My results after trying Grok's 'predictive lead scoring' - numbers inside.

So Grok finally got around to adding "predictive lead scoring" and, like a fool, I took it for a spin. The sales pitch was the usual: "AI-driven insights," "increase conversion," "maximize revenue." I was more interested in what it would cost to maximize *their* revenue.

I ran it on a sample dataset of about 50k leads from our marketing pipeline. The setup was deceptively simple, which is always a red flag. You feed it your CRM columns, it chews for a while, and spits out a "score" between 0 and 1 for each lead. Great. Now for the part they don't highlight: the operational cost to actually use this in production.

Here's the kicker from my billing analysis. To get "real-time" scores on new leads, you're hitting their inference endpoint. The pricing page talks in "units," but let's translate that to something real:

* **Batch scoring my initial 50k leads:** ~$12 (not terrible for a one-off).
* **Real-time scoring at ~100 leads/minute:** ~$180/month (starts to sting).
* **Add in the required data pipeline to keep their model 'fresh' with our CRM data:** That's another dedicated integration job, call it $50/month in compute elsewhere.
* **Vendor Lock-in Tax:** The model is a black box. You can't export it. Your scoring logic is now a permanent line item. Want to leave? Rebuild from scratch.

The accuracy? It was... fine. It identified our high-value leads about as well as the simple logistic regression model we built in-house two years ago. Which, incidentally, runs for about $22 a month on a single small instance and costs us nothing in egress fees.

So we have a marginal accuracy improvement at best, traded for a 10x cost multiplier and a dependency on Grok's API. They're not selling you intelligence; they're selling you a very specific kind of convenience that happens to come with a recurring meter that's always running.

-- cost first