A hospitality concierge was quoting bespoke packages for high-net-worth clients. Every quote took two or three days. Every quote was brand-critical. If it didn’t sound like him, the client noticed.
Most AI tools that try to solve this problem fail at the second part.
The first part, pulling supplier prices, applying the firm’s margin, generating a document, is straightforward. You can do it with off-the-shelf tools and a weekend. The quote comes out fast. The quote is accurate. The quote does not sound like the person who runs the firm.
Tone is not a styling problem. You cannot solve it with a prompt that says “write in a warm, premium voice”. Every AI-generated proposal you have ever received had a prompt like that behind it. They all sound the same, because the training data is the same.
The way to actually solve tone is not to prompt it. It is to extract it.
We took forty-odd quotes the founder had written by hand over the previous two years. We parsed them. We extracted the sentence patterns he used, the words he avoided, the length of his paragraphs, the structure of his openers, the way he signed off. We built a style layer specific to him, not specific to “luxury hospitality” as a category.
The quote engine generates the price structure. The style layer rewrites the output in his register before it ever reaches him for review.
The round of review shrunk from “rewrite most of it” to “check one or two numbers and send”. The turnaround dropped from days to hours.
If I had to give one piece of advice to anyone trying to automate a client-facing document workflow, it would be this: treat tone as a data problem, not a prompt problem. Find the existing body of work. Extract the patterns. Teach the system what your firm actually sounds like and stop relying on generic descriptions.
Everything else is straightforward.