the source for each figure and give a reviewer a way to check and sign off.
    Who you need
    Engineers, product designer, the reviewer
  • 7. Permissions and security

    What it involves
    Match each person's access to what they can see today, log activity, pass your security review, test how far instructions hidden in outside documents, such as broker memorandums, can change what the tool does, and limit what the tool is allowed to do. Make sure documents can be deleted everywhere they are copied, including search indexes, embeddings and logs, when an NDA requires it. Name who signs off outputs and who is accountable when an abstracted date or figure is wrong.
    Who you need
    IT, security, engineers, the business owner
  • 8. Rollout

    What it involves
    Train users, handle objections, support them.
    Who you need
    Business owner, a champion in each team
  • 9. Run and improve

    What it involves
    Monitor accuracy, cost and uptime, respond to incidents, fix breakages, retest when models change, add workflows, answer support questions.
    Who you need
    A standing team

Notice what is missing from most prototypes: stages 5, 6, 7 and 9. They are also the stages that decide whether anyone trusts the tool.

How long does it take to build, and what does it cost?

It depends on scope, and a fair answer has no single figure. Instead, estimate it yourself. Fill in this worksheet with your own numbers, and get prices from providers' own pricing pages and from your recruiters.

  • Team

    Build: what to estimate
    Engineers, product and design people, for how many months, at what loaded annual cost (one-off build, then a standing team each year)
    Buy: what to estimate
    Internal owner and admin time each year
  • Licences and usage

    Build: what to estimate
    Model usage for your volume of documents and questions, including peaks at quarter-end or when a data room lands
    Buy: what to estimate
    Subscription per year, and what changes at renewal
  • Infrastructure and hosting

    Build: what to estimate
    Hosting, storage, monitoring, backups, test environments; self-hosted models add hardware and specialist staff
    Buy: what to estimate
    Usually included; check limits. On-premise or air-gapped deployment puts hosting back on you
  • Data engineering

    Build: what to estimate
    Cleaning and mapping property IDs, charts of accounts and systems, and keeping that data model current
    Buy: what to estimate
    Your side of the integration work
  • Integrations

    Build: what to estimate
    How many systems must connect, and who keeps each connection working
    Buy: what to estimate
    Which connections are included, and which plan covers them
  • Evaluation and QA

    Build: what to estimate
    Building and maintaining test sets, and analyst hours checking the right answers
    Buy: what to estimate
    Running the same test set in a trial, and checking outputs after vendor changes
  • Security, procurement and vendors

    Build: what to estimate
    Internal reviews, penetration testing, policies, and the contracts and data terms of each model and cloud provider
    Buy: what to estimate
    Procurement and security review of the vendor and its model providers, each year, and how the vendor deletes deal data on request
  • Monitoring and cost control

    Build: what to estimate
    Tracking accuracy, usage and spend, with limits so one workflow cannot run up the bill
    Buy: what to estimate
    Usage reporting and any caps in the contract
  • Rollout and training

    Build: what to estimate
    Training and support hours per user, per year
    Buy: what to estimate
    The same, less any vendor service included
  • Each workflow after the first

    Build: what to estimate
    Usually cheaper than the first, but still engineering time
    Buy: what to estimate
    Configuration and service time
  • Lock-in and exit

    Build: what to estimate
    Dependence on one model provider's API and on the engineers who built it
    Buy: what to estimate
    Export terms, notice periods and the cost of moving
  • Delay

    Build: what to estimate
    Months until the first live workflow, valued in analyst hours
    Buy: what to estimate
    Months of procurement and security review before the first live workflow

Then compare the three-year total with the full cost of buying: the subscription, implementation and internal admin time, integration work on your side, price changes at renewal, and the cost of leaving. For reference, Fifth Dimension publishes its starting prices: Base starts from $139 per user per month, billed annually, with a $20,000 annual minimum, and Enterprise starts from $50,000 a year. Enterprise adds data connections shaped around your systems, and security and deployment options for your requirements, so check which plan covers what you need. See Fifth Dimension pricing. Before deciding, run a time-boxed build spike and a vendor trial against the same test set of your own documents, so the decision rests on evidence.

What are the hidden costs of maintaining an in-house AI tool?

Launch is not where the cost ends. After launch, a homemade AI tool needs ongoing care.

Models change
Providers update and retire models. Every change can alter how your tool behaves, so someone has to retest your workflows each time.
Connectors break
Systems change their formats and permissions. Each connection needs an owner.
Access rules drift
People join, leave and change teams. Permissions need to follow, or the tool shows the wrong person the wrong file.
Review never stops
Someone has to keep checking that outputs match their sources, and decide what to do when they do not.
New requests keep coming
Every team wants its own workflow. A queue forms, and the build team becomes a support desk.
Knowledge sits with a few people
If the two engineers who built it leave, the tool becomes hard to change.
Security work repeats
Reviews, policy updates and incident planning come around every year.

Some of this stays with you even if you buy: someone still checks outputs, keeps access rules right and reviews the vendor every year. A vendor can take on the engineering underneath; ask each one which parts it maintains, including connectors and model changes, and how it tells you when outputs may change.

What are the business risks of building your own AI?

You spend the time and get little back
The tool may turn out to be a thin layer over a general-purpose model. It can answer questions about a single document, but not handle the multi-document work that saves staff real time. Test vendors for the same thing.
Your team has to check everything
If the tool cannot show where each figure came from, every output needs a full manual review. That removes much of the benefit.
Falling behind
While the build team works on plumbing, the models and the products around them keep changing.
Key person risk
A tool that only two engineers understand is a risk to the firm when they move on.

Buying has risks too:

Model changes you do not control
A vendor changes its models on its own timetable, and you still need to retest your outputs.
The roadmap is not yours
Your most important workflow may not be the vendor's priority.
Price rises at renewal
Agree renewal terms before you depend on the tool.
Other parties in the chain
The vendor usually relies on model providers of its own; ask for the list and their data terms.
Vendors fall behind too
Test what a vendor ships, not what it promises.

Security and governance need owners too. AI applications have their own class of security risks, and the OWASP Top 10 for Large Language Model Applications (opens in a new tab) (2025 edition) is a good checklist for your engineers and security team: prompt injection is LLM01, and unbounded consumption, which includes runaway cost, is LLM10. Someone also has to own decisions about acceptable use, oversight and review; the voluntary NIST AI Risk Management Framework (opens in a new tab) and its Generative AI Profile (NIST AI 600-1) give a structure for this.

Why do firms buy SharePoint instead of building their own document storage?

Most firms could write a document store. Few do. Microsoft already built SharePoint (opens in a new tab) and many details around it: permissions, search, sync and governance. Building your own would not give you an edge, and you would have to maintain it forever.

The same test works for AI. Ask: is this part of how we win, or is it plumbing? Document reading, retrieval, permissions and review trails are plumbing for most firms. Your investment criteria, your underwriting method, your house report format and your institutional knowledge are not. Those are the parts worth putting your own effort into, and a good platform should let you.

The analogy has a limit. If you leave SharePoint, your files leave with you in their own formats, though metadata, permissions and version history take work. With an AI platform, check what leaves with you: your documents and extracted data, the outputs and their source trails, and the skills, templates and criteria you set up. Ask about export formats, notice periods, help with the move, and what happens to your data if the vendor is acquired or closes. If those answers are weak, the parts you most want to own are the parts you would be locking in.

When does building in-house make sense?

Building is a reasonable choice when most of these are true:

  • You have a standing engineering team with AI experience, and it has spare capacity for years, not months.
  • The use case is specific to your firm and nothing on the market comes close.
  • Your data or deployment constraints rule out outside platforms, and you have tested that with vendors rather than assumed it.
  • You are prepared to own security, governance and support.
  • You expect the tool to give your firm an edge that competitors cannot buy.
  • Your data is already clean and joined up in a warehouse your team runs.
  • The workflow is narrow and high-volume enough that per-user pricing costs more than running it yourself.

Often the best answer is both. Buy a platform for the common work, such as reading data rooms, abstracting leases and drafting reports, and put your own effort into what is unique to you: your criteria, templates and house formats. Before you commit, check how you can extend the platform and what you can export, so what you build stays usable if you change vendors.

Build vs buy scorecard

For each question, note which way your firm leans. Do it with the person who will own the tool, an analyst who would use it, and IT.

  • How common is the workflow?

    Leans build if
    It is specific to our firm
    Leans buy if
    It is common across CRE firms
  • Who will maintain it for years?

    Leans build if
    We have engineers with spare capacity
    Leans buy if
    We do not, or they are needed elsewhere
  • Which costs less over three years on our worksheet?

    Leans build if
    Our build total, including the standing team
    Leans buy if
    The vendor's full cost, including renewal and exit
  • Did the option reach our accuracy bar on our own test set?

    Leans build if
    No vendor passes it, and our build spike does
    Leans buy if
    A vendor passes it, and our build spike does not
  • Can every figure be traced to its source?

    Leans build if
    No vendor shows it, and our build spike does
    Leans buy if
    A vendor shows it in a trial
  • Who enforces access to deal and portfolio data?

    Leans build if
    We can enforce our access model in our own build
    Leans buy if
    A vendor enforces it, tested in a trial
  • Do the connections to our systems need fragile custom work?

    Leans build if
    No vendor connects to the systems we run
    Leans buy if
    A vendor already connects to them
  • Could a general AI tool we already license do it?

    Leans build if
    No, and no specialist fits either
    Leans buy if
    No, but a specialist does; if yes, extend what you have
  • How fast do we need results?

    Leans build if
    We can wait months for a first live workflow
    Leans buy if
    We need one live in weeks
  • Can we leave if we need to?

    Leans build if
    We cannot get acceptable export and exit terms
    Leans buy if
    We can
  • Will the vendor still exist in five years?

    Leans build if
    We have doubts about the vendors on our list
    Leans buy if
    We are confident, and the contract covers the risk
  • Who answers the phone when it breaks?

    Leans build if
    Our own team, and we are staffed for it
    Leans buy if
    The vendor, as part of the price

The rows are not equal. Exit terms, accuracy on your own test set and hard data constraints can rule an option out on their own; speed is a preference. If neither a vendor nor your build spike passes your test set, the workflow may not be ready for AI yet. Score each workflow separately, because deal screening, lease abstraction and asset reporting often come out differently. A split usually points to buying the common work and building the rest.

Where Fifth Dimension fits

Fifth Dimension is AI for commercial real estate, from underwriting to lease abstraction. It is built to sit across the systems you already run, not replace them. Here is what Fifth Dimension publishes; use the build stages above as your demo checklist, for us and for any other vendor:

Connections
It connects the documents, data, systems and expertise a firm already relies on. That includes Excel, SharePoint, Argus, Dealpath and more, and for asset teams it unifies Yardi, Argus, Excel and email. Workspaces are synced with your underlying data rooms.
Sources
Evidence & Controls keeps sources, assumptions, review steps and changes visible, so your team can understand how an output was produced.
Security
Fifth Dimension is SOC 2 Type II compliant, ISO 27001 certified and GDPR compliant. Each customer operates in a separate product environment, and customer data is never used to train the models that power Fifth Dimension. Every operation is logged and traceable. SAML SSO and role-based access, IP allow-listing, data residency and on-premise or air-gapped deployment are available. Enterprise adds security and deployment options for your requirements.
Automation
Tasks schedule recurring automations and repeatable workflows, with review and control built in, and the platform includes the ability to build custom skills around your processes and outputs.

Fifth Dimension takes on the work between receiving information and making a decision, while keeping your team in control of the judgement that matters. Ask any vendor, us included, what you can take with you if you leave.

For the documents your security team will ask for, see Fifth Dimension security and the Trust Center.

Together, we identify the manual processes that are slowing you down across the full lifecycle of how you invest in and manage your assets. Start with the people and work that matter most, then expand as Fifth Dimension connects more of your organisation. Fifth Dimension includes high-quality service to help your team identify valuable workflows, configure the platform around how your firm works and keep improving what you have built. Request a demo, and bring the workflow, not the files, until IT and risk have agreed what can be shared.

FREQUENTLY ASKED QUESTIONS.

Should we build our own AI for commercial real estate?
Build only if the AI is something you want to own and you can staff it for years. For common work such as reading data rooms, abstracting leases and drafting reports, many firms are better served by a platform, with their own criteria and templates added on top.
How long does it take to build an AI tool in-house?
A prototype can take weeks. A tool your team can rely on takes longer, because it needs document intake, system connections, permissions, source citations, security review and rollout. Estimate it for your own scope with the worksheet in this guide.
What does it cost to build and maintain an in-house AI tool?
The cost has parts: the build team, model usage, infrastructure, integrations, security work, rollout and upkeep. Put your own number on each line and compare the three-year total with a vendor quote.
What are the risks of building your own AI tool?
You may spend time and money on a tool that does little more than a general-purpose assistant. Your team may have to check every output if it cannot show sources. You also take on security, governance and support, and you depend on a few engineers.
Is buying AI like buying SharePoint instead of building our own document storage?
Partly. Document reading, retrieval, permissions and review trails are plumbing for most firms, and a vendor can spread that work across many customers; ask which parts it maintains. The difference is exit: check that your data, outputs and configuration can leave with you. Your criteria, templates and judgement are the parts worth owning.
Can we use a platform and still customise it?
Yes, if the platform allows it. Fifth Dimension includes the ability to build custom skills around your processes and outputs, and high-quality service to help your team identify valuable workflows and configure the platform around how your firm works.

Tools & libraries.