Rippling skipped the chatbot and stamped employees after 90 days of work

Sarah Guo opens No Priors by calling the episode spicy. Rippling, she says, is a juggernaut workforce platform unifying HR, IT, and finance, and it is launching an AI product that reads employees’ work and emits performance-management signals. Matt MacInnis, the COO, describes the company more plainly: an all-in-one platform that does the boring, important work of running a company and tries to kill the administrative burden. Most customers arrive needing payroll, or device management, and then buy the rest. Guo mentions a rumor of many hundreds of millions of dollars in revenue, growing fast. He answers, “It’s going well.” About 3,500 employees, tens of thousands of customers. Something is working.

The company has on the order of 25 unique SKUs a customer can buy. Small things ship every quarter; bigger things every couple of quarters. An applicant tracking system just landed on the HCM suite. Scheduling, unsexy and in his view actually cool, is about to ship. Guo’s summary is that this is less one company growing than 25 startups compounding from a smaller base. The hiring strategy is explicit. More than 150 people at Rippling have started companies. Entrepreneurs whose ideas did not quite work out get a safe place to land, keep pursuing what they cared about, or do something new. MacInnis raises his own hand: he has been there. He says it shows up in shipping velocity.

Customers, he adds, already want to consolidate software onto one platform, which is a tailwind for the compound-startup bet. They will keep recruiting operators and shipping. The financial picture is the part he wishes he had learned earlier as a founder. Look at IT, finance, and spend as separate suites and they are somewhat subscale. Aggregate them and the top line is beautiful, with efficiency still waiting to blossom as the suites play to one another.

MacInnis’s regret as a young entrepreneur was not learning basic corporate finance, the income statement and the balance sheet and what investors actually look at. He thinks he has mostly figured it out. For scaled SaaS, he says, unit economics converge on the cross-sell motion. New logos matter, but as you sell more products into a growing base, the economics look more like that bucket than like hunting. Rippling has a long list of things it can sell to people it already has. That is the compound-startup story as a spreadsheet, not only as a product strategy.

Consolidation’s surface benefits, fewer sales teams and lower customer spend, are in his view super basic. The magic is something common underneath the applications that gives you scale, or what he calls a vibranium advantage: a superpower at the core that makes competitors ask how the hell they did that. For Rippling it is a deep understanding of the employee graph. Everything they build runs on those rails. The question they started asking a few years ago was what happens when you marry that structure and history in one platform, then hand it to a large language model.

The answer they are shipping is Talent Signal. The system reads work product, marries it to whom you hired, at what job level, with basic job history, and yields an insight into how those people are doing. He thinks everyone knows, at some level, that AI will contribute to evaluating human performance. Traditional HR and IT systems do not necessarily hold the work itself. That break, looking at output rather than the HRIS record, is the product.

MacInnis spent time as an individual contributor at Apple. The crunch with a middle manager, he says, arrives at review time. You have an opinion, peers have an opinion, the manager has an opinion, everyone writes feedback, then they sit in a calibration room and try to hold one another to a rubric. The truth is the manager has not sat there and looked at everything you did over six or twelve months. There is not enough time. A recent *Harvard Business Review* piece he cites names the failure mode manager vibe. If the boss has a good vibe and performance is ambiguous, the hole fills with that vibe. If the vibe is negative, they drive the negativity through the crack like a Mac truck.

Talent Signal tries to get around that human process by going to the source. It does not see race, ethnicity, age, or work location. It sees the source code you wrote, or the customer interactions you had as a support agent, and it stamps a signal: high potential, typical, or in need of attention. They call the last bucket “pay attention”; it effectively means at risk, and it is there to send the manager to spend time. It also surfaces concrete examples for a coaching conversation. ICs can see the stamp when the manager lets them. That permission was debated hard. The product, he keeps repeating, is not making employment decisions. It is an independent, calibrated signal.

Job level does go into the model, and they calibrate across every company in the training data, but customers only see a localized version. In a group of 50 engineers you will always see some high-potential flags and some who need attention, even if the global model thought the whole company was strong. Otherwise the signal is not useful. Stakes are high, so V1 is an early-access program with a hard constraint: one signal, one time, for one employee, at the 90-day mark. Even a three-year veteran gets a stamp based only on the first 90 days of work. Companies can then see that the system made a pretty darn good call at day 90 while they took twelve months to learn the person was a misfit or a star. That is how they intend to build trust without stretching the Overton window too far.

A 90-day signal on someone who has already been around three years is not incrementally useful for managing them now. It is backtesting. V1 is for new hires. The more you hire, the more useful it is, so high-growth companies get value first. The sky is the limit as trust grows. His personal motive is the bad manager. If you are in the bowels of the organization on hard problems, your manager is lazy, and they show up at calibration with a better vibe on someone else, that someone else gets the promotion. Talent Signal walks in, slams your work on the table, and asks what about this. Internally, an engineer in India, early in tenure, working one of Rippling’s toughest problems, was singled out as high potential. Leadership talked to her manager. The model had no vibe. It had examples. He also wants the inverse: a signal that someone has not ramped, visible to the manager and the manager’s manager, as a team-performance tool.

Dogfooding started early, and employee feedback shaped policy. No one is allowed to make a significant employment decision using the model alone. You cannot point at Talent Signal in a promotion conversation and stop there. You need your own independent assessment. He calls the product a cheat sheet. The manager still has to do the human work of evaluating the whole person. Internally they prohibit blind following. Rippling already dogfoods hard in analog ways: CEO Parker Conrad runs every payroll himself and approves every expense above ten dollars. MacInnis sometimes eats the ten bucks rather than fight Parker on Uber Comfort versus UberX. Employees, he says, have been good sports as testers and serious about representing their own rights in the design.

Elad Gil’s objection is the classic middle manager whose job is making other people successful, collaboration and focus, with no concrete work product. V1 does not go there. Signals exist for salespeople, support agents, and individual-contributor engineers. Managers will be interesting later. When they hunted for the best r-squared, the input most predictive of later promotion, performance termination, or staying stuck at a level, the screaming signal was work product. For engineers that means source, not just volume: quality, security issues, pull requests, comments on pull requests. Foundation models already reason about code, so they turned out to be excellent engines for assessing it, including maintainability and extensibility, expressed in language a manager can then judge.

Then the alpha-customer story. A CTO saw an engineer he did not think was strong flagged as high potential. That did not jibe with his priors, or his vibe. He opened the source and recognized it: he had written most of those commits. The employee had been struggling, so the CTO had been sitting shoulder to shoulder writing and coaching, and the model had scored the high-quality contribution that only appeared when he was there. MacInnis treats it as an unknowable misattribution from the outside. The only fix is the manager going to ground. It also revealed something no prior system could see: the struggling engineer was being coached at high quality.

Gil cites Andy Grove: a manager’s output is the team’s output, so perhaps you aggregate. The product does roll up a distribution of high-potential, typical, and needs-attention people under each manager. They tell customers to use that informationally, to spot hotspots, not as a verdict. You cannot yet tease hiring quality from execution. That is why they named it a signal, a yellow light bulb, not a score that closes the case.

Why this, of all the foundation-model toys? MacInnis zooms out to the same finance lecture. A dollar goes in the front of the machine, two come out the back; do not take the two, put them back in, get four, then eight. That is why early SaaS runs a deep cash deficit. If you do not know the next technology, the next customer, or the next revenue product, you do a buyback, which means the most creative use of cash is juicing the share price. A dividend is worse: you are literally handing the money back because you cannot think of anything. Rippling, like many Valley companies, thinks it knows what to do with the next dollar and wants more than it has, so it uses equity to pump the machine. The unusual fact, easy for the team to take for granted, is a massive list of projects that would turn into revenue if they built them. The constraint is hiring enough engineers and not running out of money.

The classic objection to doing many things is focus. How do you teach the sales team? How do you keep it cohesive? His one-liner is “just work harder”: get the right people in the right jobs and enough leaders who can live with a fractal of complexity. Gil notes this is the 1990s enterprise playbook, forgotten in a 2000s decade of point products, now back via HubSpot, Datadog, and Rippling around a system of record or an identity. The Netscape-era line is that innovation is bundling or unbundling. History does not repeat itself, MacInnis says, but it rhymes, and they are in the rhyming phase of 1990s platforms.

Talent Signal does not look like bundling. That is the difference between rhyme and repeat. New technology offers new opportunity. The guiding principle is always what they alone can do. Vibranium plus AI equals what? Other companies, in his telling, had rotten roadmaps: no next SKU with a near-certain revenue shot, only features that made existing customers happier. Then AI arrived and everyone became an AI company, a chance to masquerade as a firm that knows what to do with the next R&D dollar. Rippling never had that problem, so they did not build a chatbot, a copilot, or any of the surface-level obvious caps. They will build those. Who cares. They will not sell a single extra subscription. They put hard-to-recruit AI engineers on something whose opportunity cost was actually worth it.

They have other AI work in flight, but they are not peeling people off Talent Signal to start project two. They want this one right out of the gate. Separate from AI, a high-scale data platform already sits under Rippling. MacInnis teases an “AWS moment” in roughly the next quarter. Talent Signal lives on that platform. For engineering, you install the GitHub app on Rippling and it replicates the repository into a guarded environment for analysis. Plug in Salesforce and a lot of that instance, which can be massive, comes too. The hard part is marrying HRIS data with that scale while every employee identity is cleanly dereferenced: they always know who is who across systems.

That join is why he thinks the opportunity was obvious. GitHub can grow a code-quality tool. It does not know whom you promoted, who did well, or whom you had to let go for performance. Salesforce does not know either. Workflow products will get cool evaluators. None of them, he says, will know the human outcome the way Rippling does. That is the magic.

Some companies they have pitched have said they will not be early adopters this round. He respects that. His sense is that AI in the conversation about human performance is 0.1 percent of the way there, and that it is an inevitability that LLMs will be involved in assessing people in many contexts. Gil mentions a public-company CEO friend who already uses chat products to think through employee issues. The future is showing glimmers. MacInnis’s first principle is not to go numb. If you are not in the AI world, “risk” and “ethics” sound odd next to a model that converts half a cup of oil to ounces. In this context a hallucination or a misattribution can be consequential to a life. That is why they are in early access, constrained to 90 days, and conscious of bias the system might amplify.

Advice for anyone putting tools like this into personnel decisions: understand the stakes. Even if the system is arguably bulletproof, go to ground and still do the manager’s job, inspect the context, and do not let a coaching misattribution set your view. More broadly, you need some understanding of how these systems work in order to judge output quality, and he thinks that bar is currently too high for everyone who might become a user. The companies already partnering are performance-oriented, looking for tools to compete. He uses an Olympics analogy: a coach going for gold will try a beta that assesses form, if downside is mitigated. CTOs showed up on the engineering side. Sales leaders showed up because sales is hyper-competitive and this is part of the arms race. Support teams already live on rubric adherence and weekly air-checks, so they have the culture.

Guo notes that sales, engineering, and support are also the three functions with the heaviest coaching cultures: Gong call-sharing, pair programming, script training. He agrees 100 percent. Those were the right first verticals, and they roll up to company culture: some firms wait this round out, the more hard-edged want to play ball. She also flags the AI pitchfork crowd. He is thankful for them. It is easy to inhale your own exhaust and get excited about potential without the full picture. When people ask hard questions about bias or about the unintended consequences of putting AI in decisions this important, they will listen and learn. “Feedback is a gift,” he says, and he means it. They are committed to making a tool that works for everybody.