AI server for property management companies: tenant mail, leases, defect reports in-house

What an AI server does for a property management company

An AI server for a property management company runs language and image models on hardware the firm controls, so tenant correspondence, lease contracts and owners' documents stay inside the company network. It drafts replies to tenant emails for staff to check, finds clauses across thousands of leases, prepares the documents for the annual service-charge statement, turns defect reports with photos into work orders and summarises owners' meetings into minutes. For a firm that manages a few thousand units, one server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries all of these tasks for the whole office.

This guide is for the managing director or owner of a property management firm, and for the person or service provider who looks after its IT. A few terms come up throughout. A model is the software that reads and writes the text; gpt-oss-120b from OpenAI, for example, is published under the Apache 2.0 licence and can run on your own hardware. A GPU, or graphics card, is the part of the server that does the computing for AI; its memory decides which models fit. A token is a word or part of a word, the unit in which models read and write. The legal notes below are general information on EU law as of October 2026, not legal advice; your data-protection officer or legal counsel confirms how they apply to your firm.

Tenant emails, leases and the other daily work it takes on

Property management is correspondence and documents. Five tasks make up most of what an AI server does in such a firm, and none of them replaces the person who signs the letter.

Our summary of the tasks; loads by our estimates below.

Tenant mail. Tenants write about heating, noise, rent payments, keys, moves in and out. The model reads the message, sorts it by topic and building, and drafts a reply from the tenant's file and the firm's standard texts. A staff member reads the draft, corrects it and sends it. The model does not send anything on its own. Many of these replies need only a modest model: answering when the heating is serviced does not call for a 117-billion-parameter model. The point of one company server is that the same machine also carries the heavier tasks below, so the firm buys and looks after one machine instead of a tool per task.

Lease contracts. A firm that has taken over buildings from other managers holds leases in many versions. Asked "which tenants in building 12 have a clause on garden maintenance", the model searches the scanned and typed leases and answers with the clause and the document it comes from. This is answering from your own documents, which our guide to RAG on company data explains. Our guide to preparing company documents for an AI assistant covers the scanning and sorting beforehand.

Service-charge documents. Once a year, the invoices for cleaning, caretaking, insurance, water and waste per building go into the statement for owners and tenants. The server reads each invoice into a table with supplier, date, period, amount and cost type, ready for the accounting software. The totals and the split per unit stay in that software: the model reads documents, and the sums come from code.

Defect reports. A tenant reports a leak and sends three photos. A model that reads images as well as text describes what it sees, suggests the trade (plumber, roofer, electrician) and drafts a work order with the address and access notes. Staff check the suggestion before the order goes out.

Owners' meetings. A recorded meeting is transcribed and summarised into draft minutes with the motions and votes; our guide to meeting minutes with a local AI covers the steps and the consent questions.

Tenant data and the GDPR in property management

Nearly everything a property manager handles is personal data: names, addresses, bank details, rent arrears, complaints about neighbours, and sometimes illness or family matters that a tenant mentions in a letter. Under Article 4(1) of the GDPR, personal data is "any information relating to an identified or identifiable natural person", and photos of a flat can be personal data too when they show people or identify the tenant.

The European Commission's summary of the GDPR principles, read on 25 October 2026, sets out what applies wherever the model runs. Data should be "adequate, relevant, and limited to what is necessary for the purpose", "stored for no longer than necessary for the purposes for which it was collected", and the firm "must implement appropriate technical and organisational measures to ensure the security of the personal data". The firm is also "responsible for demonstrating its compliance". An own server changes where the data sits, not these duties.

What it changes is practical. Prompts, drafts and photos stay on hardware under the firm's access rules; nothing goes to an external AI service unless the firm switches one on for a specific task. Access can follow the portfolio: the team for one set of buildings sees only those tenants. And deletion is under your control: when a tenancy ends and the retention period is over, the documents, the search index entries and the logs for that tenant can be deleted together. Our guide to local AI and the GDPR sets out what own hosting solves and what stays your duty, and whether you need a data protection impact assessment is a question for your data-protection officer.

Tenants also have rights over their data. The Commission's page on requests from individuals names "rights of access, rectification, erasure, portability", and the firm must reply "without undue delay, and in principle within 1 month of the receipt of the request", including "a copy of the personal data it holds about them". In a firm where a tenant's history is spread over mailboxes, scans and the property software, a search across all of it on your own server finds the documents in minutes; a person still decides what goes into the reply.

Our comparison of ChatGPT Enterprise and an own AI server sets out what a cloud provider's terms state, if you are weighing a subscription instead. Law firms and audit firms face similar questions under their own secrecy rules, as our guides for law firms and tax advisers and for audit and accounting firms show.

We supply the server and set up the models, the document search per portfolio and the query log through our Private AI/ML service; send us the number of units you manage and the tasks you would start with, and our team replies with a configuration.

How much work that is for the server

The figures below use example values for a firm that manages 3,000 units with about 60 staff. Replace them with your own figures; the method stays the same.

Example values (3,000 units, 2 tenant messages per unit a month, 21 working days, 150 buildings with 400 invoice pages each, 60 staff); replies from MLPerf® Inference: Datacenter v6.1, closed division, available category, gpt-oss-120b, offline scenario, entry 6.1-0050 (HPE ProLiant Compute DL380a Gen12, eight RTX PRO 6000 Blackwell Server Edition), 15,256 tokens per second, retrieved from mlcommons.org on 24 October 2026, result verified by MLCommons Association; 1,907 per card is our division of the system result, not an MLPerf metric, and we plan with half of it; pages at about 0.69 per second per card, the estimate of our document-AI guide; conversations by our users-per-card rule. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Tenant replies. 3,000 units with two messages per unit a month give 6,000 messages, about 290 per working day. A reply of about 300 tokens each makes about 87,000 tokens a day. At 953 tokens a second, half of our per-card division of the MLPerf result, that is about 91 seconds of one card, by our arithmetic. Reading the incoming message and the tenant file adds time, but the load stays small. This is why one card for mail is far more than enough, and why the same card can do other work.

Scanned documents. Pages are read by an OCR model (optical character recognition: software that turns a scanned page into text and tables). Our guide to vision-language models for document AI estimates about 0.69 pages a second for a 5B OCR model on one RTX PRO 6000 Server Edition and compares the OCR models and their licences. 45,000 lease pages then take about 18 hours on one card; 60,000 invoice pages about a full day and night, or one night of 12 hours on both cards after the office closes.

Staff questions. Our company-size guide works with example values: 40 per cent of staff use the assistant in the busiest hour, each sends 6 requests, an answer takes 30 seconds, and the busiest minutes bring twice the average. For 60 staff that is about 3 requests at the same moment. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card and less the model, gpt-oss-120b leaves 22.2 GiB on one card, room for about 19 conversations of 32K tokens.

Which server, how many staff, what it needs in the building

This is the hardware section, for you and for whoever handles your IT.

Reference configurations on our AI servers page and our small-company guide; conversations by our users-per-card rule (about 93 of 32K on a pair); power at the wall by our estimate: cards (600 W Server Edition, 300 W Max-Q), processors and 15 per cent for the rest; a 16 A circuit at 230 V carries 2.94 kW at 80 per cent.

Two cards in a rack server. The entry configuration on our AI servers page has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. Card 1 runs gpt-oss-120b, which has 117 billion parameters with 5.1 billion active per token and fits one 96 GB card. Card 2 runs Gemma 4 31B from Google, published under Apache 2.0, which takes "Text, Image" as input and supports more than 35 languages according to its model card; its checkpoint of 62.6 GB, about 58.3 GiB by our arithmetic, leaves about 24.7 GiB on the card for its conversations and the small search models. At night card 2 switches to the OCR model, and card 1 joins it once nobody is writing. If a card fails, gpt-oss-120b can run on the other one until the replacement arrives, with the photo model paused. How many photos the image model handles at once depends on the image resolution it is set to; a pilot on your own reports gives the figure.

A workstation instead. Many property firms have no server room. A workstation with two RTX PRO 6000 Max-Q cards, NVIDIA's 300 W workstation edition, runs the same models, draws about 1.1 kW by our estimate and runs on an ordinary 230 V socket in a side room with a door. As of October 2026 we found no maker's noise declaration for such a tower under load, so it belongs in a side room, not next to desks; our guide to a GPU server in an office covers heat and noise, and our page on the two-card Max-Q workstation lists the parts. For a machine that runs continuously in a rack we recommend the Server Edition.

Four cards. A firm with a large portfolio, or one that wants the archive and the service-charge invoices read during office hours, takes four cards. Two cards then hold gpt-oss-120b as one copy with about 93 conversations of 32K tokens, and the OCR card works by day. Four Server Edition cards draw about 3.4 kW and need a 32 A or three-phase feed, as our guide to GPU server power on 16 A, 32 A and three-phase feeds explains.

Practically all the power becomes heat, so the room needs cooling whenever the machine runs. Without a suitable room, our Private AI/ML service runs the models on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

Notes for the IT team or service provider

1. Drafts, never auto-send. Connect the model to the mailbox so that it creates drafts in the staff member's folder. Sending stays a human action, and the query log records which draft the model wrote.
2. Access per portfolio. Map directory roles to buildings or owners' associations; every document chunk in the search index carries the building and its access list, and the search filters before it ranks. Put an LLM gateway with keys and a query log in front of the engines.
3. Untrusted mail. Tenant emails and attachments can contain text aimed at the model. Keep the mail model without tools and extract into fixed fields, as our guide to prompt injection and LLM security explains.
4. Retention and deletion. Set a retention per document type and delete documents, index entries and logs for a closed tenancy together, on a schedule your data-protection officer approves.
5. Network and versions. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere (vLLM 0.32.0 security guide), so keep the engines on an internal network behind the gateway. Run the NVIDIA Container Toolkit at 1.17.8 or later and a driver NVIDIA lists as fixed in its bulletin of 30 September 2026.

A firm without IT staff of its own needs one person who owns the assistant and decides which documents go in and who may use it; installation and updates can sit with a service provider, as our article on an AI server without an IT department describes. Our guide for companies of 10 to 100 people covers the same decision for small firms in general. If you would like to start with one task, describe the task and the volume, and we propose a pilot and the server for it.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build rack servers with two or four RTX PRO 6000 Server Edition cards and workstations with two RTX PRO 6000 Max-Q to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, deploys the models, the search per portfolio, the mailbox connection and the query log.

FAQ