Real estate data is ready for AI when deals, listings, LOIs and leases are stored as structured fields in one system of record that every person on the team and every AI tool reads from. That system is usually a CRM. Until it’s in place, tools like Claude can work well for one broker and fall apart once a whole team depends on them.
We saw this with a commercial real estate team that had good results from Claude, one person at a time, and asked us to scale it to the whole group. Their deal data lived in folders, inboxes, and spreadsheets. A team setup built on that would have been slow, heavy on tokens, and inconsistent from one day to the next, so we advised them to fix the data first.
Key Takeaways:
Personal AI use cases work well on messy data because the broker’s memory acts as the system of record instead of a CRM. A person can tell when an answer looks wrong, point AI to where to search, and regenerate results if needed with minimal resources spent.
At the team scale, those safeguards disappear. A colleague who doesn’t know the folder conventions or the deal history can’t guide and correct AI before the output reaches a client. Regenerating results is no longer cheap either, because every user’s repeated searches add to the token bill. And the answer has to be stored somewhere the rest of the team can find it, with two people asking the same question getting the same result.
| Where context comes from | The broker's own memory of folders, clients and deal history | Stored data the AI can read, since a team does not share one person's memory |
| Who checks the answer | The person who asked and knows the deal | Often someone with less context, who is less likely to spot an error |
| Where the output goes | One person's chat history or a file on their laptop | A shared record that others find and build on |
| Consistency required | Low: the user can rephrase or re-run the question | High: two brokers asking the same question need the same answer |
| Access to data | Whatever files the person can open | Defined by role, with some records restricted |
For example, a broker asks Claude, “What rent did we propose on the last LOI for Building X?” The broker knows which folder holds this year’s proposals and recognizes the right file. An analyst who asks the same question the next day doesn’t have that knowledge, so Claude searches the wider drive, finds three similarly named versions of the LOI, and returns a figure from an earlier draft. Nothing in the folder marks which version is current, so the AI has no reliable way to choose.
“ Engaging in personal Claude sessions was benefiting Bob, Sam, and Tony, but it wasn't benefiting them as a group. But they work as a group. So we have to think about solutions that are supportive of the group, not the individual. ”
In a system of record, each LOI is a single record with a status field and its key terms stored as data. That gives the broker, the analyst, and Claude the same current answer, and it lets a team build on what one person’s AI setup started.
After implementing AI in our internal operations and for our various clients, the Ascendix team has outlined the evaluation framework: speed, cost, trust, durability, and user experience.
Of course, AI can handle large volumes of data stored in a few different systems, but this framework helps to recognize whether this approach is effective.
“ Just because AI can doesn't mean it should. ”
Claude answers a question about files in a shared drive or inbox in four steps. It searches for likely files, loads them into its context window, reads them in full, and reasons over the text. Nothing is kept between questions or between users, so the same documents are read again for every request, and AI providers bill for the tokens each of those reads uses.
Results can also differ from run to run because Claude may pick different files or passages each time, and language model output isn’t fully deterministic, so it must be verified. The setup might break completely if the folder is moved or renamed, a new file format is introduced, or there are older versions, duplicated files, or files with the same name.
The solution is to have a system of record.
| Speed(How long does the answer take?) | Whole documents and email chains read per question | Stored fields queried directly | Slower replies to inquiries; brokers go back to searching by hand |
| Cost(How many tokens does it consume?) | Spend grows with each user and query | Each document read once | AI budget outpaces value |
| Trust (Can team rely on the results?) | Results vary; old drafts slip in | Same answer, traceable source | Wrong figures reach clients, and the team stops relying on AI |
| Durability(Can it survive daily updates?) | Breaks when folders or files change | Stable records and fields | Recurring maintenance work and downtime |
| User experience(How much effort per final output?) | Going through files, inbox, prompts, and chats to verify | Ask once, get a traceable answer | Low user adoption |
A system of record is the single authoritative place where a firm stores its core data as structured records: properties, listings, deals, contacts, and commissions, each with defined fields and links to the others. In commercial real estate, this is usually a CRM. Spreadsheets and shared folders store information without fixed fields or links between records, which is why software cannot rely on them for consistent answers.
Once Claude is connected to a system of record, it answers questions by querying fields.
“ If we've got a letter of intent that has five data points in it, instead of having Claude go in there and look through that entire document to find those five data points, we would in advance have indexed that document with those five fields already stored in a system of record and point Claude to that. ”
Slatt Capital, a commercial mortgage banking firm (case study link as soon as it’s published), shows what the system of record column looks like in practice. With its deal and lender data already in Salesforce, Ascendix connected Claude through MCP with an audit log for every AI-made change, so producers now get instant answers that used to take up to a week, and data entry that took eight hours now takes one
Also, brokerages inside larger organizations often face one more constraint. An AI license managed at the organization level usually requires approval from outside the team before a custom app can connect to it. Teams in this position should request that approval before starting a build.
Book a free AI consultation with Ascendix to identify what a team-wide AI setup would need.
An AI data strategy for real estate usually includes 5 steps to build a solid data foundation:
1. List the questions and the sources. Write down the 10 to 20 questions the team asks most often and note where the answer to each one lives today: shared drive, inbox, spreadsheet, or CRM. This list later becomes the test set for the five criteria above.
2. Define the record types and fields. Each field should answer at least one question on the list. The table below shows a typical starting point for a brokerage.
| Listings | Property, size, availability, agreement expiration | Which listings expire in the next 30 days? |
| Deals | Stage, parties, assigned broker, key dates | Which deals moved to LOI stage this month? |
| LOIs and proposals | Property, tenant, size, proposed rent, term, date sent | What rent did we propose on the last three LOIs for Building X? |
| Leases | Parties, commencement, expiration, rent, commission terms | Which executed leases still need a commission voucher? |
| Contacts and companies | Role, relationships, last activity | Which tenant brokers have sent inquiries on this listing? |
| Inquiries | Source, size requirement, submarket, date received | How many inquiries came in last month, by size range? |
3. Abstract the important documents. Start with active listings, open deals, and recent LOIs and leases, and add older archives once you create the structure.
Once each record type has its key terms stored as fields, the team has AI-ready real estate data.
4. Capture new data at the source. Documents that arrive by email are the hardest to capture because they sit as attachments in individual inboxes. Ascendix built the Harvest AI email parser for this step: a user forwards an email, and Harvest abstracts the email body and the attachment data into the system of record.
5. Connect AI and measure. Connect Claude to the system of record through an MCP connector. CRM permissions decide which records each user’s AI session can see, and an audit log records changes made by AI. Re-run the questions from step 1 and compare the results against the five criteria above.
The nine questions below help a team check whether it has AI-ready real estate data before starting an AI build. Each row shows what a ready answer looks like and what indicates that data work should come first.
| Where does your deal data live? | Deals, listings and contacts are records in a CRM | Deal details sit in inboxes, desktops and a shared drive |
| Is there one current version of each document? | Each LOI or lease has one record with a clear status | Several files with names like "final_v3" exist for the same document |
| Are key document terms stored as fields? | Rent, term, size and expiration dates can be filtered without opening a file | The terms exist only inside PDFs and Word documents |
| Do records link to each other? | An LOI is connected to its property, tenant and deal | Each document stands alone, and the links exist only in people's memory |
| Does AI give the same answer twice? | The same question on two different days returns the same answer | Answers change between runs |
| Can each answer be traced to a source? | Claude's answer points to a specific record | Nobody can tell which file an answer came from |
| How does new data get into the system? | Emailed documents are captured into records as they arrive | New documents stay as attachments in individual inboxes |
Three or more answers in the right-hand column usually mean the data foundation is the place to start.
Ascendix team offers a free AI consulting call to help you understand your needs, estimate your AI readiness, and build software that fits CRE workflows.
Ascendix has worked on CRM and data systems for real estate firms since 1996. We are a Summit Salesforce Partner and have our own CRE-focused AI-powered CRM, AscendixRE.
Teams that already use Claude individually and want a setup the whole group can rely on can start with a free AI consulting call.
One person can use AI on files without a CRM. A team needs a system of record, usually a CRM, where deals, listings, and client information are stored as structured fields. Every user and every AI tool then works from the same source, which keeps answers consistent and cheaper to produce at a team scale.
Yes, for individual work, Claude can search spreadsheets and shared-drive folders and return useful answers. At team scale the approach slows down, uses more tokens and gives inconsistent results, because Claude has to re-read raw files on every request.
Large language models are non-deterministic. Asked to search unstructured files, Claude decides each time which documents to open and which passages matter, so results can shift between runs. Connecting Claude to structured fields, such as the key terms of an LOI stored in a CRM, limits what it has to choose between and makes answers repeatable.
An AI data strategy for real estate is a plan for where deal, property, document, and client data is stored and how AI tools access it. It typically covers the system of record, the fields abstracted from documents, how new data is captured, who controls AI access, and how AI-made changes are logged.
Start with the data foundation for AI: audit where deal and document data lives, move it into a shared system of record, and abstract key document terms into fields. Then connect Claude to that system and share skills across the team. Confirm who controls the AI environment and log every change AI makes to a record.
No. Most teams start with the records they use every day: active listings, open deals, and recent LOIs and leases. Older archives can be abstracted later, once the record types and fields are set. New documents should be captured as they arrive, so the volume of unstructured files does not grow again.
It can be safe with controls in place. CRM permissions should limit which records each user’s AI session can see and change, and an audit log should record whether a person or the AI made each change. A common approach is to start with read-only access and extend write access to a small pilot group once it proves to be reliable.
Kateryna is an experienced writer focused on topics related to Salesforce consulting and optimizing operations within the real estate industry via CRM implementation. Through her insightful articles, Kateryna aims to help organizations make the most of their Salesforce adoption.
(88 votes, average: 4.99 out of 5)
Loading...