Rippling’s AI Bet: The Data Graph Is the Moat
Rippling’s whole AI bet comes down to one thing: a single connected database under every product it sells.
Luke Prokopiak, who leads Rippling’s AI product team, spent the first two minutes of his demo on that database before he touched a model. Rippling’s employee graph is the data layer beneath all 25+ products: payroll, HR, recruiting, performance management, benefits, IT, and spend. Rippling built every one of them from the ground up. No acquisitions, no bolted-on systems, no data sitting in disconnected silos that can’t talk to each other.
Most competitors acquired their way into a product suite, and the data underneath was patched together. Rippling’s argument is that AI can only do real work when it’s sitting on one clean, connected graph. With more than a million queryable fields covering everything from a person’s title and pay to how you handle taxes for a part-time hourly employee in Germany, the data alone isn’t enough. The harder problem is the layer on top: understanding the relationships between fields, enforcing who has permission to see what, and choosing which fields actually answer the question. That middle layer is where accuracy and trust come from.
Insights, Then Actions, Then Proactive Workflows
The demo moved through three stages, and the progression is the right product arc for any company building AI on top of business data.
Stage one: insights. A single prompt to build a company dashboard by department and tenure produced a shareable artifact in seconds to minutes. Employee count, number of managers, departments, work locations, average tenure, levels in use. The output is a durable artifact a team can pull up in weekly or monthly planning sessions.
The deeper version was a top-performer report. The prompt found 17 full-time individual contributors with the highest potential and overall ratings, then broke them down by department, level, and tenure. The insight that surfaced: 71% of top performers had six or more years of tenure. Rippling AI flagged the same pattern Luke noticed reading the data, which points at the obvious retention question. If your best people are your longest-tenured people, the job is getting more people to that threshold.
The natural follow-up was attrition risk. Another prompt found 9 full-time employees who hit the criteria: top performance, high potential rating, multiple years of tenure, no promotion, and more than one manager. Five in customer support, four in engineering, with each person’s current manager and manager count attached. That’s a list a leader can act on the same day.
Stage two: actions. This is where it stops being a reporting tool. Luke promoted an employee to IC7, staff software engineer, effective Monday, in plain language. Rippling AI resolved who the person was, mapped IC7 to the leveling system, identified the new title, built the promotion packet, and filled the right fields. Before executing anything sensitive, it required multiple confirmations and showed a before-and-after view. The reason it can do this safely is what Luke called strong types. The system understands every single data point in the database, so it knows exactly what it’s changing and who is allowed to change it.
Stage three: proactive workflows. The last prompt created a recurring workflow: a monthly high-performer growth review that runs once a month at 9 a.m. and emails the HR business partner. This is the shift from pull to push. You stop digging for insights and the system surfaces them across a million fields you’d never query by hand.
The Lessons for Founders Building on AI
These apply well beyond HR software.
The data layer is the moat, not the model. AI sitting on fragmented, acquired, disconnected data struggles with context and produces answers you can’t trust. A unified graph is what makes accurate answers and trusted actions possible. If you’re building AI into your product, the question isn’t which model you use. It’s whether your underlying data is clean and connected enough for AI to act on.
Permissions and strong typing aren’t glamorous, but they’re the difference between insight and liability. Knowing who can see each field, and knowing exactly what each field means, is what lets an agent take action without breaking something. Most AI demos skip this part. Production never can.
Confirmations are correct for the current trust phase. Putting multiple confirmation steps in front of any sensitive action, with a clear before-and-after, is the right call while teams are still getting comfortable letting AI act on their behalf. That comfort builds with each successful action.
And the product arc itself is the lesson: answer questions, then take actions with guardrails, then surface things proactively. Each stage builds the trust required for the next one.
From System of Record to System of Intelligence
Luke closed on the strategic framing, and it’s the line worth keeping. Rippling spent a decade being the system of record, the place where the data lives. The next phase is becoming the system of intelligence, where the data does work. One workspace where you find the data, trust the permissions on it, and take action on it.
The early signal: Rippling AI went to early release about two months ago, moved into a full 30-day free trial, and has already generated 400+ LinkedIn posts from users about the impact on how they run their businesses.
Clean Data Beats Clever Models
Every company is racing to add AI right now. The winners of this phase will be the companies whose data is connected enough that AI can be trusted to act on it. Rippling’s decade of building one graph, refusing to acquire and patch, was a deeply unsexy bet for years. In an AI-native world, it’s the thing that lets the product do work instead of just storing it.
If you’re a founder, the takeaway is direct. Audit your data layer before you audit your AI roadmap. The model is the easy part. The graph underneath is what decides whether any of it works.
Top 5 Takeaways
- The data layer is the moat, not the model. Rippling’s decade of building one connected employee graph, with no acquisitions to patch together, is the thing that lets AI actually do work. Audit your data before your AI roadmap.
- The right product arc is insights, then actions, then proactive workflows. Answer questions first, take actions with guardrails second, surface things proactively third. Each stage earns the trust the next one requires.
- Permissions and strong typing are what make AI safe to act. Knowing who can see each field and exactly what each field means is the difference between an insight and a liability. Boring, and non-negotiable for production.
- Confirmations are correct for this trust phase. Multiple confirmation steps with a clear before-and-after on sensitive actions is the right call while teams get comfortable letting AI act for them. The comfort builds one action at a time.
- The same thing that’s your strength is your friction. Rippling’s all-in-one graph powers the AI and drives the complexity, learning curve, and reporting complaints. Whatever your moat is, know where it costs you, and make AI the thing that pays that cost down.
