Sell the data work as the AI project. Don't position it as a prerequisite you have to get past, because that framing invites the client to skip it. The deliverable is a scored pipeline, and phase one is making outcomes recordable, which is where most of the hours are.
Concretely, with 4,000 records and 60 percent blank dispositions, you need a closed outcome taxonomy, maybe eight values, mandatory on stage exit, plus backfill on whatever subset is recoverable from call logs and notes. That's real work with a real number attached and it produces something the client can see, source-level conversion reporting they've never had. Scoring rides on top afterward.
The distinction worth making in the sales conversation is between generative features and predictive ones. A model that drafts follow-up copy or transcribes and summarizes a call needs no historical data at all and works on day one. Predictive scoring needs a few hundred clean closed-won and closed-lost outcomes minimum to be worth anything, and in real estate the label lag is brutal because a seller lead can close eleven months out. Give clients the generative wins immediately and they stop comparing you to the competitor's dashboard.
On losing deals to the placebo: some you will lose, and the ones you lose are the accounts that would have churned at month four when the scores turned out to be random. What protects you is asking for read access to their data during discovery and putting one slide in the proposal showing the actual field completeness. That's not an argument, it's their own numbers.
Watch what you promise in writing. Language implying a lift in conversion is a performance claim you can't control, since it depends on whether their people work the leads at all.
