Situation
A real estate agency in Vilnius has about 35 employees: brokers, a few people in the back office, a lawyer and an accountant. Brokers work in the field, at properties, with clients, at notaries, and for most of the day they need a quick answer: which contract template to use for a lease with a legal entity, how commission is calculated when two brokers sold the same property, what to do when a client asks for a reservation without a deposit.
All of this information existed in the company: procedures, commission rules, templates and checklists kept as Word and PDF documents on a shared drive, plus a few instructions buried in emails. The problem was that nobody searched for them. It was faster to message the administrator. The back office staff estimated that they received several dozen such messages a day, and most of the time the answer was a link to a document the broker could have found on their own. Some answers depended on who replied: two administrators sometimes explained the same thing differently because they were working from different versions of a template.
The second pain point was new brokers. Onboarding took about six weeks, because a person had to read documents they could not understand without context and then keep asking colleagues questions for the first few months. The managing director wanted a newcomer to be able to ask at any time and get the same answer the back office would give.
What was needed
- One assistant that answers questions about company procedures, templates and commissions using only company documents.
- Every answer with a link to the document and its version, so the broker can verify it.
- Access rules: commission and compensation documents are not visible to everyone, and personal data from contracts does not enter the knowledge base.
- Company data is not used to train models.
- Deployment within a month, without an IT department, which the company does not have.
What we did
Week 1: document inventory
We started with the shared drive. Out of roughly a hundred documents we selected about forty that were current: old template versions were removed, duplicate instructions were merged, and every document was assigned an owner and a review date. Without this step the assistant would have answered from outdated documents, which is worse than not answering.
Week 2: workspace and assistant
We set up a ChatGPT workspace for the company under business terms: data stays under the company’s control and is not used to train models. In the workspace we created an assistant with the company’s documents and one instruction: answer only from the attached documents, always cite the source, and if the answer is not in the documents, say so plainly and refer the person to the back office. This is the standard principle in our ChatGPT assistant deployments: no guessing, every answer either comes with a source or is “I don’t know”.
We solved access with two assistants: a general one for all brokers with procedures and templates, and a separate one for managers with commission and compensation rules. Access is managed through workspace groups linked to the company’s Microsoft 365 accounts. We did not put signed contracts into the knowledge base, only templates and procedures, so client personal data is not available to the assistant.
Weeks 3-4: trial and fixes
For two weeks the assistant was used by five brokers and the back office. We collected the questions the assistant answered with “not in the documents”, which turned out to be the best list of what was missing from the procedures. The back office extended four procedures and rewrote two. The hardest part was the commission calculation rules with their exceptions: we rewrote them as a table with examples, because from a long text the assistant gave ambiguous answers. We also checked that the assistant does not answer questions about specific clients. It has no such data and redirects those questions to the deal manager.
Result
Questions to the back office dropped by about 55%. We measure this by the number of messages in the shared back office channel per week, comparing three months before and after. The remaining questions are mostly about a specific deal rather than a procedure, and these are exactly the questions the back office is there for. Once a month the back office reviews the assistant’s question log and updates the procedures accordingly, which takes about an hour.
Onboarding a new broker shortened to about two weeks. The newcomer gets the assistant on day one and asks it instead of colleagues; the managing director sees which questions they ask and plans conversations accordingly.
Answers without a link to the source: zero, because that is how the assistant is configured. When there is no source, the assistant does not answer but refers the person to a human. Brokers say this is what allowed them to trust it: every answer can be checked with one click.
We did not expect the biggest benefit to be not the assistant, but that we finally got our procedures in order.
What we would advise in a similar situation
- Organise your documents first. The assistant will not hide that you have three versions of a template. It will show it on day one.
- Set a boundary for the assistant: answer only from attached documents and always cite the source. Without this boundary it will be convincing but not reliable.
- Assign access by document sensitivity from the start, not after someone sees the compensation table.
- Collect the questions the assistant could not answer. It is the cheapest way to find out which procedures your company is missing.
