When Software Becomes the Operator

Operational complexity does not disappear. It changes owner.

Every company in this study crossed the same line. It stopped selling a tool that a customer used to do the work, and became the party that performs the work and answers for it. Payroll gets filed. Money moves. A container clears customs. When it goes wrong, the vendor is the one holding it.

The lazy version of what follows is that these companies discover they were secretly service businesses all along. That is not what the record shows, and the best-documented case here contradicts it directly. What the record shows is narrower and more useful.

Companies that made this transition intact tended to build the system of record before they took on the obligation, and to grow the obligation slowly enough that the operating apparatus could be built alongside it. Companies that ruptured either took the obligation on as the business model, before any apparatus existed, or took it on top of real infrastructure and then let the volume outrun the people.

That second half is where the record thins out badly. Nearly every company that got through this grew its responsibility-bearing book over years, so the question facing the current generation of AI-native operators is one the past cannot answer. It is being tested right now, in public, by companies that have not finished.

Two other findings came out of the work that were not obvious going in. The ladder runs in both directions, and one of the largest companies here climbed back down. And the recurring failure across a decade of cases is not an operational failure at all. It is a measurement failure, and it repeats with unusual precision.

Five different businesses hiding under one word

"Software company" covers at least five distinct positions, each with a different cost structure and a different failure mode. Companies move between them without announcing that they have, which is why the transition is usually visible only in hindsight. This is a lens for reading the cases rather than a finding proved by them. No company is obliged to climb.

The distinction between rungs three and four matters more than it looks. At rung three, an unhandled exception is a support ticket. At rung four, the same exception is a late filing with a statutory penalty attached. The work is identical. The consequence is not.

Most of the mechanism in this study lives at rungs three and four. Rung five is a smaller and more specialized problem, handled separately near the end.

Who crossed, when, and what happened

Eleven cases, chosen so that a majority did not rupture. A cohort assembled only from wreckage proves its thesis by selection, and postmortems get written while quiet competence does not. Six came through intact. Two corrected course and are still operating. Three ruptured.

Bright Health Group and Oscar Health are treated separately in the rung-five section. Their evidence bears only on companies carrying balance-sheet risk, which most operators do not.

When the obligation arrives before the machinery

Zenefits sold free HR software to small businesses and made its money as a licensed insurance broker on the health plans those businesses bought through the platform. That structure put it on rung four in year one, before it had built anything that belonged there.

The consequence was that growing revenue and growing regulatory exposure were the same activity. Sales headcount was the growth lever. Sales headcount was also the licensing obligation. There was no configuration in which one could safely outrun the other, and nothing in the product would catch the gap, because the product was not the thing that was breaking.

Conrad's own statement on the way out named the same thing from the other side: the company's management infrastructure and policies had not kept pace with its growth. Both men are describing an apparatus problem rather than a software problem. That framing recurs in every rupture in this study, which is the first thing worth carrying forward.

Zenefits to Rippling, and what the comparison can prove

The same founder built a second company in the same category and did not repeat the same failure. That is irresistible as a story and treacherous as evidence, so it is worth separating what the record supports from what hindsight supplies.

What is documented

Fact Rippling was founded in 2016 and launched in 2017, starting with employee onboarding. Conrad described the architecture to TechCrunch in October 2022: an employee graph at the bottom of the stack, middleware above it for reports, permissions, and workflow automation, and products on top of that. Rippling PEO was announced to the broker market around November 2020, four years in, on top of payroll and benefits that already ran. Global payroll came in October 2022, six years in.

Fact The November 2020 PEO launch was aimed at brokers rather than around them. It offered brokers real-time visibility into client data and renewal information so they could advise clients on whether a PEO remained the right fit, one-click data export so a client could be taken back to market, and moving clients off the PEO in days rather than months. Zenefits had made itself the broker and taken the commission. Rippling made itself something the incumbent channel could place, and built the exit into the product at launch.

What hindsight is supplying

Inference No public statement from Conrad establishes that PEO was sequenced later because of Zenefits. The compound-startup thesis he has argued repeatedly produces exactly this ordering as a byproduct of building many products on one data model, so the compliance benefit may be a consequence rather than a motive. Conditions also differed enormously: roughly $1.85 billion raised into a mature software market, against a company that had been building the category.

And Rippling is not a clean success

Fact Rippling conceded in writing three separate times in its 2026 Better Business Bureau responses that capabilities confirmed during the sales process failed after migration had already begun. It was recruiting foundational members of a Global Implementation team at approximately $1 billion in annual recurring revenue while selling enterprise deals, and in February 2026 launched a forward-deployed engineering organization to internalize work its own announcement described customers otherwise hiring third-party consultants to do. Conrad has publicly acknowledged discovering that the 90th percentile support experience was far worse than the median.

Fact Rippling does not appear on the IRS list of active Certified Professional Employer Organizations dated April 15, 2026. Justworks has been certified since January 1, 2017. Careful sequencing and full regulatory credentialing are separate decisions, and these two companies made them differently.

The supported lesson is narrower than the story and more useful. Compliance risk and delivery risk are different operator problems with different causes. At Zenefits, responsibility was embedded in the revenue model, regulated exposure scaled with the sales floor, and controls failed. At Rippling, regulated responsibility was layered onto a mature system of record, the compliance failure did not repeat publicly, and implementation complexity surfaced anyway, later and at far greater scale.

Why the apparatus tends to arrive after the revenue

Across multiple cases here, the licensing, implementation, underwriting, or exception-handling function was built after the volume that required it. The mechanism is unglamorous. Revenue-generating functions get resourced first because they are visible and measurable. Back-office functions get resourced when they break. The gap between those two moments is set by growth rate rather than by company size, which is why it shows up at a billion in revenue as readily as at fifty million.

This is a tendency and not a law, and the exceptions matter. Justworks does not show it. Deel does not appear to, though it publishes nothing that would let anyone check. Presenting the lag as universal would be wrong.

The reliable tell is documentary. In several of these cases the company was asking new hires to write the process documents for work already running at scale. Implementation checklists, underwriting playbooks, standard operating procedures, and onboarding templates normally predate volume. When a job posting asks a new hire to build them, the volume got there first.

What separated the survivors was rarely whether they had the lag. It was how the gap surfaced. In the clearest rupture cases, the operating gap became visible externally before management had demonstrated it was under control. Zenefits learned its licensing was broken from a journalist. Olive's customers reported the promised savings had not materialized before the company conceded it. Bright Health's cost structure emerged through reserve development. Internal awareness is not resolution and guarantees nothing, but the sequence in which a problem becomes public is itself informative when reading any company's current position.

Automation removes the middle of the distribution, not the edges

The question is not whether automation works. In every case here it worked on the thing it was pointed at. The question is which work goes first, which stays human longest, and how much the labor curve actually bends once the vendor is answerable for the result.

Regulated, correctness-critical work has a long tail where the rule is ambiguous, the data is missing, or two obligations conflict. Automation handles the well-specified center of that distribution efficiently. The tail needs judgment. Because cost per item is highest in the tail, removing half the item count removes considerably less than half the labor.

The mature comparison is instructive because nothing was walked back. TriNet launched an AI assistant in spring 2026. By its second-quarter earnings call on July 30 it reported that the assistant had addressed 50 percent of customer-initiated chat sessions and lowered service case volumes. In the same quarter the company was growing its sales force by approximately 20 percent, running more than 100 new representatives through a training program, and accelerating investment in its service model.

Half the sessions handled, service capacity redeployed, and the organization still growing. That is what a working deployment looks like at the operator boundary. It is a real gain, correctly sized.

Deel supplies the blunt version of the same point. A company reported to have crossed a billion dollars in annual recurring revenue employs on the order of nine thousand people to hold employment liability in a hundred and fifty countries. Reported Those figures come from secondary sources and are unaudited, and they are doing real analytical work here, so treat them as directional. What they suggest is that fast operator scaling is survivable, and that the price is paid in headcount.

This finding repeats with unusual precision. Klarna measured coverage, speed, and repeat-contact rate. What ultimately determined the outcome was quality on complex and emotionally loaded cases, which those metrics did not capture. Rippling published first response time. Every 2026 customer complaint in the record concerns time to resolution and ownership. Olive AI told hospitals its technology would cut administrative spending by a factor of five, and an April 2022 Axios investigation reported that the savings calculations rested on rough estimates and were tracked only when a client asked.

Three companies, three different technologies, a decade apart. In all three, the measured variable was throughput and the failing variable was resolution.

Gross margin falls. Operating margin can rise. Both are in the same filing.

The lazy claim is that operator businesses have service economics and are therefore worse. The best-documented case in this study says something more specific, and Shopify says it about itself.

At Shopify, the gross-margin difference between the two revenue lines is roughly a factor of two. The point is not that operator revenue always cuts gross margin in half. It is that adding responsibility can materially alter the economics of the revenue line while still improving operating leverage elsewhere, and that a company can hold both effects at once without either one settling whether the move was correct.

What actually changes by rung is more specific than a margin. At rung three, labor intensity and cost to serve become the binding constraints, and they scale with transaction volume. At rung four, the new costs are licensing, certification, audit, controls, and legal exposure, and they behave like a fixed floor that has to be in place before the volume rather than after it. At rung five, working capital and reserving appear, and a bad year becomes a capital event rather than an income statement event.

The mature reference point is TriNet, which publishes the full picture quarterly. In the second quarter of 2026 it reported total revenue of $1.2 billion, of which professional services revenue, the fee for administering the work, was $159 million. Adjusted EBITDA was $128 million at a 10.9 percent margin. Most of the top line is pass-through. Any read of an operator's gross revenue as though it were software revenue will be wrong by close to an order of magnitude.

The ladder runs down, too

Almost every discussion of this transition treats it as one-way. The largest company in this study demonstrates otherwise, and did it deliberately.

Shopify announced the Shopify Fulfillment Network in June 2019, acquired 6 River Systems that year for $450 million, and acquired Deliverr in May 2022 for $2.1 billion, its largest acquisition. That is a software company climbing to rung three and administering physical logistics on merchants' behalf.

On May 4, 2023 it sold the logistics business to Flexport in exchange for roughly 13 percent of Flexport and a board seat, sold 6 River Systems to Ocado, and cut 20 percent of its workforce. The transaction completed on June 6, 2023.

Fact Shopify reported a first-quarter operating loss of $193 million on the same day, against $98 million in the comparable prior period. Shares rose sharply following the combined earnings, restructuring and logistics announcement, gaining as much as 24 to 29 percent in the following session depending on the exchange and the window measured. The three items were announced together, so the move cannot be attributed to the divestiture alone.

Harley Finkelstein's framing at the time was that the services could be delivered more effectively through integration than through ownership.

The uncomfortable symmetry is that Flexport, the receiving party, absorbed this expansion in the middle of its own contraction, and was cutting 20 percent of its own staff five months later. The work did not become easier by changing hands. It went to a party whose cost structure was built for it, which is a different claim and the one that matters.

Once you own the outcome, the channel is part of operations

In a rung-one business, a channel produces leads and the analysis stops at cost per acquisition. Once the vendor owns the outcome, the cost of a customer includes getting their data right, setting them up, and handling their exceptions for as long as they stay. A channel that delivers customers who are already advised, already structured, and already carrying clean records reduces all three. That makes channel design an operating decision.

TriNet is the clearest instance. Fact On the July 30, 2026 earnings call, Simonds noted that more than 90 percent of small and medium businesses offering health care obtain it through a health insurance broker. TriNet's broker channel represented 32 percent of new sales in the quarter, with requests for proposals up 54 percent year over year. He described retention-based incentives to brokers as critical to alignment, and the operating design around it as putting the right people at the local level against the right broker producers, giving brokers trusted-adviser access as standard procedure, including them in renewal discussions, and assigning dedicated client service staff against the broker block rather than customer by customer. That is a delivery architecture rather than a marketing program.

Gusto shows the same mechanism outside the PEO model. Reported by the company More than 23,000 accounting firms held active partner accounts as of April 30, 2026. The program performs implementation work directly: partners receive onboarding help, free payroll transfers, automated import from prior providers, and a multi-client dashboard. The accountant becomes the first line of exception handling for their own book and holds the client relationship through the year.

Rippling's 2020 PEO launch gave brokers renewal visibility and one-click export, which reduces friction at entry and at exit. An exit path lowers the service burden of clients who should not be on the product in the first place. Zenefits is the negative instance: by becoming the licensed broker itself, it collapsed distribution and regulatory exposure into a single activity, so the channel could not absorb any operating load because the channel was the operating load.

The transferable principle is that in an operator business, the best channel may lower the cost of delivering the outcome as well as sourcing it. Mechanisms observed across these cases include structured data arriving at intake, a trusted adviser absorbing first-line questions, better-qualified customers, standardized implementation, easier renewals, and a cleaner path out for customers who do not fit. The limit is worth stating. TriNet expects direct sales to remain more than half of new acquisition, so the broker channel is an addition rather than a replacement, and the mechanism is strongest where the reason for purchase is itself intermediated.

What actually tells you whether the transition is working

No thresholds are offered here, because the cohort does not support them. What the cohort does support is which quantities matter, and they are consistently not the ones companies publish at the moment they cross the line.

The pattern running through the top half of that table is the same measurement failure described earlier. Throughput is easy to instrument and gets reported. Resolution quality is hard to instrument and is what breaks.

Separating what history establishes from what is still being proved

Much of the automation evidence in this study is a generation old. Olive AI ran robotic process automation. Babylon ran probabilistic symptom modeling built in the 2010s. Treating those as evidence about current models would be a category error, so the split has to be explicit.

What history establishes

That exception-heavy regulated work has a long tail, that the tail carries the cost, and that automation reaches the tail last. That claim is supported across a decade and across technologies, and current systems do not obviously escape it. Klarna is the strongest current-generation instance and points the same way, with the qualification that its assistant did work and the walk-back concerned scope rather than capability.

What is genuinely open

How fast the tail shrinks with systems that can read primary sources, reason across conflicting rules, cite what they relied on, and abstain when the inputs are thin. Nothing in the historical cohort had those properties. The most interesting design pattern appearing in current operator companies is not higher automation coverage but explicit abstention, versioned prompts, evaluation suites, and decision logs built so an AI-assisted output can be reproduced for an auditor. Whether that produces a materially different labor curve is not yet visible in any published operating metric this research could find.

What has not changed at all

Rungs four and five are legal and financial positions rather than technical ones. A model cannot hold a license, post a bond, or be the named party on a filing. Certification timelines, audit requirements, carrier relationships, and capital adequacy move at the speed of institutions. Whatever automation does to the cost of the work, it does nothing to the calendar of becoming permitted to do the work.

The financial risk case, kept separate on purpose

Rung five applies to a minority of operators, so it is held apart rather than folded into the general argument. Where a company's balance sheet absorbs the outcome, additional mechanisms appear, and two well-documented cases sit on either side of them.

Fact Bright Health Group announced in January 2022 that it was positioned to pass one million members. It reported a loss of roughly $1.2 billion for 2021, raised $175 million in October 2022 to carry the business to profitability, then exited the individual market entirely, giving up approximately 970,000 commercial members and releasing roughly $250 million in regulated capital. That last figure measures how much balance sheet the book had been consuming without anyone having planned for it.

Fact Oscar Health faced the same mechanism in the same market in overlapping years. Its medical loss ratio deteriorated to 87.4 percent for full year 2025 and 95.4 percent in the fourth quarter alone, producing a $443.2 million net loss against a $25.4 million profit in 2024. It repriced rather than exiting, and reported second-quarter 2026 net income of $361.8 million with full-year operating earnings guidance raised to between $500 and $700 million.

What separated them was capital sized to absorb a full reset year, a market with a scheduled annual repricing mechanism, and the willingness to take the loss year rather than leave. Fourteen years passed between Oscar's founding and that result.

The general point is the one worth carrying: financial risk changes the severity and the economics of the operator transition. It does not change whether the transition has occurred. Zenefits failed at rung four carrying no claims risk at all.

Seven questions, in the order the evidence suggests asking them

Not a recommendation about whether to climb. A set of questions the cases indicate are load-bearing, arranged so the cheap ones come first.

Two further questions sit behind these and are worth holding in mind rather than scoring. The institutional calendar for licensing, certification, and carrier relationships runs on its own clock and is the constraint automation does not touch. And descending a rung remains available, with precedent, though it is almost never in the strategy deck.