How a title company put its own voice on the phone
An ElevenLabs-powered voice workflow combined cloning, context-aware RAG retrieval, and React conversation management for natural, personalized AI responses.
Every closing generates the same handful of questions, asked by different people at different times. Has the earnest money arrived. Are we clear to close. What is still outstanding. The answers are simple, but someone has to stop working to give them, and that someone is usually the person handling the file.
This title company wanted those calls answered without pulling staff off production, and without the caller feeling handed to a machine. We built a conversational AI agent that speaks in a cloned copy of a real voice from the business, answers from the company’s own material, and escalates anything that needs judgement.
Challenge
Routine calls interrupted skilled work
Answering a status question is not difficult work, but it is disruptive work. A closer stops mid-file, opens the production system, finds the file, checks which milestones are complete, relays the status, then goes back and rebuilds their concentration. Multiply that by every buyer, agent, and lender on every open file.The same answers, repeated
Most of the questions were not unique. Several parties to the same transaction would call separately across the same week, and each one received the same information assembled by hand. Industry commentary describes this as a status update tax: when parties cannot see the file, they call, and production staff become a call centre.
Calls outside office hours went unanswered
Buyers and agents do not only think about their closing between nine and five. Calls that arrived in the evening or at the weekend reached voicemail and were returned the following working day, which turned a thirty second answer into an overnight wait.Solution
The company adopted a conversational AI agent built around two decisions: it should sound like them, and it should never invent an answer.
The voice belongs to the business
Rather than a stock assistant voice, we cloned a real voice from the company. Callers hear something that sounds like the firm they are dealing with. A separate set of persona controls governs tone, manner, and the boundaries of what the agent will discuss, so behaviour can be adjusted without rebuilding anything.
Answers are retrieved, not generated
The agent is connected to the company’s own knowledge base. Each answer is grounded in retrieved source material rather than produced from a general model’s memory. In title work this distinction matters more than it does in most industries, because a confidently invented answer about somebody’s closing is worse than no answer at all.It knows where to stop
The agent answers what the record supports and declines what it does not. When a caller asks something that depends on information the file does not yet hold, it says so, offers a callback from the responsible person, and captures the number. Play the call below to see how that plays out.
Show full transcript
Hi, I’m calling about the closing on Maple Street. Do you have everything you need from me?
Let me check. I have the file. We’ve received your signed disclosures and the earnest money. We’re still waiting on the payoff statement from your lender.
How long does that usually take?
Payoff statements typically come back within three to five business days of the request. Ours went out on Tuesday, so we’d expect it by the end of this week.
And can you tell me exactly what my final cash to close will be?
I can’t confirm a final figure until the payoff is in. I’ll have your escrow officer call you as soon as we have it. Is this the best number to reach you?
An example call showing how the agent answers and where it hands over to a person.
What changed
Routine calls answered without interrupting staff
Closers stay on the file. The agent handles the questions that only needed a lookup.
Available outside office hours
Evening and weekend callers get an answer when they ask, instead of a voicemail and a next day callback.
Escalates instead of guessing
Anything the record does not support goes to a named person, with the callback number captured. 
Related case studies
More stories from teams using XtractSol.
Case study · Title insurance
Intelligent Deed & Mortgage Matching
A name search returns every deed attached to a name. We score each one against the subject parcel on address, parcel number, and owner, so examiners see the records that matter first.
Read moreCase study · Title insurance
AI-Powered Source-to-Decision Traceability
Extracted fields are fast to work with, but a decision has to be defended against the original. We keep the source document, the search context, and a findable reference beside every value.
Read moreCase study · Real estate compliance
Automating FinCEN Compliance
We read a purchase contract, work through the FinCEN reporting tests, and return a clear reportable-or-not decision with the exact contract wording behind it.
Read more
Ready to transform your workflow?
Book a 30-minute discovery call with our engineering team to map how custom AI agents can automate your title and escrow operations.