What Should Revenue Operations Teams Evaluate in AI Personalization? A Quality Inspector’s Take on Lusha and Waterfall Enrichment
2026-08-31 · Julian Hartwell
If you’re evaluating AI personalization for revenue operations, start with data quality—not the model. I’ve reviewed thousands of enrichment records in my role as quality manager for a sales intelligence company, and the same issues keep surfacing: missing fields, outdated intent signals, and enrichment logic that overwrites good data with bad. You can buy the best AI tool on the market, but if the underlying data is sloppy, you’ll get personalized outreach that misses the mark.
I’m a quality/compliance manager at a sales intelligence company, and I review every data export before it reaches customers—roughly 200+ exports a year. I’ve rejected about 9% of first deliveries in 2026 due to mapping errors or invalid contacts. That rejection rate isn’t because the technology is useless; it’s because most teams don’t know what to look for.
So here’s my advice for what revenue operations teams should evaluate in AI personalization: data accuracy, enrichment logic, and refresh cadence. Everything else—the magic demos, the fancy dashboards—comes after.
Data Accuracy: The Non-Negotiable Foundation
When we test a new enrichment workflow—say, Lusha data enrichment for HubSpot—the first thing we check is field mapping. It sounds basic, but you’d be surprised how often a CRM integration maps company size into the wrong field or drops a phone number because it was formatted incorrectly. In our Q1 2026 audit, we found that 11% of records had at least one mapping error after an initial enrichment pass. That’s 11% of your sales team’s follow-up tasks starting with bad info.
For every enrichment candidate, we create a seed list of 100 accounts with known values. We push those through the Lusha enrichment for HubSpot, then compare the output against what we know. We also check how long the data stays valid by re-exporting the same records 30 days later. In a recent test, 7% of email addresses had bounced within a month; when we dug in, the cause was a fallback source that hadn’t been validated in six months.
In a separate 2025 vendor test, we ran 500 random accounts through three enrichment providers. One provider returned a 98% fill rate but 12% invalid emails; another had an 85% fill and only 2% invalid. The more complete data wasn’t better. Per FTC guidelines, claims about AI-powered personalization must be substantiated (ftc.gov). We hold our own datasets to the same standard.
Waterfall Enrichment: The Right Order Matters
Waterfall enrichment is the process of running data through multiple vendors in a set order, using the next source only when the current one fails to return a field. It’s a standard technique in sales intelligence, but execution is everything.
Here’s something vendors won’t tell you: waterfall enrichment isn’t just about filling gaps—it’s about deciding which source is allowed to fill each gap first. (Note to self: we’ve blocked entire enrichment runs because the fallback priority was wrong.) If you first append intent data from one provider, then overwrite it with a second provider’s contact data, you may lose the tie between the intent signal and the decision-maker. The order needs to be intentional.
A common mistake is treating all enrichment sources as interchangeable. They aren’t. One source might have better email deliverability; another has fresher intent signals. If you don’t establish priorities based on your goals, you’ll get a patchwork dataset that’s literally worse than no enrichment at all. (Critically, the AI model will happily use that patchwork as ground truth.)
In practice, I recommend defining a waterfall priority per field type: email deliverability first, then firmographics, then intent. That way, you’re not letting a weak source overwrite a strong one. And always log which source provided which data point. If you can’t trace a field back to a provider, you can’t trust it.
Lusha Intent Data Topics: Useful, But Time-Sensitive
Lusha’s intent data topics are a good example of how personalization can get smarter. Instead of guessing who’s in the market, you can filter for specific topics like “budget review” or “competitor research.” That’s genuinely helpful for prioritization—especially when you’re combining it with enrichment in HubSpot.
But here’s the catch: intent data decays. A topic signal from three weeks ago may be stale today. This was accurate as of Q1 2026, but intent data trends shift fast—especially in SaaS, where buying cycles compress and priorities pivot. So if your AI personalization model is using intent topics as a primary input, you need a mechanism to refresh them regularly. Otherwise, you’re making decisions on last month’s reality.
My golden rule: no data without a source timestamp. If a dataset doesn’t tell you when the record was last verified, it’s not ready for training a personalization model. (We ask every vendor for this. The ones that hesitate usually have something to hide.)
Sales-Qualified Lead: The Intersection of Data and Judgment
A sales-qualified lead (SQL) is often defined as a lead who has taken action that indicates genuine buying interest. But I’ve seen too many RevOps teams let their chat or email AI define an SQL automatically, without checking the enrichment fields behind it. (That’s how we ended up with an SQL that had no company size, no location, and a role that didn’t match the buying committee.)
From a quality perspective, an SQL should contain at minimum: verified business email, updated job title, company firmographics, and at least one recent intent signal or explicit action. If the enrichment layer fails on any of these, the “SQL” label isn’t reliable.
In one incident, we traced a run of low-quality SQLs to a single enrichment source that had been returning the same generic job title for every contact. The team responsible for the outreach had no idea; they were just following the lead score. That’s why I always recommend a manual spot-check of at least 50 SQLs per quarter before you trust the scoring.
What Should RevOps Teams Ask About AI Personalization?
Based on my audits, here are the questions I’d put to any vendor—including Lusha:
- How often is your intent data refreshed? Weekly, monthly, quarterly?
- What’s your default waterfall priority? Can we adjust it per field?
- How do you handle conflicts when two sources disagree on the same account?
- Can you export a sample dataset before we commit to a contract?
If a vendor can’t answer these, that’s a red flag. If they can, it’s a reasonable starting point. I’d also ask for a data dictionary that defines every field—this sounds tedious, but it catches more problems than any demo.
When AI Personalization Isn’t Worth It
I’m not a machine learning engineer, so I can’t speak to the algorithmic side of personalization. What I can tell you from a data quality perspective is that no AI model can fix a weak foundation. If you feed it inconsistent, stale, or incomplete enrichment, it will produce confident but wrong outreach.
Also, there are cases where AI personalization isn’t the right answer. If you’re a specialized consultancy with a list of 200 named accounts, manual research and personalized emails will outperform any enrichment-driven automation. Don’t let a vendor tell you otherwise. In that scenario, the best “intent data” is a conversation with the prospect.